Coordination Theory and Collaboration Technology (Volume in the Computers, Cognition, and Work Series)

Citation preview

.(Eds.) ~acilitating the Development and of Use Interactive Learning Environments Q

S.)

ideo-~ediated Communication

Q

x The ~ u m a n Tutorial Dialogue Project: Issues the in Design ofInstruct~onal Systems Q

(Ed.) Com~utersas Assistants: A ~ e ene w era ti on of Support Systems Q

n (Ed.)

Q

CSCL: Theory and Practiceof an Eme~ing

mm11(Eds.) * Design Rationale: Concepts, Techniques,

and Use

~ Q n e / S(Eds.) ~ i ~ ~oordin~tion Theory and Collaboration Technolo~y Q

a n (Ed.) Adaptive User support: Ergonomic Design of ~anually and A~tomaticallyAdapt~bleSoftware Q

Collective Intelligence in Comput~r-~ased Collaboration

~ n i v e r s oif~~ o r t hCa

2001

~

a Jersey ~ New w ~

,

Copyright

@ 2001 by Lawrence Erlbaum Associates, Inc.

All rights reserved. No part of the book may be reproduced in

any form, by photostat, microform, retrieval system, or any other means, without the prior written permission of the publisher.

Lawrence Erlbaum Associates, Inc., Publishers 10 Indus~alAvenue M ~ wNew ~ Jersey , 07430

Coordination theory andcollabor~iontechnology / edited by Gary M. Olson, Thomas W. Malone, John B. Smith. p. cm. Includes biblio~aphicalreferences and index. ISBN 0-8058-3403-6 (cloth:alk. paper) 1. Group work in research-United States. 2. ~ese~ch-UnitedSt Information technology-United States. I. Olson, Gary M. 11. Malone, Thomas W. 111. Q180.55.G77 C66 2001 ~ 1 . 4 ’ ~ 8 ~ c 2 1 2~1018986

Books published by Lawrence Erlbaum Associates are printed on acid-free paper, and their bindings are chosen for strength and durability. Printed in the United States of America l 0 9 8 7 6 5 4 3 2 1

is uo~umeis de icated to the m ,whose enthusiasm ny researchers to take on the e u a ~ ~ a t information in~ s y s t e ~ in s re~atio~ to their social, Q~~anizatiQnal, and ~cQnomic cQ~texts.

This Page Intentionally Left Blank

1

1.

ACHES TO COO

5

"he Interdiscip~aryStudy of Coordination

7

Thomas W. alone and Kevin Crowston

2.

Communication and Collaboration in Distributed Cognition 51 V. ~ e n ~ s i

Richard J. Boland, Jr. and Ramkris~nan

3.

Coordination as Distributed Search

67

~ d m u n dH. Durfee, Daniel Damouth, Piotr J .Gmytrasiewicz, J. Huber, Thomas A. ~ o n t g o m e and ~ , Sandip Sen

arcu us

4.

Strategic Negotiationin ~ u l t i a ~ e n t E n v i r o ~ e n t s

93

Sarit Kraus and Jonathan Wilkenfeld

5.

sign Princi~lesfor Co~aborationTechnology: Examples of Semiformal Systems and Radical Tailora~ility Thomas W. alone, Kum-Yew Lai, and Kenneth

6.

On Econoxnies of Scope in ~

125

R. Grant

o ~ ~ c a ~ o n

161

Tho~as ~arschak

7.

Knowledge, Discovery, and Growth

193

Stanley Reiter

261

8.

~nfra§tructureand ~ ~ ~ l i c a tfor i o~~o sl l a ~ o r a t i v ~ oftw ware Engineering

263

Praszen Dewan, ~ a ~ ~i ads h a y e k h i ,and John Riedl

vii

viii 9.

Cooperative Support for DistributedS u p e ~ Control ~ o ~

311

Christo~herA. Jasek and Patricia M. Jones

10.

Trellis: A Formally Defined Hypertextual Basis for ~ t e g r a t ~ Task g and~ ~ o ~ a t i o n

341

Richard ~ u r u t aand P. David Stotts

11.

Problems of Decentralized Control: Using ~ a n d o m i ~ e d ~oordinationto Deal With Uncertainty and Avoid Conflicts

369

JosephPasquale

12.

The Architecture and Implementation of a Distributed ypermedia Storage System D o ~ g l a sE. S h a c ~ ~ o r John d , B. Smith, and F. Donelson Smith

391

409

13.

~ o m m ~ i c a t i and o n ~ o o r d ~ a i~n Reactive on Teams R o n a l ~C. Arkin and Tucker B a l c ~

411

14.

Seeding, Evolution~yGrowth, and ~crementalDevelopment of Collab E n ~ ~ o ~ ~ t s

447

Gerhard Fischer, Jonathan Grudin, Raymond M ~ C a l l ,Jonathan O s ~ a l d David , Redmiles, Brent Reeves, and Frank S h i ~ m a n

15.

i s ~ i ~ u Group ~ e d Support Systems: Theory ~ev~lopment and Exper~entation

473

Starr Roxanne Hiltz, Donna Dufner, Jerry ~ermesta~, Yo~n~in Kim, Rosalie Ocker, Ajaz Rana, and Murray ~ u r o ~ 507

16.

, G ~ r ~ ~ aand n iMargaret , Elliott Yannis ~ a k Q sUjay 17.

uter Support for Distri~utedCollaborative

A ~ o o r d ~ ~ tScience i o n Perspective

Christine M. ~ e u w i r t h ,David S. ~ a ~ R ~~ i nrd e,r C h a n d h oand ~ James H. Morris

535

CONTENTS 18.

TechnologySupportforCollaborativeWorkgroups

559

Gary M. Olson and Judith S. Olson

585

IZ 19.

Central Coordination of DecentralizedI n f o ~ a t i o n in Large Chains and Franchises

587

Toby Berger and ~ i c h o l a sM. Kiefer

20.

~ r g a n ~ a t i o nPerformance, al Coordination, and~ognition Kathleen M. Carley

595

21.

Computational Enterprise Models: Toward Analysis Tools for Designing~ g a n i z a t i o ~

623

extend in^ Coordination Theory to Deal With Goal Conflicts

651

R ~ m o n dE. Levitt, Yan Jin, Gaye A. Oralkan, John C. Kum* and Tore R. Christiansen

22.

C l ~ t o nLewis, Rene Reitsma, E. Vance Wilson, and Ilze Zigurs

23.

Modeling Team Coordination and Decisions in a Distributed Dynamic Enviro~ent

673

Wei-Ping Wang, David L. ~ l e i n m a n and , Peter B. Luh

ES 24.

71 1

SocialTheoreticalIssuesintheDesignof Co~aboratories: Customized Software for Community Support Versus Large-scale Infrastructure

713

A Path to Concept-Based Information Access: From National Collaboratories to Digital Libraries

739

Geoflrey C. Bowker and Susan Leigh Star

25.

~ s i n c h u nChen and Bruce R. Schatz

26.

Technology to Support Distributed Team Science: The First Phase of the Upper Atmospheric Research Collaboratory (UARC) Gary M. Olson, Daniel E. Atkins, Robert Clauer, Terry Weymouth,

761

~ h o A. m Finholt, ~ Atul Prakash, Craig Rasmusse~, and Farnam Jahanian

Author Index Subject Index

785 801

This Page Intentionally Left Blank

The global revolutionin human interconnectedness goes back at least to the last century, By the turn of the century, we had telegraphy and telep~ony that startedus on the pathto a wired world, networked transportation (e.g., railroads) that fundamentalls altered how we organized our communities (e.g., Stilgoe, 198!j), and office technology that began altering the ways we were ableto organize ourselvesVates, 1989). Throughout the 20th century the pace of innovation and change accelerated, with each generation lauding the next apparent steps toward a better world. However, by the latter decades of the 20th century we began to realize thatthetechnologies of interconnectednesswerenot by themselvesa full of surprises. When new technologies panacea. Rather, the century was of connectedness emerged the initial inclination was to think of possible gains, suchas efficiency gainsin speed or accuracy (Sproull& Kiesler, 1991). But repeatedly we are surprised by second-order effects that no one had predicted. For instance, electronic mail provides what seems like an effective asynchronous communication medium of considerable fle~ibility,making it possible for people anywhere ona relevant network to communicate easily. But new problems arise. First, making it easy to communicate makes in efforts to filter it too easy to become flooded with information, resulting to reasonable levels (e.g., Malone, Grant, Turbak, Brobst, the message stream & Cohen, 1987). Second, something about the medium itself appears to make users ofemail cast their messages in ways that come across as hostile or negative to recipients (Sproull& Kiesler, 1991), adding a further unpleasant

dimension to email usage. Tobe sure, email use is flourishingas this century comes to a close, butit is a more complex and surprising medium than first thought. Repeated surprisesof these sorts haveled many to call for multidisciplinary research that would bring together experts on human cognitive, social, and organizational behavior with the technologists whoare designing and of numerous deploying the new tools. Indeed, the 1980s witnessed the birth new conferencesto explore the issues invohed, suchCHI, as CSCW, ECSCW, and I N T E ~ C TNumerous . smaller workshops and seminars furtheraccelerated the paceof multidisciplinary discussion. Itwas in thiscontextthattheNationalScienceFoundation(NSF) announced a funding program in the area of Coordination Theory and Collaboration Technology(CTCT, or CT squared). This initiative was announced in calls for proposals issued in 1989 and 1990. It grew out of a seriesof workshops fundedby NSF that began the process of defining thisarea of research and creatinga community of researchers whose work might be funded. Two workshops on Coordination Science were heldMIT at (June 1987 and under the leadership of Thomas Malone, and one on Open Systems at Xerox PAW (June 1988) that was organized by Bernard0 Huberman. These workshops gave strong indication that the questions about coordination asked in many disciplines are similar and that ideas and methods from each could inform the other. Two workshops were sponsored by NSFon issues of technologyand by Jolene Galagher at the Universits of Aricooperative work, one organized zona (February 1988), the otherNew at York University (June 1988) under the leadership of ~argrethe Olson. These workshops focused on the behavioral science aspectsof coordination and collaboration relating to the interaction of individua~s, small groups, and formal organizations and on computerbased systems and software design for collaboration. Two influe~tialbooks came from these workshops (Galegher, Kraut&C Egido, 1990; M. Olson, 1989). announcement of the CTCT initiativecapturesmany of the themes exploredin these anticipatory workshops. The initiative focuses on processes of coordination and cooperation among autonomousunits in human systems, in computerandcommunicationsystems, and in hybrid o r ~ ~ i z a t i o n ofsboth systems. ...This initiative is motivated by threescientificissues which havebeenthefocus of separate research efforts, but which may benefit from collaborative research. The first is the effort to discover the principles underlying how people collaborate and coordinate work efficiently and productively in environments characterized by a high degree of decentralized computation and decision-making. The second isto gain a betterfundamental understandingof the structureand outputs of organizations,industriesandmarkets which incorporatesophisticated,

decentralizedinformationandcommunicationstechnology as animportant component of their operations. The third is to understand problems of coordination in decentralized, or open, computer systems. [NSF 88-59, p. 21

Two rounds of CTCT competition were held, one in 1988 (with results announced in June of 1989) and a second in 1990. From these two roundsof competitions,20 projects were funded out of 136 applications. In subsequent years, additional projects were funded out of the regular NSF unsolicited proposal competition, so that by the time the CTCT group held a seriesof three grantee workshopsin 199%1992, and 1993, nearly 30 different projects participated. These annual workshops gave the investigators in this program an extended opportunity for interdisciplinary exchange. A closely related activity took placein March 1989, several months after the submission deadline for the firstCTCT Initiative. A major workshop on scientific collaboration was held at Rockefeller University under the direction of Joshua Lederberg and Keith Uncapher. This workshop focused on the feasibility of designing network dependent multi-purpose systems to supin specific disciplines. The components port remote scientific collaborations of such systems would include multimedia communications, remoteaccess to instrumentation, digital journals and libraries, aand varietyof services to support science. This initial workshop led to a seriesof follow-on meetings that by 1993 resulted in an influential National Research Council report on “collaboratories”(cf. NAS report).Several of theoriginal CTCT projects clearly fit into this “collaboratory” concept, and additional projects were included in the CTCT workshops described earlier. A recent review showed that collaboratories were becominganincreasinglywidespreadformof sociotechnical support for distributed team science (Finholt & Olson, 1997). Laurence Rosenbergof the National Science Foundations wasa key player in all of these activities. Throughout this period he was managerof the Information Technology and Organizations program, anda deputy director of the Division of Information, Robotics, and Intelligent Systems(IRIS) in the directorate for Computer and Information Science Engineering (CISE). He was pri~arilyresponsible for funding these various workshops that ledto in the closely related collaboratory the CTCT initiatives, and was also active arena. In particular, heserved as the NSF program manager for the nearly 30 projects funded under theCTCT initiative. In the springof 1994 Larry Rosenberg died of cancer after a relatively brief illness. The three editors of this volume decided that a fitting tribute to Larry’s efforts N atSF would be for those of us who had been funded through a book that reportedon his interdisciplinaryCTCT initiatives to put together the work we had done. This effort was facilitated by the fact that those of us so funded had cometo know eachother quite well through the annual work-

shops of theCTCT program in the 1991-1993 period. We had become familiar with each other’s work, and had explored a number of themes and crossconnections among the projects. These links are reflected in a number of ways in the individual chapters in this volume. The multiyear projects funded by the CTCT program were all invited to submit chapters for this volume, and almost all of those invited agreed. The initial drafts of chapters were carefully reviewed by other authors,and revisions were obtained. Thus,we used the collegiality established by Larryin the program and its annual workshops to facilitate the production of this volume. We offer these chaptersto honor the memory of his efforts on all of our behalf. Special thanks must go to Sue Schuon, Deborah Jahn and Dawn Nugent for their diligent assistancein bringing this volumeto completion.

--Gary M. Olson -~homas W Maio~e -John B. ~ m i t h

Finholt, T. A.,& Olson, G. M. (1997). From laboratories to collaboratories:A new organizational form for scientific collaboration.Psych and technical foundations Galegher, J., Kraut, R. E., & Egido, C. (1990). Intellec~al teamwork:S~ial of c~peratjue work. Hillsdale, NJ: Lawrence ErlbaumAssociates. Malone, T.W., Grant, K R., Turbak, :F.A., Brobst, S. A., & Cohen, M. D. (1987). Intelligent information-sharingsystems. Communicationsof the ACM,30,390-402. support for work group colla6oration. Hillsdale, NJ: Lawrence Olson, M. H. (1989). ~echnol~ical Erlbaum Associates. Stilgoe, J.R. (1985). Me~opolitancorridor. New Haven, CT Yale University Press. Sproull, L., & Kiesler, S. (1991). Connections: New ways o f w o r kin i ~the networked o ~ a n ~ a t i o n . Cambridge, MA:MIT Press. Yates, J. (1989). Control t h ~ communication: ~ h The rise of system in American man~ement.Baltimore: Johns Hopkins UniversityPress.

P A R T

This Page Intentionally Left Blank

C H A P T E R

Thomas W. Malone Kevin Crowston

Massachusetts Institute of Technology

*

In recent years, there hasbeen a growing interest in questions about'how the activities of complex systems can be coordinated (e.g., Bondlk Gasser, ;Huberman, 1988b; Huhns & Gasser, 198%Johansen, 1988, NSF, 1991; NSFIRIS,1989; Rumelhart, et al., 1986; Winograd lk Flores, 1986). In some cases, this work has focused on coordinationin parallel and distributed computer systems; in others, on coordination in human systems; and in many cases, on complex systems that include both people and computers. Our goal in this chapter is to summarize and stimulate development of theories thatcan help with this work. This newresearch area-the interdisciplinary studyof coordination~rawsupon a varietyof different ~isciplines including computer science, organization theory, management science, economics, linguistics, and psychology. Manyof the researchers whose efforts can contribute to and benefit from this new areaare not yet awareof each other's work. Therefore,by summari~ing this diverse body of work in a way that emphasizes its common themes, we hope to help adefine communityof interest andto suggest useful directions for future progress. There is still no widely accepted name for this area, so we will use the t ~ Qtonrefer to theories about how coordination can term c ~ ~ ~ i n ~theory 'Reprinted (with permission) fromACM C o ~ ~ uSurveys, t j ~ Vol. 26, No. l, March 1994, 87119. Copyright 0 1994 Association for Computing Machinery,

MALONE AND CROWSTON

occur in diverse kinds of systems. We use the termtheory with some hesitation because it connotes to some people a degree of rigor and coherence that is not yet presentin this field. Instead, the field today aiscollectionof intriguing analogies, scattered results, and partial frameworks.We use the term t~eory,however, in part to signifya provocative goal for this interdisciplinary enterprise,and we hope that this chapter will help us move closer toward that goal.

Webeginwithoneofthequestionsthatcoordination theory mayhelp e ways answer: How will the widespread useofinformationtechno lo^ c h a ~ the ~ e o p l work e together?This is not the only possible focus of coordination theory, but it isa particularly timely question today for two reasons. First, in recentyears,largenumbers of peoplehaveacquireddirect access to computers, primarily for individual tasks like spreadsheet analysis and word processing. These computersare now beginning to be connected to each other. Therefore, we now have, for the first time,opportunit~ an for vastly larger numbers of peopletousecomputingandcommunications capabilities to help coordinate their work. For example, specialized new software has been developed to (a) support multiple authors working together on the same document,(b) help people display and manipulate information more effectively in face-to-face meetings, and (c) help people intelligently route and process electronic messages (see detailed references in Section 3.3).

It already appears likely that there will be commercially successful products of this new type (often called ‘~omputer-supported cooperative work” or “groupware”), and to some observers these applications herald a paradigm shift in computer usage as significant as the earlier shifts to time-sharing and personal computing. It is less clear whether the continuing development of new computer applications in this area will depend solely on the intuitions of successful designers or whether it will also be guided by a coherent underlying theory of how people coordinate their activities now and how they might doso differently with computer support. in the costs and capaSecond,in the long run, the dramatic improvements bilities of information technologiesare changing-by orders of magnitud~the constraints onhowcertainkinds of communication and coordination can occur. At the same time, thereaispervasive feelingin businesses today that global interdependencies are becoming more critical, that the ofpace change is accelerating, and that we need to create more flcxible and adaptive organizations. Together, these changes may soon lead across us a threshold where entirely new waysof organizing human activities become desirable.

1. I ~ ~ ~ I S C I P STUDY ~ I N OF ~ COO~INATION Y

For example, new capabilities for communicating informationfaster, less expensively, and more selectively may help create what some observers (e.g.,Toffler, 1970) havecalled ~dhocracies”-rapidlychangingorganizations withhighly decentralized networksof shifting project teams. As another example, lowering the costsof coordination between firms may encourage more market transactions (i.e., more “buying” rather than ~aking”} and, at the same time, closer coordination across firm boundaries (such as “justin time” inventory management).

If we believe that new forms of organizing are likely to become more common, how can we understand the possibilities better? What other new kinds of coordination structureswill emerge in the electronically connected world of the near future? When are thesenew structures desirable? What is necessary for them to work well? To some estent, we can answer these questions by observing “leading edge” organizationsas they experiment with new technologies. But to understand the experiences of these organizations, we may need to look more deeply into the fundamental constraints on how coordination can occur. And to imagine new kinds of organizational processes that no organizations have tried yet,we may need to look even further afield for ideas, One way to do both these things-to understand fundamental constraints and to imagine new possibilities-isto look for analogiesin how coordination occurs in very different kindsof systems. For example, could we learn something about tradeoffs between computing and communicating in distributed in computersystemsthatwouldilluminatepossibilitiesforcoordination humanorganizations?Mightcoordinationstructuresanalogous to those used in beehives or ant colonies be useful for certain aspects ofhuman organizations?And could lessons learned about coordination in human systems help understand computational or biological systems, as well? For these possibilities to be realized, a great deal of cross-disciplinary interaction is needed. It is not enough justto believe that different systems are similar; we also need an intellectual framework for “transporting” concepts and results back and forth between the different kinds of systems. In the remainder of this chapter, we provide the beginnings of such a framework. We first define coordinationin a way that emphasizes its interdisciplinary natureand then suggest an approach for studying it further. Nest,we describe examples ofhow a coordination perspective can be applied in three domains: understanding the effects of information technology on human organizations and markets, designing cooperative work tools, and design in^ distributedandparallelprocessingcomputer sys-

~

~ AND CROWSTON O ~

E

tems. Finally, we briefly suggest elements of a research agenda for this new area.

~e all have an intuitive sense of what the wordc~rdinationmeans. M e n we attend a well-run conference, when we watcha winning basketball team,or when we see a smoothly functioning assembly line we may notice how well coordinated the actions of a group of people seem to be. Often, however, good coordination is nearly invisible, and we sometimes notice coordination most clearly when it is lacking. When we spend hours waiting on anairport runway because the airline can’t find a gate for our plane, when the hotel where we thought we had a reservationis fully booked,or when ourfavorite word processing program stopswo~kingin a new version of the operating system, we may becomevery aware of the effectsof poor coordination. in trying For many purposes, this intuitive meaning is sufficient. However, to characterize a new interdisciplinary area, it is also helpful to have a more precise ideaof what we meanby coordination. AppendixA lists a number of definitions that have been suggested for this term. The diversity of these definitions illustrates the difficulty of defining coordination, and also the variety of possible starting points for studying the concept. For our purposes here, however, it is useful to begin with the following simple definition: C~rdinationis ma nag in^ dependencies b e ~ e e na~tivities.

This def~nition is consistent with the simple intuition that,if there is no interdependence, there is nothingto coor~inate.It is also consistent with a long history in organization theoryof emphasizing the importance of interdependence (e.g., Galbraith, 1973; Hart & Estrin, 1990; Lawrence &C Lorsch, 1967; Pfeffer, 1978 Roberts & Gargano, 1989; Rockart &Short,1989; Thompson, 1967). As the definition suggests, we believe it is helpful to use the wordcoordinatio~in a fairly inclusive sense. For instance, it is clear that actors performing interdependent activities may have conflicting interests and that what might be called “political processes’’ are ways of managing them (e.g., Ciborra, 1987; Ming, 1980;Schelling,1960;Williamson,1985).Furthermore, eventhoughwordslike“cooperation,”“collaboration,”and omp petition" ‘This definition was particularly influenced by Rockart and Short (1989) and Curtis (1989). The importance of coordination in thisvery general sense was perhaps first recognized byHolt (1980,1988).

1. INTE~ISCIP~INARY STUDY OF ~ O O ~ I N A T I O N

each have their own connotations, they can each be viewed, in part, as ways of managing dependencies between activities. It should also be clear that coordination, as we have defined it, can occur in many kinds of systems: human, computational, biological, and others. For instance, questions about how people manage dependencies among their activities are central to parts of organization theory, economics, ma nag^ ment science, sociology, social psychology, anthropology, linguistics, law, and political science.In computer systems, dependencies between different computationalprocessesmustcertainly be managed,and, asnumerous observers have pointed out, certain kinds of interactions among computational processes resemble interactions among people (e.g., Fox, 1981; Hewitt, 1986; Huber~an,1988a, 1988b; Miller& Drexler, 1988; Smith& Davis, 1981).To give a senseof the approaches different fields have taken to studying coordination, we summarizein Appendix B examples of results about coordination from computer science, organization theory, economics, and biology. Even though we believethere aremore similarities amongthese different kinds of systems than most people appreciate, there are obviouslymany differences as well. One of the most important differences is that issues of incentives, motivations,and emotions are usuallyof much more concern in human systems thanin other kinds of systems. In computer programs, for example, the “incentives”of a program module are usually easyto describe and completely controlledby a programmer. In human systems, on the other hand, the motivations, incentives, and emotions of people are often extremely complex, and understanding them is usually an important partof coordination. of systemsmay Even in human systems, however, analogies with other kinds help us understand fundamental constraints on coordination and imagine new kinds of organizations that might be especially motivational for people. A primary vehicle for facilitating transfer among these different disciplines is identifying and studying the basicprocesses involved in coordination: Are in all coordinated systhere fundamental coordination processes that occur

tems? How can we represent and analyze these processes?Is it possible to characterize situationsin a way that helps generate and choose appropriate coordination mechanisms for them? One of the advantagesof the definition we have used for coordinationis that it suggestsa direction for addressing these questions. If coordination is defined as managing dependencies, then further progress should be possible byc h a r a c t e r ~different i~ kindsofdependencies andi d e n ~ ~ ithe n gcoordination processes that canbe used to manage them.

Table 1.1 suggests the beginningsof such an analysis (see Malone, Crowston, Lee, & Pentland, 1992, for more details). For example, one possible kind

0G ONEAND CRO~STON Examples of C o ~ o n

TAB= 1.1 ~ ~ n d eBetween n c i e sActivities and Alternative~ o ~ i nProcesses ~ o n for ~ a n ~ Them ~ n g ~ ~ p lof eCoor~i~tion s Processes for ~ u ~ g i ~n ~ g pe~ency

Shared resources Task ~ s i ~ ~ n P ~ d u ~ r ~ o n s u relationships mer ~ ~ u i s i t e c o n s ~ n ~ Inventory

“Fmt codfirst serve,” priority order, budgets, managerial decision, ~ ~ t - l ibidding k e ~ (Sas for “’Shared resour~s’,)

Notification, sequencing, tracking Inventory ~ g e ~(e.g., n “Just t In Time”; “Ekonomic Order Q u ~ t i ~ ” ) Us~i~ity S ~ d ~ d i ~ t ask ion , p ~ i c i p a tdesign users, o~ Design for m ~ u f ~ t u r ~ i l i t y C o n c ~ n t e n g i n ~ ~ n g S i m u l ~ e ~ct y o n s ~ n ~ Sch~uling,s y n c ~ n i ~ o n T ~ k ~ u b t ~ k Goal selection, task decomposition

Nure. Inden~tionsin the left column indicate more specialized versionsof general d e ~ n ~ n c y types.

of dependency between different activities is that they require the same (limited) resources.The table shows that shared resource constraints can ed by a varietyof coordination processes such as “first come/first serve,” priority order, budgets, managerial decision, and market-like bidding. If three job shop workers need to use the same machine, for instance, they could usea simple first come/first serve mechanism. Alternatively, they could use a form of budgeting with each worker having preassi~nedtime slots, or amanager could explicitly decide what to do whenever two workers wanted to use the machine at the same time. In some cases, they might even wantto “bid” for use of the machine and the personwilling to pay the most would get it. The lists of dependencies and coordination processes in Table 1.1 are by no means intendedto be exhaustive.It is important to note, however, that many in particular kindsof systems (suchas “design for specific processes that arise manufacturability”) can be seen as instances of more generic processes (such as managing “usability” constraints between adjacent steps in a process). In fact, we believe that oneof the most intriguing possibilities for coordination theory is to identify and systematically analyzea wide variety of dependencies and their associated coordination processes. Such a “handbook”of coordination processes could not only facilitate interdisciplin~ytransfer of knowledgeaboutcoordination;itcouldalsoprovide a guideforanalyzingthe coordination needsin particular situations and generating alternative ways of fulfilling them (see Malone, Crowston, Lee, & Pentland, 1992). One question that arises immediately is how to categorize these dependencies and coordination processes. Table 1.1 provides a start in this direc-

1. INTERDISCIPLIN~YSTUDY OF C O ~ - ~ I N A T I O N

tion. Crowston (1991) suggesteda more structured taxonomy based on all the possible relationships between “tasks” and “resources.” To illustrate the possibilities for analyzing coordination processes, we discuss in the remainderof this section the coordination processes listed in Table 1.1 and how they have been analyzedin different disciplines. Whenever multiple activities share some limited resource (e.g., money, storage space, or an actor’stime), thena resource aZZ~ationprocess is needed to manage the interdependencies among these activities. Resource allocation is perhaps the most widely studied of all coordination processes.For in economics, organization theexample, it has received significant attention ory, and computer science. conomics. Much of economics is devoted to studying resource allocation processes, especially those involving market-like pricing andbidding mechanisms. As economists have observed, for instance, markets have a number of interestingpropertiesasresourceallocationmechanisms (Simon,1981).Foronething,theycanbe very ~ e c e n ~ a ZMany ~ e ~independent decision-makers interacting with each other locally can produce a globally coherent allocationof resources without any centralized controller a built-in set of incentives: (e.g., Smith,1’7’76).For another thing, markets have When all participantsin a perfect markettry to maximize their own individual benefits, the overall allocationof resources is (in a certain sense) globally “optimal” (e.g., Debreu, 1959).

nj~a~ion eo^. Organization theory has also paid great attention to resource allocation issues.For instance, controlof resources is intimately connected with personal and organizational power: those who control resources have power and viceversa (e.g., Pfeffer& Salancik, 1978). In general, organizationtheoristsemphasizehierarchicalresourceallocation methods where managers at each level decide how the resources they control will be allocated among the people who report to them (e.g., Burton& Obel, 1980a, 1980b). In practice, however, resource allocation in organizations ismuch more complex thana simple hierarchical model suggests. For instance,managersmay try toincreasetheirownpower by attracting resources (e.g., employees and money) away from other possible activities (Barnard, 1964) or by using their resources in a way that isvery suboptimal from the pointof view of the whole organization. Howcanwe choosebetweendifferentresourceallocationmethods? cost t ~ e addresses o ~ partof this questionby anaRecent workin ~ansffction lyzing the conditions under whicha hierarchy isa better wayof coordinat-

MALONE AND CROWSTON

ing multiple actors thana market (e.g.,~illiamson,1975, 1985). For example,

if there are extra costs associated with a market transaction (such as exten-

sive legal and accounting work), then the costs of internal transactions within a hierarchical firm may be lower and therefore preferable. A related question involves the conditions under which it is desirable to use market-like resource allocation mechanisms (such as transfer pricing) withina hierarchical organization (Eccles,1985).

ce. Resource allocation issues also arise in computer systems and much work has been done on these topics (e.g., Cytron, 1987; Halstead, 1985). For instan€e, operating systems require algorithms for allocating resources-such as processors and memory-to different processes and for scheduling accesses to input/output devices, such as disks (e.g., Deibeen esamplesof cross-fertel, 1983). As we see shortly, there have already tilization of ideas about resource allocation between computer science and other fields. For example, in Section 2.3.3, we see how ideas about distributed computer systems helped understand the evolutionof human organizations, and in Section 3.4, we see how analogies with human markets have generated novel resource allocation schemes for computer systems.

~~~~. One very important special case of resource allocation is task assignment, that is, allocating the scarce time of actors to the tasks theywill perform. An insight of the approach weare taking here, therefore, is that all the resource allocation methods listed in Table 1.1 are potentially applicable for task assignment, too. in a human For instance,in trying to imagine new coordination processes organization, one might consider whether any given situation requiring task assignmentcouldbebettermanaged by managerialdecision, by prior assignment accordingto task type, or by a pricing mechanism. To illustrate the surprising ideas this might lead to, consider Turoff’s (1983) suggestion that employees within a large organization should be able to “bid”for the wish to work, and that teams could be selectinternal projects on which they ed using these bids. Thereare obviouslymany factors to consider in determining whether such an arrangement would be desirable in a particular situation,butitisinterestingtonotethatonepotentialdisadvantage-the significantly greater communication required-would be much less important in a world with extensive computer networks. Another extremely common kind of relationship between activities is a ‘~roducer/consumer~ relationship, that is, a situation where one activity produces something that is used by another activity. This relationship clear-

1, INTE~ISCIPLINARYSTUDY OF ~ O O ~ I N A T I O N

1y occurs in all physical manufacturing processes, for instance, where the

output of one step on an assembly line is the input to the next.It also occurs with information whenever one personin an organization uses information from anotheror when one partof a computer program uses information produced by another. Producer/consumer relationships often lead to several kindsof dependencies: s ~ r ~ A~ very ~ ~common s . dependency between a "pro-

a onsu sum er" activity that is the producer activity must be completed before the consumer activity can begin. When this dependency exists, theremust at least be somen o ~ ~ c f fprocess ~ o n to indicate to the consumeractivitythatitcanbegin.Forinstance,whenanautomobile designer delivers a completed drawing of a part to the engineer who will design the manufacturing process for that part, the arrival of the drawingin the engineer's in-box "notifies" the engineer that her activity can begin. Managing prerequisite dependencies also often involves explicit se~~encing and ~ f f cprocesses ~ i ~ to be sure that producer activities have been completed before their results are needed. For instance, techni~uesfrom operations research, such as PERT charts and critical path methods, are muloften usedin human organizationsto help schedule large projects with tiple activities and complex prerequisite structures. These and other project tracking systemsare also often usedby managers to identify activities that are late and then use their authority to "motivate" the people responsible for the late tasks. What alternatives can we imagine for managing this dependency? One possibility would be computer-based tracking systems that made it easy for everyone in the project to see status information about all the other activities andtheirdependencies. In thiscase,latetaskscouldbevisible to everyone throughout the project, and '~uthoritarian" motivation by managers might become less important. ~equencingproblems arise frequentlyin computer systems, as well. For instance,one of thekeyissues in takadvantage of parallel ~rocessing in parallel and computers is determining which activi can be done ,ones must wait for the completion of others (Arvind &C Culler, 1986 ;Peterson, 1977,1981). Some of these ideas from computer science have also been used to help streamline processes in human or~aniz ing advantage of their latent parallelism (e.g., Ellis, Gibbons, &C du

~~tisfying a rer requisite constraint means that the pro~ucer too late. But in some cases, the producer activity might also be to too early (e.g.,if the item producedis perishable or there is no placestore hen it is important for an item to be neither too late nor too early, we

I

MALONE AND CROWSTON

call this an “inventory dependency.” One way ofmanagingan inventory dependencyis to carefully control the timing of both activitiesso that itemsare s ttime” to be used. This technique, for example, is becoming delivered ‘ ~ f f in increasinglycommon in manufacturingenvironments(Schonberger, 1982, 1986). A more common approach is maintain an inuento~of finished items, ready for the second activity to use, as a buffer between the two activities. Operations researchers,forinstance,havedevelopedtechniquesfor at what stock levels andby how much to replenish an inventoryin order to minimize costs {e.g., the “economic order~ f f f f ~ t iMcClain, ~,’; Thomas, & Mazola, 1992). Managing this dependency is also importantin certain parts of computer science. For example, in parallel processing systems, the rateof execution of processes must sometimes be regulatedto ensure that the produc1986).et al., er does not overwhelm the consumer or vice versa (e.g., Arvind As our framework would suggest, a common approachto this problem isto place a buffer between the two processes and allocate space in the buffer to one process or the other. Network protocols manage similar problems between communicatin~ processes that do not share any memory (Tannenbaum, 1981). i l i ~ ~Another, . somewhat less obvious, dependency that must often be managed in a producer/consumer relationship is that whatever is produced should be usable by the activity that receivesOne it. common way of creating uniformly intermanaging this dependency is by stffndffrd~fftion, changeable outputsin a form that users already expect. This is the approach on assembly lines, for example. Another approachtoisask users what characteristics they want. For instance, in human organizations this might be done by market research techniques such as surveys and focus (Kingroups near & Taylor, 1991). A third, related, alternative is pffrticipffto~ design,that is, having the users of a product actively participate in its design (Schuler& Namioka, 1992). This is a widely advocated approach to designing computer systems, for example, and it is interesting to note that the increasingly common practice of “concurrent engineering” (Carter & Baker, 1991) can also be viewed asa kind of participatory design. In concurrent engineering, people who design a product do not simply hand the design “over the transom” to those who design its manufa~turing process. Instead, they work together concurrently to create designs that can be manufactured more easily. In computer systems, the usability dependency occurs whenever one part of a system must use information producedby another. In general, this dependencyismanaged by designingvariouskinds of interchangelangua~esand other standards.

1, I N T E ~ I S C I P L I N ~STUDY Y OF COO~INATION

Another commonkind of dependency between activities is that they need to occur at the same time (or cannot occur at the same time). Whenever peo ple schedule meetings, for instance, they must satisfy this constraint. Another exampleof this constraint occursin the designof computer systems in which multiple processes (i.e.? instruction streams) can be executed simultaneously. (These systems may have multiple processors or a single processor which is shared between the processes.) In general, the instructions of the different processes can be executed in any order. Permitting this indeterminancy improves the performanceof the system (e.g., one process can be executed while another waits for data to be input) but can cause problems when the processes must share data or resources. System designers must therefore providemechanisms that restrict the possible orderings of the instructions by synchronizing the processes (i.e., ensuring that particular instructions from different streams are executed at the same time; Dubois, Schuerich,& Briggs, 1988). ~ynchronizationprimitives can be used to control sharing of data between a producer and consumer process to ensure that all data is used exactly once (the~ r ~ u c e r / c o n s uproblem) ~er or to prevent simultaneous ~ ~ ~cZusion ~ a problem). Z For example,if writes to a shared data item (the two processes simultaneously read and then update the same data (adding a deposit to an account balance, say), one processoverwrite might the value stored by the other. One example of interdisciplinary transfer involving this concept is the work of Singh and colleaguesin using computer science concepts about synchronized interactions to model process in human organizations (Singh & Rein, 1992).

e c o m ~ o s i t i o ~A. common kind of dependency among activities is thata group of activitiesare all “subtasks” for achieving some overall goal. h discussed in more detail shortly, there is a sense in which some overall ev~uationcriteria or “goals” are necessarily implied by the definition of coordination. The most commonly analyzed case of managing this dependency occurs when an individualor group decides to pursue a goal, and then will achieve decomposes this goal into activities (or subgoals) which together the original goal. In this case, we call the process of choosing the goal~ ~ Z ~ e Z e c on, and the processof choosing the activitiesgod ~ e c o ~ ~ ~ i ~ o n . For example, the strategic planning process in human organizations is often viewed as involving this kind of goal selection and goal decomposition

MALONE AND CROWSTON

process. Furthermore, an important role for all managers in a tradi~ionally conceived hierarchy is to decompose the goals they are given into tasks that are, in genthey can,in turn, delegate to people who work for them. There eral, many waysa given goal can be broken into pieces, and a long-standing topic in organization theory involves analyzing different possible decompositions such as by function, by product, by customer, and by geographical region (Mint~berg,1979). Someof these different goal decompositions for human organizationsare analogous to ways computer systems can be structured (e.g., Malone& Smith, 1988). In computer systems, we usually think of the goals as being predetermined, but an important problem involves how to break these goals into activities that can be performed separately. In a sense, for example, the essence of all computer programming is decomposing goals into elementary activities. For instance, programming techniques such as subroutine calls, modular programming, object oriented programming, so and forth can all be thou~htof as techniques for structuring the processof goal ~ecomposition (Liskov & Guttag, 1986).In these cases the goal decomposition is performed by a human programmer. Another exampleof goal decomposition in computer systemsis provided by work on planning in artificial intelligence (e.g., Allen, Hendler, & Tate, 1990; Chapman, 1987; Fikes & Nilsson, 1971). In this case, goalsare decomposed by a pl~nningprogram into a sequence of elementary activities, based on knowledge of the elementary activities available, their prerequisites, and their effects. In some cases, techniques for goal decomposition used in computer systems may suggest new ways of structuring human organizations. For examoses (1990) suggested that human organizations might sometimes be off not as strict hierarchiesbut as multilayered structuresin which any actor at one level could direct the activities of any actor at the next levelThismultilayered str~ctureis analogous to successiv la of s or"virtualmachines" in a computersystem (see 1990).

Eventhoughthemostcommonlyanae ase~uentialprocess of goal selection steps do not necessarily happenin this order. noth her possibility, for instance, is that several actors realize that the thin~sthey are already doing(withsmalladditions)couldwork together to achieve a new goal. For example, the creation of a new interdisciplinary research groupmay have this character. In human systems, this bottom-~pprocess of goal selection can often engender more comof responmitme~t ~rom the actors involved than a top-down assignment sibility. Q

1. INTE~ISCIP~INARY STUDY OF COO~INATION

As noted before, the dependencies discussed so far are only a suggestive list of common dependencies. We believethere are many more dependencies to be identified and analyzed. For instance, when two divisionsa company of a ~ o ~ both deal with the same customer, there isa shared r e ~ ~ t dependency between their activities: What one division does affects the customer’s perception of the company asa whole, including the other division.As another example, when several people in the same office want to buy a new rug,a key problem is not howto alZ~atethe rug, but what color or othercharac~eris~~s it should have.We might call this,therefore, ashared c h a r ~ c ~ e r ~depends~cs ency (another special case of the shared resource dependency). In general, there may be many ways of describing different dependencies, coordination processes, and their relationships to each other, and we suspect that there are many opportunities for further useful work along these lines. sic

Table 1.2 loosely summarizes our discussion so far by listing examplesof how common coordination processes have been analyzedin different disciplines. The key point of this table, and indeed of much of our discussion,is thattheconcepts of coordinationtheorycanhelpidentifysimilarities among concepts and results in different disciplines. These similar~ties,in turn, suggest how ideas can be transported back and forth across disciplinaryboundariesandwhere op~ortunitiesexist to develop even deeper analyses.

So far, the examples we have described have mostly involved a single fie1 or analogies that have been transported from one discipline to another. To illustrate the possibilities for developing abstract theoriesof coordination S stems, let us conthat can apply simu~taneously to many different ofkinds (Maartz, d in more detailel sew her^ alone, 1992), these analyses illustrate the kind of interdisciplinary interaction that our search coor~ination for theory encourages:The models grew Qri~inally out of designing distributed computer systems, they results from operations research, and they led eventually to new insi~hts about the evolutionof human organizations.

MALONE AND CROWSTON

TABLE 1.2 Examples of How Different Disciplines Have Analyzed Coordination processes

Coordi~tion Process ~ a n a ~ shared ng resources (including task assignments)

techniques for processor scheduling and memory allocation

analyses of markets and other resource allocation m e c ~ i s ~ , scheduling algorithms and other opti~~on tec~iques

analyses of different org~i~tional structures; budgeting processes, organizational power, and resource dependence

~ ~ a producer~ n g data flow and Petri consumer ~ l a t i o n s ~ p snet analyses (including prenquisites and us~ilityc o n s ~ n ~ )

PEBT charts, critical path methods; scheduling techniques

P~cipatorydesign; marketresearch

~ ~ a s i~m unl ~ge i ~synchronization constraints techniques, mutual exclusion

scheduling techniques

meeting scheduling, certain kinds of process modeling

~ a n a ~ task-subtask ng ~lations~p

economics of scale and scope

strategic pianning; ~ a g e m e n by t objectives;m e t h ~ s of grouping people into units

modul~zation techniques. in Prog~ng; pl~inm g artificial intelligence

Consider the following task assignment problem:A system is producing a set of “products,” each of which requires a set of “tasks”to be performed. The tasks are of various types, and each type of task can only be performed by “server” actors specialized for thatkind of task. Furthermore, the specific tasks to be performed cannot be predicted in advance; they onlybecomeknownduring the course of the process and then only to actors we will call “clients.” This descriptionof the task assi lern is certainly not universally applicable, but it an is abstract description thatcanbeapplied to manycommon taskassignmentsituations.For instance, the tasksmight be (a> designing, manufacturing, and marketing different kinds of automobiles, or(b) processing stepsin different jobs on a c o m ~ u ~network. er

1

IN~~ISCIPLINARY STUDY OF COO~INATION

I

One (highly centralized) possibility for solving this task assignment problem is for all the clients and servers tosend all their information toa central decision maker who decides which servers will perform which tasks and then notifies them accordingly, Another (highly decentralized) possibility is suggested by the competitivebidding scheme for computer networks formalized by SmithandDavis(1981). In thisscheme, a clientfirstbroadcastsan ffnnounce~entmessage toallpotential servers. Thismessageincludes a description of the activity to be performed and the qualifications required. The potentialservers then use this information to decide whether to submit a bid on the action.If they decide tobid, their bid message includes a descrip tion of their qualifications and their availability for performing the action. The client uses these bid messages to decide which server should perform the server thatis selected. activity and then sends award an message to notify the Malone and Smith (Malone, 1987; Malone & Smith, 1988) analyzed several alternative coordination mechanisms like these, eachof which is analogous to a mechanism used in human organizations.In particular, they developed ets and decenformal models to represent various forms of ~ f f r ~(centralized (based on products or functions). tralized) and various formsof ~jerffrc~jes Then they used techniques from queuing theory and probability theory to analyze tradeoffs among these structures in terms of ~ r ~ u c costs, ~ o n coordij l j ~For instance, they showed that the cenn f f ~ ocosts, n and ~ ~ l n e r f f 6costs. tralized schemes had lower coordination costs, but were more vulnerable to processor failures. Decentralized markets, on the other hand, were much less vulnerable to processor failures but hadhigh coordination costs.And decentralized hierarchies (“product hierarchies”) had low coordination costs, but they had unused processor capacity which led high to production costs.

Even though these models omit many important aspectsof human organizations and computer systems they help illuminate a surprisingly wide range of phenomena. For instance, as Malone and Smith (1988) showed, the models are consistent witha number of previous theories about human organization1973; March &Simon,1958; ~illiamson, al design (e.g., Galbraith, major historical changesin the org~izationalforms of both hu tions handler, 1962, 1977) and computer systems. These models also help analyze design alternatives for distributed scheduling mechanisms in computer systems, and they suggest ways of analyzing the structural changes associated with introducing new information technology into organizations (see Section 3.2 of this chapter; row st on, Malone, & Lin, 1987; Malone & Smith, 1

MALONE AND CROWSTON

In addition to the processes already described for managing specific depend-

encies, two other processes deserve specific attention: group ~ e c i s i o~ ~ f f ~ i and c o ~ ~ ~ n i cItf isf sometimes ~~n. possible to analyze these processes as ways of managing specific dependencies. For instance, communication can be viewed as a way of m~agingproducer/consumer relationships for inform^" tion. However, because of the importance of these two processes in almost all instances of coordination, we describe them separately here. any coordination processes require making decisions that affect the activities of a group. For instance, in sharing resources a group must somehow ecide” how to allocate the resources; in managing taswsubtask dependencies, a group must decide howto segment tasks.In all these cases, the alternative ways of makinggroupdecisionsgiverisetoalternativecoordination in principle,bemade by processes. For example, any group decision can, authority (e.g.,a m ~ a g edecides), r by voting, orby consensus resulti in^ from negotation). in coordination, Becauseof theimportance ofgroupdecisionmaking answers to questions about group decision making (e.g., Arrow, 1951; Simon, 1976) will be important for developing coordination theory. For instance, what are the decision-ma~ng biases in groups (e.g., Janis & Mann, 1977)as opposed to individuals ~ h n e m a n& Tversky, 1973)? How do computer-based group decision-making tools affect these processes (e.g., Dennis, JJoey, maker, & Vogel, 1988; Kiesler, Siegel,& ~ c ~ u i 1984; r e , Kraemer Can we determine optimal ways of allocating tasks and shari for making group decisions (Miao,Luh, Kleinman, 1992)? How do (or should) decision-making processes changein situations where both rapid response & Halpern, 1994). and high reliability are required (Roberts, Stout,

ecision making, there is already a great dealof theory about comboth froma te~hnicalpoint of view (e.g., Shannon &Weaver, 1949) of view (e.g., Allen, 197’7; Roger from an organizat~onal point ers, 1976 Weick, 1969). One obvious way of generating new processes, for example, is by considering alternative forms of ~ommunication ronous vs. as~chronous,paper vs. electronic) for all the placesin a S where information needsto be transferred. ination framework also highlights new aspects of these problems. le, when we view communication as a way of managing pro er/cons~~ relationships er for information, we may be concerned about how

1. I N T ~ ~ I S C I P L I N A RSTUDY Y OF C O O ~ I ~ A T ~ O N

to make the information “usable.” How, for instance, can actors establish common langua~esthat allow them to communicatein the first place? This question of developing standards for communication ofiscrucial concernin designing computer networksin general (Rertouzos,1991) and cooperative work tools in particular(e Lee &C alone, 1988).Theprocess by which standards are developed i also of concerntoeconomists,philosophers, and others (e.g., Farrell& Saloner, 1985; Hirsch, 1987). A related set of questions arises when we are concerned about how a group of actors cancome to have ‘~ommonknowledge,” that is, they all know something, andthey also all know that they all know it. There is a growing literature about this and related questions in fields as diverse as com&C Levesque, puterscience,economics,andsties(Aumann,1976;Cohen 1991; ray, 1978; Halpern, 1987; m, 1981; Shoham, 1993).

Any scientific theory (indeed, any statement about the world) must neglect Kling (1980) some things,in order to focus on others. For example,

how different perspectives (such as rational, structural, and po~itical) on the use of info~mation systems in orga~izations each illuminate aspectsof reality of these perspecneglected by the others.In some situations? one or another tives may be most important, and all of them are involved to some degreein any real~ituation, In a p p l ~ coordination ~g theory to any particular system, therefore,it may be necessaryto consider many other factors as well. For instance, in designing a new computer system to help people coordinate their work, “details” about screen layout and response time may sornetimes be as important as the basic functionality of the system, and the repuin a particular organ~zation tation of the manager who introduces the system may have more effect on the motivation of people to use itin t tion than any incent~ve structuresdesi~nedinto the system. d~signing distributed a processi ferent kinds of commu~ications bemay the p r ~ m a r ~ erations about how

There are at least two ways an ~nter~iscipli~ary theory can he1 ifferencesliketheseamon systems~~ a r a ~ ea~~ ia cl ~ s ias ,

u~aiysis.

MALONE AND CRO~STON

~alysis. In parametricanalysis,theabstracttheories include parameters which may be different for different kinds of systems. For instance, the principles of aerodynamics apply to both birds and airplanes, even though parameters such as size, weight, and energy expenditure are very different in the two kinds of systems. Similarly, abstract models of coordinationmayincludeparametersforthingslikeincentives, cognitive capacities, and communication costs which are very different in human,computational,andbiologicalsystems.Ekamples of modelsthat have been applied to more than one kind of system in this way are summarized in Sections 2.3 and 3.4.2. ~alysis. In baseline analysis, one theory is used as a baseline for comparisonto the actual behaviorof a system, and deviations from the baseline are then explained with other theories. For example,in behavioral decision theory (e.g., Kahneman & Tversky, 1973), mathematical decision theory is used to analyze the ways people actually make decisions. In the cases where people depart from the prescriptionsof the normative mathematical theory, new theoriesare developed to explain the differences. Even though the original mathematical theory does not completely explain peoby the newtheories could not ple’s actual behavior, the anomalies explained even have been recognized withouta baseline theory for comparison. This suggests that an important part of coordination theory will be ~ e ~ a v i o r a l c~r~inatio t n~ e inowhich ~ careful observations of actual coordinationin human systems are used to develop, test,and augment abstract modelsof coordination.

In order to analyze a situationin terms of coordination, it is sometimes important to explicitly identify the components of coordination in that situanag aging ation. According to our earlier definition, coordination means dependenciesbetweenactivities.”Therefore,becauseactivitiesmust, in somesense,beperformed by “actors,”thedefinitionimpliesthatall instances of coordination include actors performing activities that are interoften ~ e ~ e n It~is ealso n ~ ~ useful to identify eval~ationcriteria for judging how well the dependenciesare being ‘~anaged.” For example, we can often entify some overall “goals” of the activity (such as producing automobiles or printing a report) and other dimensions for evaluating how well those oals are being met (such as minimizing time or costs). Some coordination ‘SeeBaligh(1986);BalighandBurton(1981);Barnard(1964);Malone(1982’);Maloneand ;McCrath (1984); and Mintzberg (1979) for related decompositions of coordination.

1. INTE~ISCIPLINARYSTUDY OF C ~ O ~ I N A T I O N

processes may be faster or more accurate than others, for instance, and the costs of more coordinationare by no means always worthwhile. It is important to realize that there is no single “right” way to identify these componentsof coordination in a situation. For instance, we may somedivision as one times analyze everything that happens in a ma~ufacturing “activity,” while at other times, we may want to analyze each station on an assembly line as a separate activity. As another example, when we talk about muscular coordination, we implicitly regard different partsof the same person’s body as separate “actors” performing separate “activities.” is. One important case of identifying evaluation criteria occurs when there are conflicting goals in a situation. In analyzin nation in human organizations, it is often useful to simply ask people what their goals are and evaluate their behavior in terms of these criteria.However, some amount of god con~ictis nearly always present (e.g., Ciborra, 1987; Schelling, 1960; ~illiamson,1985), and people may be unableor unwilling to accurately report their goals, anyway. To understand these situations, it is often usefulto try to identify the conflicting goals and also to analyze the behavior of thesystem in terms of someoverallevaluationcriteria. For instance,~ifferentgroups in a company may compete for resources pee and ple, but thisvery competition may contribute to the company’s overall abil1981). ity to produce useful products (e.g., Kidder, Another imp~rtantexample of confli~tinggoals occursin mark~ttran§actions: As we saw earlier, all participantsin a market might have the goal of m ~ i m i z i n gtheir own individual benefits, but we,observers, as can evaluate the market as a coordination mechanism in terms of how well it satisfies g utilities (e.g., Debreu, 1959) or overall criteriasuch as m ~ m i z i n consumer “fairly” distributing economic resources. ,

In the remainder of this section,we describe examples of how concepts about coordination have been applied in three different areas: (1) understanding the new possibilities for humanor~anizations and markets provi ed by information techno lo^, (2) designing cooperative work too designing distributed and parallel computer systems. The early use very general notionsof coordination; the later onesare more c com~onentsof coordination. theiridentification of S This list is not inten to be a comprehensive list of all ways that theories of coordination CO e applied. In fact, most of the work we describe here did not explicit~yuse the term c ~ ~ ~ i n ~ t i o n~e have t ~ echosen o~. examples, however,to illustrate the wide rangeof applications for inter ciplinary theories about coordination.

M ~ O N AND E CROWSTON

~anagers,organ~zation theorists, and others have long been interested in how the widespread use of information technology (IT) may change the ways human organizations and markets will be structured (e.g., Leavitt & ~hisler, 1958; Simon, 1976). One of the mostimpo~antcontributionsof coordination theory may be to help understand these pQssibilities better. To illustrate how the explicit study of coordination might help with this endeavor, we begin witha very general argument that does not depend on so far in this chapany of the detailed analyses of cQordination we have seen ter? Instead, this argumentstarts with the simple observation that coordination is itself an activity that has costs. Even though there are many other forces that may affect the way coordination is performed in organizations and markets (e.g., global competition, national culture, government regulation, and interest rates), one important factoris clearly its cost, and that is the focusof this argument.In particular, it seems quite plausible to assume that information technology is likely to significantly reduce the costs of certain kinds of coordination (e.g., Crawford,1982). Now, using some elementary ideas from microeconomics about substitution and elasticityof demand, we can make some simple predictions about the possible effects of reducing coordinatiQn costs.It is useful to illustrate these effectsby analogy with similar changesin the costsof trans~ortation induced by the introductionof trains and automobi~es: 1.

first order” effectof reducing transportation costs with trains and omobiles was simply some substitution of the new transportation ide on trains more and in sportationcostswasto nd convenientlyin trains

Ies of new strucconvenient tr~ns~ortation. no

expect several effects from using newinfor~ationtechcostshe of coQrdination:

3See Malone (1992) and Malone and Rockart (1991) for more detailed ment in this section.

versions of the ar

1. INTE~ISCIPL~NARY STUDY OF COO~INATION

1. A “first order” effect of reducing coordination costs with information technologymaybe tosubstituteinformationtechnologyforsomehuman coordination. For instance, many banks and insurance companies have substituted automated systems for large numbersof human clerks in their back offices. It has also long been commonplace to predict that computers will lead to the demiseof middle management because the communication tasks perby computformed by middle managers could be performed less expensively ers (e.g., Leavitt& Whisler, 19%). This prediction was not fulfilled for several decades after it was made, but many people believe that it finally began to ha pen with large numbers of middle management layoffsin the 1980sand 1990s. 2. A “secondorder”effect of reducingcoordinationcostsmaybeto increase the overall amount of coordination used.In some cases, this may in one case we studied, a comoverwhelm the first order effect. For instance, puter conferencing system was usedto help removea layer of middle managers (see Crowston et al., 1987). Several years later, however, almost the same number of new positions (for different people at the same grade level) had been created for staff specialistsin the corporate staff group, manyof whom were helping to develop new computer systems. One interpretation of this outcomeis that the managerial resources no longer needed for simple communication tasks could now be applied to more complex analysis tasks that would not previously have been undertaken. 3. A “third order” effectof reducing coordination costs may be to encourage a shift toward the use of more ‘~oordination- intensive^'structures.In other will now words, coordination structures that were previously too “expensive” become more feasible and desirable. For example, as noted before, i n f ~ r ~ a ~ gTof, tion techno lo^ can facilitate what someobservers (e.g.,~ i n t z b ~1979; fler, 1970) have calledf f ~ ~ ~ r f f Adhocracies ci~~. are veryflexible or including many shifting project teams and highly decentralized communication am on^ relatively autonomous entrepreneurial groups. Oneof the disadvanta~esof adhocraciesisthattheyrequirelarge amo~ntsof unp~annedcommunicationandcoordination throu~houtan o However, tec~nologies such as electronic mail and computer confe communication, and alone, Grant, Turbak, help make this communication more effective at much larger scales.

any specificde~ende~cies. Instead, it compares two nat~onmechanisms that can manage many such de

MALONE AND CROWSTON

transactions vs. internal decision making with firms, and(2) centralized vs. decentralized managerial decisions. alone, Yates, and Benjamin (1987) have used ideas from transact~on cost theory to systematically analyze how information technologywill affect firm size and, more generally, the use of markets as a coordination structure. They conclude thatby reducing the costs of coordination, information techy lead to an overall shift toward smaller firms and proportionatee of markets-rather than internal decisions within firms-to coordinate economic activity. This t has two parts. First, because market transactions often dinationcoststhaninternalcoordination(Malone,‘Yates, & have hi enj jam in, 1987; ~illiamson,1985), an overall reductionin th in situations coor~inationshould lead to markets becoming more desirable where internal transactions were previously favored. This, in turn, should lead to less vertical integration and smaller firms. r example, after the introduction of computerize airline reservation ms, the proportionof reservations made through avel agents (rather than by calling the airline directly) went from 35% to 70%. Thus, the function ~eservationswas ‘~isintegrated” from the airlines and movedto a firm-the travel agents. Preliminary econometric analyses of the overall U.S. economy in the period 1975-19~are also consistent with these predictions: The use of information technology appears to be correlated h firm size and vertical integration ~ r ~ j Q l f s s o Man, If we extrapolate this trend toa possible long-run extreme, it leads us to speculate that we might see increasing use of “firms” containing only one person. Forinstance,MaloneandRockart (1991) suggest that there may in which somedaybeelectronicmarketplaces of “intelmercenaries~ it is possibleto electronica~lyassemble“overrmies” of thousands of articularproblemand peoplewhoworkfor a fewhours ordaysto S n disband. Flexible arrangements like this might appeal especially to pead a strong desire for autonomy-the freedom to choose their own orking situations.

ani and hang (1991) have used ideas from agency the tematically analyze the effects on centrali~ation of the reduction enabled by IT. They conclude thatIT can lead to eit ecentralization,dependingonhoitisused. ~ l t h o ~ gthis h cQnclusion may not be surprising, the~tructureof their analysis helps us

1. INTE~IS~IPLIN~Y STUDY OF ~OO~INATION

understand the factors involved more clearly: (1) When IT primarily reduces decision infor~utioncosts, it leads to more centralization. For instance, the Otis elevator company used IT to centralize the reporting and dispatching functions of their customer service system, instead ofhaving these functions distributed to numerous remote field offices (Stoddard, 1986).(2) On the other hand, when IT primarily reduces agency costs, it leads to more decentralization.As used here, agency costsare the costsof employees not acting in the interestsof the firm. For instance, when one insurance company developed a system that more effectively monitored their salespeople's overall performance, they were able to decentralize to the salespeople many of the decisions that had previously been made centrally & (Bruns Mc~arlan, 1987). Overall, this bidirectional trend forIT and centralization is consistent with empirical studiesof this question (Attewell& Rule, 1984). An alternative approach to this question is provided by (Danziger, Dutton, KIing, & Kraemer, 1982). In a sense, thiswork can be considereda kind of "behavioral coordination theory," In studies of computerization decisions in 42 local governmentsin the UnitedStates, theyfound that changesin centralization of power were not best explained any of the formal factors one might have expected. Instead, they found that because people who already have power influence computerization decisions, the newofuses computers tend to reinforce the existing power structure, increasing theofpower those who already have it. There has recently beena great dealof interest in designing computer toolsto help people work together more effectively (e.g., Ellis,& Gibbs, Rein, 1991; Greif, 1988; Johansen, 1988; Peterson, 1986; Tatar, 1988, 1990; additional references in Table 1.3). Using terms such asco~puter-suppo~ed c~perutive work and ~ o u ~ wure these systems perform functions such as helping people collaborate on writing the same document, managing projects, keeping track of tasks, find-and ing, sorting,and prioritizing electronic messages. Other systems in this categoin face-tory help people display and manipulate information more effectively face meetings and represent and share the rationales for group decisions. In this section, we describe how ideas about coordination have been he1 ful in suggesting new systems,classi~ingsystems, and analyzing how these systems are used.

One way of generating new design ideas for cooperative worktools is to look toother ~isciplinesthatea1withcoordination.Forinstance,even did not explicitly use the term c ~ r d i n ~ tti~oe~though the following authors

MALONE AND CROWSTON

WY, they each used ideas about coordination from other disciplines to help

develop cooperative work tools.

for analyzing group action based heavily on ideas from linguistics (e.g., Searle, 1975). This perspective emphasizes different kindsof speech acts, such asrequestsandcommitments.Forexample, ~inogradand Flores in terms of the possible states analyzed a generic “conversation for action” and transitions involvedwhen one actor performs a task at the request of another. An actor may respond to a request, for instance, by promising to fulfil^ the request, declining the request, reporting that the request has already been completed, or simply acknowledging that the request has been received.The analysis of this conversation type (and several others) ed a primary basis for designing the Coordinator, a computer-based cooperative work tool. For example, the Coordinator helps people make and keep track of requests and commitments to each other. It thus suports what we might call the “mutual agreeing” part of the Cask a s ~ ~ ~ ~r~~ess.

tool for routing i~formationwithin organizations. In the “blackboard archite~ture~~’ program modules interact by searching a global blackboard for theirinputsandpostingtheiroutputsonthesameblackboard ~rman, th, Lesser,& Reddy, 1980; Nii, 1986) his provides very flexible patterns ofcommunicationbetweendifferentogrammodules: Any module can communicate with any other module, e n when this interaction is not e~plicitlyanticipated by the program designer. In adhocracies, as we saw ear lie^, just this kind of unplanned, highly decentralized communication is essential for rapidly responding to new situations(~intzberg,1979; Toffler, bla ated, in part, by this need for an ‘~rganizational designedtheInformationLenssystem Cohen, 1987).A central componentof this systemis an ‘ ~ n y oserver” ~ e thatletsspecifyrulesaboutwhatkinds of electronicmessagesthey are int in seeing.Thesystemthenusesthese in th rules to route all non-private electronic messages to everyone . . might want to see them. sages, another partof the oritize the messages people receive.)

~

I. I N T E ~ I S C I P L I N ~STUDY Y OF COO~INATION

Two cooperative work tools,gIBIS (Conklin & Begeman, 1988) and Sibyl (Lee, 1990) are designed to help groups of people make decisions more effectively. To do this, they explicitly represent the arguments (and counterarguments)diffor ferent alternatives a group might choose. Both these systems are based on ideas from philosophy and rhetoric about the logical structure of decision making. For example, the basic elements in the gIBIS system (issues, positions, and arguments) are taken from a philosophical analysis of argumentation by Ftittel and Kunz (1970). The constructs for representing arguments in Sibyl are based on the work of philosophers like Toulmin (1958) and Rescher (1977). Holt (1988) describeda theoretical language used for designing coordi~ationtools that is based,in part, on ideas about Petri nets,a formalism used in computerscience to representprocessflows in distributed or parallelsystems (Peterson, 198,1977). This language is part of a larger theoretical framework called ‘~oordination mechanics” and has been used to design a ‘~oordination en~ronment” to help people work together on computer networks. les. Clearly,usingideasaboutcoordinationfrom other disciplines provide any guarantee of developing useful cooperative work tools. However, we feel that considering these examples within the (1) it sugcommon framework of coordinationtheory provides two benefits: gests that no oneof these perspectivesis the completestory, and (2) it suggests that we should look to previous work in various disciplines for more insights about coordination that could lead to new cooperative work tools.

01s

As shown in Table 1.3, the framework we have suggested for coordination provides a natural way of classifying existing cooperative work systems according to the coordination processes they support. Some of these systemsprimarilyemphasizeasinglecoordination-relatedprocess. For instance, electronic mail systems primarily support the message transport part of communication, and meeting schedulingtools primarily support the synchronization process (i.e., arranging for several people to attend a meeting at the same time). There is a sense,of course, in which each of these systems also support other processes (e.g., a simple electronic mail system can be used to assign tasks), but we have categorized the systems here according to the processes they explicitly emphasize. Some of the systems also explicitly support several processes. For example, the Information Lens system supports both the communication routing

MALONE AND CROWSTON

TAB= 1.3 A Taxonomy of Cooperative Work Tools Based on the Processes They Support

~ r n System ~ ~ e

Process ana aging shared resources (task assignment

Coordinator ( W i n o g ~& Flares, 1986) Info~ationLens alone, Grant, Turbak, Brobst, & Cohen, 1987)

~ a n a ~ producer~onsumer ng relationships (sequencing p ~ r ~ u i s i t e s )

Polymer (Croft & kfkowitz, 1988)

Mana~ngsimultaneity consmints (synchronizing)

Meeting sc~duling tools (e.g., Beard et al., 1990)

and prioriti~tion)

a n a ~ n gtask-subt~k la ti on ship (goal decomposition) Group decisionm ~ i n g

gIBIS (Conklin & ~ e g e1988) ~ , Sibyl (h, 1990) electronic meetingrooms (e. 1988; DeSanctis & Gallup, 1987)

Electronic mail, Computerconfe~ncing(e.g.,

Lotus, 1989)

electronic meeting rooms (e.g., Denniset al., 1988; DeSanctis & Gallup, 1987;Stefic et al., 1987)

I n f o ~ ~ Lens o n (Malone, Grant, Twbak, Brobst, & Cohen, 1987) collaborative authoring tools (e.g., Ellis et al., 1990;Fish, b u t , Leland, Neuwi~h,Kaufer, Chmdh

process (by rules that distribute messagesto interested people) anda form of resource allocation process(by helping people prioritize their own activities using rules that sort messages they receive).And the Polymer system helps people decompose goals into tasks and sequence the tasks (e.g., to prepare a monthlyreport, firstgather the project reports and then write a summary paragraph). One possibility raisedby this framework is that it might help identify new opportunitiesforcooperativeworktools.Forinstance,theCoordinator focusesonsupportingonepart of the task assignment process (mutual a~reementon commitments). However, it does not provide much help for the earlier part of the process involving selecting an actor to perform the task in the first place (see Section 2.3). New tools, such as an “electronic yellow pages” or bidding schemes like those suggested by Turoff (1983) and alone ( 1 9 8 might ~ be useful for this purpose. Another intriguing possibility suggested by this framework is that it might be possible to implement “primitives”a number for of different coordination-

1. INTE~ISCIPLINARYSTUDY OF COO~INATION

related processes in the same environment, and then let people combine these primitivesin various ways to help solve particular coordination problems. This is one of the goals of the Oval system (Lai, Malone, &Yu, 19 lone, Lai,& Fry, 1992).

Another use for coordination theory in designing cooperative work tools can be to help systematica~ly evaluate proposed or actual systems. For example, Markus and ~onnolly(1990) systematically analyzed how the payoffs to individual users ofa cooperative work system depend on how many pee other ple are using the system. They did this by usinganeconomicmodelfrom ) about the incentives to use Schelling (1978) to extend ~rudin’s( 1 9 ~insights cooperative work systems. For instance, online calendars and many other cooperative work applications involve “discretionary databases” which users can view or updateas they see fit.For each individual user, however, the benefits of viewing the database can be obtained without contributing an~hing. Thus, it is often in the interests of each individual user to use the database without making the effort required to contribute to~nfortunately, it. the equilibrium state of a system like this is for no one to ever contribute an~hing! An interesting empirical illustration of this phenomenon occuredin a study of how one large consulting firm used the Lotus Notes group conferencing system. In this study, Orlikowski(1992) found that there were surprising inconsistencies between the intended uses of the system and the actual incentives in the organization. For instance, Orlikows~observed that this organization (like many others) was one in which people were rewarded for being the “expert” on something-for knowing things that others did not. Should we be surprised, therefore, that many people were reluctant to spend much effort putting the things they knew into a database where everyone else could easily see them? These observations do not, of course, mean that conferencing systems like in organizations. What they do mean, however, is that this one cannot be useful we must sometimes be sensitiveto verysubtle issues about things like incenin order to obtain the full benefits of such systives and organizational culture tems. For instance, it might be desirable in this organization to include, as part of an employee’s performance appraisal,a recordof how often their contributions to the Notes database were usedby other peoplein the organization.

a varietyof Much recent activityin computer science has involved exploring distributed and parallel processing computer architectures.In many ways, physically connecting the processors to each other is easy compared to the

MALONE AND CROWSTON

difficulty of coordinating the activities of many differentprocessors working on different aspectsof the same problem. In this section, we describe examples of work that have addressed these issues in an explicitly interdisciplinary way, drawing on insights from other disciplines or kinds of systems to design or analyze distributed or parallel computer systems. In particular, we consider examples of analogies with social and biological systems as a source of design ideas, and quantitative tools for analyzing alternative designs.

is

i ~ i v ~ of basic One the systems n de to assign tasks to processors, and several distributed computer systems have addressedthisproblemwithcompetitive bidding mechanismsbasedon analogies with human markets. For example, the Contract Nets protocol (Davis & Smith, 1983; Smith & Davis, 1981) formalizes a sequence of messages to be exchanged by computer processors sharing tasks in a network.The“conby tracts” are arbitrary computational tasks that can potentially be performed any of a number of processors on the network, the “clients”are machines at which these tasks originate, and the “contractors” are machines that might process the tasks (i.e., the servers). The sequence of announcement,bid, and award messages usedby this protocol was already described in our analysis of the task assignment process (Section2.3). One of the desirable featuresof this system is its great degree of decentralization and the flexibility it provides for how both clients and contractors can make their decisions. For instance, clients may select contractors on the basisof estimated completion time or the presenceof specialized data; contractors may select tasksbidtoon based on the size of the task or how long the task has been waiting. Using these or similar ideas, a number of other bidding systems have been developed (e.g., Kurose& Simha, 1989; Stankovic, 1985). For instance, several bidding systems have been developed to allow personal workstations connectedby a local area network to share tasks (Malone, Fikes, Grant, & Howard, 1988; Waldspurger, Hogg, Huberman, Kephart, & Stornetta, 1988). In this way, users can take advantageof the unused processing capacityat idle workstations elsewhere on the network. Furthermore, the local bidding ‘~egotiations” can result in globally coherent processor scheduling according to various priorities (e.g., Malone et al., 1988). (For a review of several relatedsystemsandananalysis of a variety of bidding algorithms, see Drexler & Miller, 1988; Miller & Drexler, 1988.) The notion of competitive bidding markets has also been su gested as a technique for storage management by Miller and Drexler (19

1. INTE~~SCIPLINARY STUDY OF COO~INATION

ille er, 1988). In their proposal, when object A wishes to maintain a pointer to of the spacein which object object B,object A pays “rent” to the “landlord” €3 is stored. These rentsare determined by competitive bidding, and when an object fails to pay rent, it is “evicted” (thatis, garbage collected). Their proposal includes various schemes for how to determine rents, howto pass rents alonga chain of references, and howto keep trackof the various costs and payments without excessive overhead. They conclude that this proposal is not likely to be practical for small scale storage management (such as garbage collectionof individual Lisp cells), but that it may well be useful for sharing large objects in complex networks that cross “trust boundaries” (e.g.,interorganizationalnetworks).Theschemealsoappearsusefulfor managing local caching and the migration of objects between different forms of short-term and long-term storage. i Another ~ ~ central . problem that arises in distributed and parallel processing systems is how and when to route information between processors. in artificial intelFor instance, one interesting example of this problem arises ligence programs that search a large space of possibilities, the nature of which is not well known in advance. It is particularly useful,in this case, for processors to exchange information about intermediate results in such a way that each processor can avoid performing work that is rendered unnecessary by work already done elsewhere. One solution to this problem is suggested by the Scientific ~ommunity ~etaphorembodied in the Ether system (Kornfeld,1982; Kornfeld & Hewitt, 1981). In this system, there are a numberof “sprites,” each analogous to an individual scientist, that operate in parallel and interact through a global database. Each sprite requires certain conditions to be true in the global database before it is “triggered.” When a spriteis triggered, it may compute new results that are added to the global database, create new sprites that await conditions that will trigger them,or stifle a collection of sprites whose work is now known to be unnecessary. In one example use of this system, Kornfel~(1982) showed how sharing intermediate results in this way can dramatically improve the time performance of an algorithm (evenif it isexe He calls this effect‘~ombinatorial cuted by timesharing a single processor). implosion.” This system also uses the scientific community metaphor to suggest a solution to the resource allocation problem for processors. Each sprite is “supported”by a “sponsor,” and without a sponsor, a spritewill not receive any processing timeto do its work. For instance,a sponsor may sometimes support both work directed toward proving some proposition and also work directed toward proving the negationof the proposition. Whenever oneof these linesof work is successful, support is withdrawn from the other.

~

MALONE AND GROWSTON

Another wayof applying coordination concepts isto help evaluate alternative designsof distributed and parallel processing computers stems. For instance, Huberman and his colleagues~uberman& Hogg, 19 Huberman, 1990) applied mathematical techniques like those usedin chaos theory to analyze thedynamic behavior of distributed computer networks. In one case they analyze, for example, heavily loaded processorsin a network transfer tasks to more lightly loaded processors according to a probabilistic process. When any processor in such a system can exchange tasks with any other processor, the behaviorof the system is unstable for large numbers of processors (e.g., more than 21 processors in a typical example). However, when theprocessors are grouped hierarchically into clusters that exchange tasks frequently among themselves and only occasionally with other clusters, the system remains stable forar~itrarilylarge numbers of processors, Thishierarchicalarrangementhasthedisadvantagethatit takes a longtime to reach stability. In anintriguinganalogywithhuman organizations, however, Huberman and his colleagues find that this disadvantage can be eliminated by having a few “lateral links” between different clusters in the hierarchy (Lumer& Huberman, 1990). As summarized in Table 1.4, the examples we have described show how a coordination perspective can help analyze alternative designs, and suggest new design ideas.In each case, these applications depended upon interdisciplinary useof theories or concepts about coordination. *

~e have seen howa number of different disciplines can contribute to answering the questions about coordination, and how theories of coordination can, in turn, be applied to the concerns of several different disciplines. What is needed to further develop this interdisciplinary study of coordination? As we suggested earlier, a central concernof coordination theory should be identi~ngand analyzingspecificcoordinationprocessesandstructures. Therefore, a critical item on the agenda for coordination research should be developing these analyses. For example, the followingofkinds questions arise. Howcanwe represent coordination processes? When should we use flowcharts, Petri nets, or state transition diagrams? Are there other notations that are even more perspicuous for analyzing coordination?How can we classify

1. INTE~ISCIPLINARYSTUDY OF COO~INATION

TABLE 1.4 Sample Applic~onsof a Coordination Perspective

Applic~tionArea

~ ~ lof Ae~ l ysz i n g A l t e ~ tDesigns i~~

~ p l e ofs~enerating New Design Idem

C~ting ~mpor~ the effects of O r g a n i ~ t i o ~ l s tund ~ c ~ r eAnalyzing s “intellectu~~ k e t p l a c e s ”to decreasing coordination costs infor~tion tec~ology on firm size, c e n ~ i ~ i o n , solve specific problems and internal structure Designing new tools fortask a s s i ~ ~ ninfo~ation t, routing, and group decision making;

Cooperative work tools

An~yzinghow the payoffs to individual users of a system depend on the number of other users

Dist~butedand p~rallel computer s y s t e ~

Using com~titivebiddin Analyzing stability properties of load sharing a l g o in~ ~ mechanisms to allocate processors and m e ~ ino ~ computer networks computer systems. Using a scientific unity metaphor to organize parallel problem solving

different coordination processes? For instance, can we usefully regard some coordination processes as “special cases” of others? How are different coordination processes combined when activitiesare actually performed? Anotherset of questionshastodowithhowgenericcoordination processes really are:How far can we get by analyzing very general coordination processes, and when will we find that most of the important factors are specific to coordinatinga particularkind of task? For example,are there general heuristics for coordination that are analogous to the general pro in cognitive science and artificial intelligence? lem-solving heuristics studied At least as important as these general questions are analyses of specific processes. For example, how far can we goin analyzing alternative coordination processes for problems such as resource allocation? Can we characterize an entire “design space” for so~utions to this problem and analyze the major factors that would favor one solutionover another in specific situations? Could we do the same thingother for processes such as goal selection or managing timing dependencies? Are there other processes (such as manin aging other kinds of dependencies) that could be analyzed systematically ways that have not yet been done?

MAtONE AND CROWSTON

In analyzing alternatives processes for specific problems, we might considin terms er various kindsof properties: Which processes are least “expensive” of production costs and coordination costs? Which processes are fastest? Which processes are most stable in the face of failures of actors or delaysof information? Which processes are most susceptible to incentive problems? For instance, how does the presence of significant conflicts of interest among actors affect the desirability of different resource allocation methods? How do information processing limitations of actors affect the desirability of different methods? For example, are some methods appropriate for coordinating people that would not be appropriate for coordinating computer processors, and viceversa?Whatnewmethodsforcoordinatingpeoplebecomedesirable when human information processing capacities are augmented by computers?

A critical part of the research agenda for thisarea is developing coordinationtheory in thecontext of variousdifferentkinds of systems. For instance, in Section 3, we suggested numerous examplesof these possibilities for human organizations and computer systems. In some cases, this work may involve applying previously developed theIn many cases, however, we expect that new ories to these application areas. systems or new observations of these systems will stimulate the developmentofnew theories. Forexample,allofthefollowingmethodologies appear likelyto be useful in developing coordinationtheory: (a) empirically studying coordinationin human or other biological systems (e.g., field studies, laboratory studies, or econometric studies), (b) designing new technologies for supporting human coordination, (c) designing and experimenting with new methods for coordinating distributed and parallel processing computer systems, and(d) formal modeling of coordinationprocesses (e.g., mathematical modelingor computer simulation).

Clearly, the questions we have just Iistedare only the beginning of a set of research issuesin the interdisciplinary studyof coordination. However, we believe they illustrate how the notion of coordination provides a set of abstractions that help unify questions previously considered separately in a variety of different disciplines and suggests avenues for further exploration. lthough much work remains to be done, it appears that this approach can build upon much previous work in these different disciplines to help solve a varietyof immediate practical needs, including: (1) designing computer and communication tools that enable people to work together more effectively,(2) harnessing the powerof multiple computer processors work-

1. INTERDISCIPLINARY

STUDY OF COO~INATION

ing simultaneously on related problems, and (3) creating more flexible and

more satisfying waysof organizing collective human activity.

This work was supported, in part, by Digital Equipment Corporation, the Natio~lScience Foundation (Grant Nos. I R 1 ~ 0 ~ ?and 9 8 IRI~9030~), and other sponsorsof the MIT Center for Coordination Science. Parts of this chapter were included in three previous papers (Malone, 1988; Malone & Crowston, 1990; Malone& Crowston, 1991).We are especially grateful to Deborah Ancona, John Carroll, Michael Cohen, Randall Davis, Rob Kling7 John Little, and Wanda Orlikowski for comments on earlier versions of the paper, andto participants in numerous seminars and workshops at which these ideas have been presented.

“The operationof complex systems madeup of components.” (NSF-IRIS, 1989) “Theemergentbehavior of collections of individualswhoseactions are based on complex decision processes.” (NSF-IRIS, 1989) ‘‘Information processing withina system of communicating entities with distinct information states.”(NSF-IRIS, 1989) “The joint efforts of independent communicating actors towards mutually defined goals.” (NSF-IRIS, 1989) “Networksof human action and commitments that are enabled by computer and communications technologies.” (NSF-IRIS, 1989) “Composing purposeful actions into larger purposeful wholes.” (A. Holt, personal communication, 1989) “Activities required to maintain consistency within a work product or to manage dependencies within the workflow.” (Curtis, 1989) “Theintergrationandharmoniousadjustment of individualwork efforts towards the accomplishmentof a larger goal.” (Singh, 1992) “The additional information processing performed whenmultip~e, connected actors pursue goals that a single actor pursuing the same goals would not perform.’’ (Malone, 1988) “The actof working together.” (Malone& Crowston, 1991)

M ~ O N AND E CRO~STON

Even though use of the term c~r~infftion t ~ e iso quite ~ recent,a great deal of previous work in various fields can contribute to the interdisciplinary understanding of coordination. In this appendix, we briefly describe examples of such work from several different disciplines. These examples focus on cases where coordination has been analyzedin ways that appear to be generalizable beyond a single discipline or type of actor. We have not, of course, attemptedto list all such cases; we have merely triedto pick illustrative examples from several disciplines. s o ~ r c ~ sMuch . research in computer science focuses on how to manage activities that share resources, such as processors, memory, and access to input/output devices (e.g., Deitel,1983). Other mechanisms have been developed to enforce resource allocations. For example, semaphores, monitors, and critical regions for mutual exclusion are programming constructs that can be used to grant a process exclusive accessto a resource (e.g., ~ijkstra,1968; Hoare,1975).Researchers in databasesystemshave developed numerous other mechanisms, suchas locking or timestamping, to allow multiple processes to concurrently access shared data without interference (e.g., Bernstein& Goodman, 1981).

ors. In addition,protocolshavebeendeveloped to ensure the reliability of transactions comprising multiple reads or writes on different processors (e.g., Kohler,1981). In particular, these protocols ensure that either all a transaction’s operations are performed or none are, even if some of the processors fail.

One of the important problems in allocating work to processors is how to divideup the tasks. For example, Carriero & Gelernter (1989) discuss three alternative waysof dividin~parallel programs into units: according to the type of work to be done, accordingto the su~partsof the final output, or simply according to which processor is available. ws. Another important set of issues involves ion. For instance,researchers in artificial intelrticularly in distributed artificial intelligence(DAI; e.g., Bond Huhns & Gasser, 1989) have used “blackboard architectures” to allow processes to share information without having to know precisely

1. INTE~ISCIPLINARYSTUDY OF COO~INATION

which other processes need it (Ermanet al., 1 9 8 Nii, ~ 1986), and ”partial global plans” to allow actors to recognize when they need to exchange more information (Durfee& Lesser, 1987).

In a sense, almost allof economics involvesthe study of coordination, with a special focus on how incentives and information flows affect the allocation of resources among actors. For example, classical micr~cono~ics analyzes how differentsources of supply and demand can interact locally in a market in waysthat result in a globally coherent allocation ofresources. ~ m o n the g major results of this theory are formal proofsthat (under appropriate mathematical conditions)if consumers eachmaximize their individual ‘~tilities~ and firms each maximize their individual profits, then the resulting allocain the sense that no one’s utilities tion of resources will be globally “optimal” can be increased without decreasing someone else’s (e.g., Debreu, 1959). Some more recent work in economics has focused on the limitations of cost markets andcontracts for allocatingresources. For instance, ~ff~saction t ~ e analyzes o ~ the conditions under which a hierarchy is a better way of coor~inating m~ltjple actors than a market (e.g., ~illiamson,1975). Agency t ~ efocuses o ~ on howto create incentives for some actors (agents) to act in a way that advances the interests of other actors (principals) even when the principals cannot observe everything their agents are doing (Ross, 1973). One result of this theory is that there are some situations where no incentives can motivate an agent to perform optimally fromthe princi~al’s point of view (Jensen& Mec~ling,1976). Finally, some parts of economics focus explicitly on information flows. For o ~its descendants analyze how information should example, team t ~ e and be exchanged when multiple actors need to make interdependent decisions butwhenallagentshave the same ultimategoals(e.g.,Hurwicz, 1973; m t ~ e also o ~anaMarschak & Radner, 1972; Reiter, 1986). ~ e c ~ a n i sdesign lyzes how to provide incentives for actors to reveal information they possess, even when they have conflicting goats.For example, this theory has been appliedto designing and analyzing various forms of auctions. In a “second price auction,” for instance, each participant submits a sealed bid, and the highest bidderis only requiredto pay the amount of the second highest bid. It can be shown that this mechanism motivates the bidders to each reveal the true value they place onthe item beingsold, rather than trying to “game the system” by bidding only enoughto surpass what they expect to be the next highest bid (Myerson,1981). Operations research analyzes the properties of various coordination mechanisms, butoperations research also includesa special focus on developing optimal techniques for coordination decisions. For instance, opera-

W O N E AND CROWSTON

tions research includes analyses of various scheduling and ~ueueingpolicies andtechniques such as linear programming and dynamic programming for makingresource allocation decisions optimally (e.g., Dantzig, 1963).

Research in organization theory, drawing on disciplinessuch as sociology and psycho lo^, focuses on how peoplecoordinate their activities in formal organizations. A central theme in this work has involved analyzing general issues about coordination (e.g., ~albraith,1977; Lawrence & Lorsch, 1967; Simon, 1958 Simon,1976;Thompson, 1967; summarized by Malone, Mintzberg, 1979). We can looselyparaphrase the key ideas of this work as follows: All activities that involve more than one actor require (1) some way of dividin~activities among the different actors and (2) some way of managing the interdependencies between the different activities (Lawrence & Lorsch, 1967; March & Simon, 1958). Interdependencies between activities can be of (at least) three kinds: pooled, where the activities share or prol, duce common resources but are otherwise independent; s e ~ u e ~ t i awhere some activities depend on the completion of others before beg inn in^, and reciprocal, where each activity requires inputs from the other~ h o m p s o n , 1967). These different kinds of interdependencies can be managed by a where prevariety of coordination mechanisms, such as: standard~utio~, determined rules govern the performance of each activity; direct supervision, where one actor manages interdependencies on a case-by-casebasis, and ~ u u ~ u~ s t ~~ewhere ~ t l, each actor makes on-going adjustments to manage the interde~endencies (Galbraith, 1973;March & Simon,1958; Mintzberg, 1979). These coordination mechanisms can be used to manage interdependencies, not only between individual activities,also but between groups of activities. One criterion forgroup in^ activities into unitsis to minimize the difficulties of managing these intergroup interdependencies. For example, activities with the strongest interdependencies are often grouped into the smallest units,then these units are grouped into larger units with other units with which they have weaker interdependencies. Various c~mbinationsof the coordination mechanisms, together with different kinds of grouping, ive rise to thedifferent organizationalstructures common in human organizations, including functional hierarchies, product hierarchies, and matrix organizations. For instance, sometimes all activitiesof the same type (e.g., manufacturing) might be grouped together in order to take advanta economies of scale; at other times, ail activities forthe same product (e.g., marketing, manufacturing, and engineering) might be groupedtogether to simplify managingthe interdependencies betweenthe activities.

l . INTE~ISCIPLINARYSTUDYOF COO~INATION

iolo

Many parts of biology involve studying how different partsof living entities interact. For instance, human physiology can be viewed as a study ofhow the activitiesof different partsof a human body are coordinated in order to keep a person alive and healthy. Other parts of biology involve studying how different living things interact with each other. For instance, ecology can be viewed as the study of how the activitiesof different plants and animalsare coordinated to maintain a “healthy” environment. Someof the most intriguing studies of biological coordination involve coordination between different animals in a group. For example angel and Clark (1988) discusses the optimal hunting pack size for lions, who trade the benefitof an increased chanceof catching something against the cost of having to share what they catch. Deneubourg and Gross (1989) point out that the interaction between simple rules-such as‘‘dowhat my neighbour is doing”-and the environment may leadto a variety of collective behaviors. The most striking examples of such group behaviors are in social insects, such as honey bees or army ants, where the group displays often quite complex behavior, despite the simplicity of the individuals (e.g., Franks, 1989; Seeley, 1989). Usinga varietyof simple rules, these insects “allocate” individual workers at economically efficient levels to a variety of tasks-including searching for new food sources, gathering nectar or pollen from particular sources (bees), carrying individual food items back to the bivouac (ants), guarding the hive (bees) and regulating the group temperature. For example, in honey bees, the interaction of two simple local rules controls the global allocationof food collectors to particular food sources. First, nectar storing bees unload nectar from foraging bees returning to the hive a rate at that depends on the richnessof the nectar. Second,if bees are unloaded rapidly, they recruit other bees to their food source. The result of these two rulesis that more bees collect food from better sources. Seeley (1989) speculates thatthisdecentralizedcontrolmayoccurbecauseitprovidesfaster responses to local stresses (Miller, 1978),or it may be simply because bees have not evolved any more global means of communication.

Allen, J.,Hendler, J. and Tate,A. (Eds.). (1990). Readings in Planning.San Mateo, CA Morgan Kaufmann, Allen, T. J.(1977). ~ a n atheg flow ~ ~of Technology.Cambridge, MA MIT Press. Arrow, K. J.(1951). Social Choice and individual Value (Vol. 12). New York Wiley. Arvind and Culler, D. E. (1986). Dataflowarchitectures.Annual Reuiewsin C o ~ ~ u ~ e ~ S (Vol. c~ence 1). Palo Alto, CA Annual Reviews, Inc.

MALONE AND CROWSTON Arvind, Nikhil,R. S. and Pingali, K.K.(1986). I-Structures: Datastructures for parallel computing. In P r ~ e e d i ~ sthe oGraph f ReductjonWor~hop. Santa Fe, NM, October. Attewell, P. and Rule, J. (1984). Computingand organizations: What we know and whatwe don’t know. Communicationsof the ACM, 27,1184-1192. Aumann, R. 3. (1976). Agreeing to disagree. Anna~s Statjs~cs, of 4, 1236-1239. Baligh, H. H. (1986). Decision rules and transactions, organizations and markets.Man~ement Science, 32, 1480-1491, Baligh, H.H.and Burton, R. M. (1981). Describing and designing organizational structures and processes. InternationalJ o u ~ a l Policy of Analysis and Information Systems, 5,251-266. . M A Allyn & Bacon. Baligh, H. H. and Richartz,L. (1967). Vertical MarketS ~ c t u r e sBoston, Barnard, C. I. (1964). The unctions of the Executive.Cambridge, MA: Harvard University. Beard, D., Murugappan, P., Humm, A., Banks, D.,Nair, A. and Shan,Y. -P. (1990). A Visual Calendar for Scheduling Group Meetings. In D.Tatar (Ed.), P r ~ e e d iof ~ the s Third Conference on Co~puter-supportedCooperative Work(pp. 279-2~).Los Angeles,C A ACM Press. Bernstein, P. and Goodman, N. (1981). Concurrency control in distributed database systems. ACM Computi~ Surveys (June). Bond, A. H. and Gasser,L. (Eds.). (1988). R e a d i ~ in s D~tr~buted ArtificialIntell~ence, San Mateo, C A Morgan Kaufman. Bruns, W. J, and McFarlan,F. W. (1987). Information technology puts powerin control systems. ~arvardBusiness Review(Sept,-Oct.), 89-94. Brynjolfsson, E., Malone, T., Gurbaxani, J. and Kambil, A. (1989). Does Information Technology Lead to Smaller Firms?(Technical report 106). Center for Coordination Science, NIT. Burton, R. M. and Obel,B. (198Oa). A computer simulationtest of the M-form hypothesis. Administrative Science Quarterly, 25,457-466. Burton, R. M. & Obel, B. (198Ob). The efficiency ofthe price, budget, and mixed approachesunder Management Science,26. varying a priori information levels for decentralized planning. Carriero, N. and Gelernter,D,(1989). How to write parallel programs:A guide to theperplexed. ACM Computi~ Surveys, 21(3), 323-357. Carter, D. E. and Baker, B. S. (1991). Concurrent ~ ~ i n e e rThe i ~Product : Development Environment for the 19901s.Reading, MA:Addison-Wesley, in the History ofthe American Industrial Chandler, A. D., Jr. (1962).Strategy andS~cture: Chapte~ ~nterprise.Cambridge, M A MIT Press. Chandler, A. D. (1977). Admin~trativecoordination,all~ationand monitori~:Concepts and com~arisonsork king paper 77-21). Brussels: European Institute for Advanced Studiesfn Management. 32,333-377. Chapman, D. (1987). Planning for conjunctive goals.A~ificial Int~ll~ence, Ciborra, C. U. (1987). Reframing the role of computers in organizations: The transaction costs approach. Office:techno lo^ and People,3,17-38. Cohen, P. and Levesque, H. J. (1991). Team~ork(Technote No. 504). Menlo Park, C A SRI International. Conklin, J. and Begeman, M. L. (1988). gIBIS: A hypertext toolingfor exploratory policy discussion. In ~ e e d i n g of s GCW ‘88 Conference on Computer~upported CooperativeWork (pp. 140-152). Portland, OR ACM Press. Crawford, A. B., Jr. (1982). Corporate electronicmail-A communication-intensive applicationof information technology.MIS Quarterly,6(September), 1-13. Croft, W. B. and Lefkowitz, L.S. (1988). Using a plannerto supportoffice work. InR. B.Allen (Ed.), Proceedings of the ACM Conference on O&ce Information Systems (pp, 55-62). Palo Alto,CA. Crowston, K. (1991). Towards U C~rdination C~kbook: Recipes for~ u l t i - ~ eAction. nt Ph,D. Dissertation, MIT SIoan Schoolof Management, Cambridge,MA. Crowston, K, Malone, T. W and Lin, F. (1987). Cognitive science and organizational design: A case studyof computer conferencing.~ u m a Computer n Interaction,3,59-85.

1. I~E~ISCIPLINARY STUDYOF COO~INATION

Curtis, B. (1989). Modeling coordination from field experiments. In O ~ a n ~ a t ~ o n a l C o m p u t i ~ J CoordinationandCollaboration: meories andTechnologies For Computer-Suppo~edWork. Austin, TX. Cytron, R (1987). Limited processor scheduling of doacross loops. In Proceedings of the I987 International ConFerence on Parallel Processing (pp. 226-234) Pennsylvania State University, University Park, Pennsylvania. Dantzig, G. B. (1963). Linear P r ~ a m m and i ~ Extensions. Princeton, NJ: Princeton University Press. Danziger, J, N., Dutton, W. H., Kling, R. and Kraemer, K L. (1982). Computers and Politics: ~~h techno lo^ in American Local Governments.New York: Columbia UniversityPress. ArtiDavis, R. and Smith,R. G. (1983). Negotiationas a metaphor for distributed problem solving. ficial Intell~ence,20,63-109. Debreu, G. (1959). Theory oFva1ue:An~ i o m a t i analysis c oFeconomic equilibrium.New York Wiley. Deitel, H.M. (1983). An I n ~ ~ u c t to ~ oopera n ti^ Systems. Reading, MA:Addison-Wesley. Deneubourg, J.L. and Gross,S. (1989). Collective patterns and decision-making.E t h o l ~EcoloJ g y & Evolution, I, 295-31 1. Dennis, A. R., Joey, E G,, Jessup, L. M., Nunamaker, J.E and Vogel, D. R. (1988). Information Technology to Support Electronic Meetings.MIS ~ u a ~ t eZ24(December), r~, 591-619. Dertouzos, M.L. (1991). Building the Information Marketplace. Technology Review (January), 29-40, DeSanctis, G. and Gallupe,R.(1987). A foundation forthe study of group decision support systems. Management Science,33(5),589-609. Dijkstra, E. W, (1968). The structureof the T. H. E, operating system.Communicat~onsof the ACM, 11(5), 341-346. Drexler, K E. and Miller, M. S. (1988). Incentive engineering for computational resource management. In B. A. Huberman (Ed.), The Ecology OF Computation (pp. 231-266). Amsterdam: Elsevier. Dubois, M,, Scheurich, C. and Briggs,F. A. (1988). S~chronization,coherence, and event ordering inmultiprocessors. IEEE Computer, 21(2), 9-21. Durfee, E, W; and Lesser,V. R. (1987). Using partial global plans to coordinate distributed problem sobers. In ~ o c e e d i of ~ sthe TenthInte~ationalJoint ConFerence onArtificialIntell~ence ~ I J ~ (pp. I ~875-883). 7 ~ Eccles, R. G. (1985). The TransferPricing Problem:A Theory forPractice.Lexington, MA: Lexin~on Books. Ellis, C. A., Gibbons, R. and Morris, P. (1979). Office ~treamlining.In Proceedi~sOFthe Internatio~alWorksho~on rnte~ratedOffices.Versailles, France. Institut de Recherche d’Informatique de d’Automatique. Ellis, C. A., Gibbs, S. J. and Rein, G. L.(1990). Designand useof a groupeditor. In G. Cockton (Ed.), E ~ i n e e rFior~~uman&omputerrnteraction (pp. 13-25) North Holland/Elsevier, Amsterdam. Erman, L. D., Hayes-Roth, E., Lesser, V, R. and Reddy, D. R. (1980). The HEARSAY-I1 speech understanding system: Integrating knowledge to resolve uncertainty. ACM Com~utingSurveys, 12(2), 213-253. Farrell, J. and Saloner,G. (1985). Standardization, compatibiity, and innovation. R a n ~ ~ o uof ~al Economics, 16(Spring), 70-83. Fikes, R E. and Nilsson,N. J. (1971). STRIPSA new approach to the applicationof theorem proving to problem solving.Artifi~ialrntell~ence,2, 198-208. Fish, R, Kraut, R., Leland, M. and Cohen, M. (1988). Quilt: A collaborative tool for cooperative writing. In ~ o c e e d i oF ~ the s Conference on Oflice InFormation Systems (pp. 30-37). Palo Alto, CA. Flores, F., Graves, M., Hartfield, B, and Winograd,T. (1988). Computer systems andthe design of Systems, 6(2), 153-172, organizational interaction.ACM Transactions on OEce InFor~a~on

MALONE AND CROWSTON Fox, M. S. (1981). An organizational view of distributed systems.IEEE Transactions on Systems, Man and Cy6ernetics,11(1), 70-79, Franks, N. R. (1989). Army ants: A collective intelligence.American Scientist, 77(March-April), 139-145. Galbraith, J.R. (1973). Des~ningComplex O ~ a n ~ a t i o nReading, s. M A Addison-Wesley. Galbr~th,J.R. (1977). O ~ a n ~ a t iDesign. on Reading, M A Addison-Wesley. Gray, J.(1978). Notes on data base operating systems (Research ReportRJ 2188). IBM. Greif, I. (Ed.). (1988). Computer Supported Cooperative Work. Los Altos, C A Morgan Kaufmann. Grudin, J. (1988). Why CSCWapplications fail: Problems inthe design and evaluationof organizational interfaces. In D. Tatar (Ed.), Proceedings of the Second Conference onComputer-su~ ported CooperafiveWork (pp. 85-93). Portland, OR ACM Press. Gurbaxani, V. and Whang, S. (1991). The Impact of Information Systems on Organizations and Markets. Communicationsof the ACM, 34(1), 59-73, In AnnuHalpern, J.J,.(1987). Using reasoning about knowledgeto analyze distributed systems. al Review of Computer Science(Vol. 2;pp. 37-68). Palo Alto,C A Annual Reviews Inc. Halstead, R. H. (1985). Multilisp: A language for concurrent symbolic computation.AGM Transactions on ramming Languages and Systems,7,4,501-538. Hart, P. and Estrin,D. (1990). lnter~rganizationcomputer networks: Indications of shifts in interdependence, In Proceedings of the ACM Conference onOfficeInformation Systems.New York ACM. 4(3), 271-287. Hewitt, C. (1986). Offices are open systems.ACM Transactions on Omce Systems, Hirsch, E. D.(1987). CulfurafLiteracy: Whateve^ A~ericanNeeds to Know. Boston, M A Houghton Mifflin, Hoare, C. A. R. (1915). Monitors: An operating systems structuring concept.Co~mun~cations of the ACM, 17(10), 549-557. Corrigendum, CACM, 18(2): 95. Holt, A, W. (1980). Coordinator Programs~npublishedtechnical report).Wakefield, MA: Massachusetts Computer Associates, Inc. Holt, A. W. (1988). Diplans: A new language forthe study and implementationof coordination. ACM Transactions onOfficeInformation Systems,6(2), 109-125. Huberman, B. A. (Eds.). (1988a). The Ecology of Computatioff.Amsterdam: No~h-Hoiland. Huberman, B, A. (1988b). Open systems: The ecology of computation (PARC working paper P880~74).Palo Alto,C A Xerox PARC. Huberman, B. A. and Hogg, T. (1988). The behaviourof computational ecologies,In B. A. Huber(pp. 77-116). Amsterdam: Elsevier, man (Ed.), The Ecology of Compu~ation (Vol. 3). San Mateo,C A Huhns, M. N. and Gasser,L. (Eds.). (1989). Dis~ibutedArtificialrntell~ence Morgan ~ u f m a n n . Hurwicz, L. (1973). The design of resource allocation mechanisms.American Economic Review Papen and Proceedings,58(May), 1-30. Janis, I. L. and Mann, L. (1977). Dec~ion-making.New York Free Press. Johansen, R. (1988). group ware^ Co~puterSupportf o r ~ u s i f f eTeams. ~ s New York.The Free Press. Kahneman, D.and Tversky,A. (1973). On the psychologyof prediction. Psychol~icalReview, 80, 237-251. Kidder, T. (1981). The Soulof a New Machine.Boston: Little, Brown. Kiesler, S., Siegel, J. and McGuire, T. W. (1984). Social psychological aspects of computer-medi39,1123-1134. ated communication.American Psycholog~t, Kling, R. (1980).Social analyses of computing: Theoretical perspectives in recent empirical research. ACM Computing Surveys,12(l), 61-1 10. Kohler, W, (1981). A survey of techniques for s~chronizationand recovery in decentralized Surveys, 13(2), 149-183. computer systems.AGM Comp~ting Kornfeld, W. A, (1982).CombinatoriallyImplosiveAlgorithms. Communications of the ACM, 25( lo), 734-738.

1, INTE~ISCIPLIN~Y STUDY OF COO~INATION Kornfeld, W. A. and Hewitt, C, (1981). The scientific community metaphor. IEEE Transactions on Systems, Man and ~bernetics, SMC-I I, 24-33. Kraemer, K. and King, J. L. (1988). Computer-based systems for cooperative work and group decision-making.ACM C o m p u t i ~ Surveys, 20(2), 115-146. Kurose, J.E and Simha,R. (1989). A microeconomic approachto optimal resource allocationin distributed computer systems.IEEE Transaction on Computers,38(5), 705-717. Lai, K.Y., Malone,T. and Yu, K . 4 . (1988). Object Lens:A spreadsheet for cooperative work. ACM T~nsactionson OfficeInformation Systems, 6(4), 332-353. Lawrence, P. and Lorsch, J.(1967). O ~ a n ~ a t i oand n Environment. Boston, M A Division of Research, Harvard Business School. Leavitt, H. J, and Whisler, T. L. (1958). Management in the 1980's. Haruard Business Reuiew, 36(November~ecember),41-48. Lee, J. (1990). Sibyl: A qualitative decision management system. InP. Winston (Ed.), Artificial i ~ (Vol. 1). Cambridge,M A MIT Press. Intell~enceat M n ~ p a n dFrontiers Lee, J. and Malone, T. W. (1988).How can groups communicate when they use different languages? Translating between partially shared type hierarchies. In R. B. Allen (Ed.), Proceedings of the ACM Conference on Oflice information Systems(pp. 22-29). New York ACM. Liskov, B. and Guttag, J. (1986). Abs~actionandSpecificationin Program ~evelopment.Cambridge, M A MIT Press. Lotus. (1989). Lotus Notes Users Guide. cambridge, M A Lotus Development Corp. Lumer, E. and Huberman,B. A. (1990). Dynamics of resource allocationin d ~ ~ i b u tsystems ed (SSL9045). Palo Alto,C A Xerox PARC. Malone, T. W. (1987). Modelingcoordination in organizations and markets. Management Science, 33,1317-1332. n g #2051-$8). Cambridge,MA:MIT Malone, T. W. (1988). What is cmrdination t h e o r y ~ ~ o r k i paper Sloan Schoolof Management. Malone, T, W, (1990). Organizing information processing systems: Parallels between organizations and computer systems. In W. Zachary, S. Robertson andJ.Black (Eds.), Co~nition,Computation, and Cmperation (pp. 56-83). Norwood, NJ: Ablex. Malone,T. W. (1992). Analogies Between Human Organization and Artificial Intelligence Systems: ed ~erspectives Two Examples and Some Reflections.M.InMasuch (Ed.),~ i s ~ i b u tIntell~ence: ofArtificia1Intell~enceon O ~ a n ~ a t iand o n ~anagementTheory. Amsterdam: Elsevier. Malone, T. W. and Crowston,K. (1990). What is coordination theory andhow can it help design cooperative work systems? In D. Tatar (Ed.), Proceeding of the Third Conference on Computersupported Cmperative Work (pp. 357-370). Los Angeles, C A ACM Press. of coordination (TechMalone, T.W. and Crowston,K. G. (1991). Toward an interdisciplinary theory nical report #120). Cambridge, M A Massachusetts Institute of Technology, Center for Coordination Science. Malone, T. W., Crowston, K., Lee, J. and Pentland, B. (1992). Tmls for i n v e n t i ~ o ~ a n ~ a t i o n s ; onal ~npublishedWorking Paper). Cambridge, Toward a h a n d b ~ k o f o ~ a n ~ a t iprocesses M A MIT Center for Coordination Science. Malone, T. W., Fikes, R. E., Grant, K. R. and Howard, M. T. (1988). Enterprise: A market-like task scheduler for distributed computing. In B. Huberman (Ed.),The Ecology of Computafion. New York North"Hol1and. Malone, T. W., Grant, K. R., Turbak, F, A., Brobst, S. A. and Cohen, M.D. (1987). Intelligent information-sharing systems.Communications of the ACM, 30,390-402. .Xand Fry, C. (1992). ~perimentswith Oval:A radically tailorable toolfor cooperative work.In ~ ~ e e d i nof g rhe s ACM Conference on Computer~upportedC~perative Work (CSCW '$2).Toronto, Ontario. Malone, T. W. and Rockart, J, F. (1991). Computers, networks, and the corporation. Scientific Sept.), 128-136, A ~ e r i c a n265(3, ,

MALONE AND CROWSTON Malone, T. W. and Smith, S. A. (1988). Modeling the performance of organizational structures. Operations Research,36(3), 421-436. Malone, T. W., Yates, J.and Benjamin,R. I. (1987). Electronic markets and electronic hierarchies. CommunicationsoftheACM, 30,484-497. ~ Ecolo~. Princeton, NJ: PrinceMangel, M.and Clark, C. W, (1988). Dynamic M ~ e lini~ehavioral ton University Press. March, J.G. and Simon,H. A. (1958). O ~ a n ~ a t i o nNew s . York Wiley. Markus, M.L. and Connolly,T. (1990). Why CSCW applications fail: Problems inthe adoption of interdependent work tools. In D. Tatar (Ed.), ~ e e d i n gof s the Third Conferenceon Computer-supported Cooperatiue Work(pp. 371-380). Los Angeles,C A ACM Press. R. (1972). Economic Theoryof Teams.New Haven: Yale University Press. Marschak, J. and Radner, McClain, J., Thomas, L. J. and Mazola, J. (1992). Operations~ a n ~ e m e(3rd n t ed.). Englewood Cliffs, NJ: Prentice-Hall, McGrath, J,E. (1984). Groups: Interaction and Performance. Englewood Cliffs,NJ: Prentice-Hall. Miao, X., Luh, P.B. and Kleinman,D.L. (1992). A normativedescriptive approach hierarchical to team resource allocation. IEEE Tra~actionson Systems,Man and ~bernetics,22(3), 482-497. rom, P. (1981). An axiomatic characterizationof common knowledge, Econometr~cQ, 49(1),

17-26.

Miller, J,G. (1978). Living Systems. New York McGraw-Hill. Miller, M. S, and Drexler, K. E. (1988). Markets and computat~on:Agoric open systems. In B. A. Huberman (Ed.), The Ecology of ComputQtion(pp. 133-176). Amsterdam: North-Holland. Mintzberg, H. (1979). The S t r u c t ~ r i ~ o f O ~ a n ~Englewood a t i o n s . Cliffs,NJ Prentice-Hall. Moses, J. (1990). O ~ a n ~ a t i and o n Ideology ~npublishedmanuscript). Cambridge,MA:Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technolo-

gy.

Myerson, R.B. (1981). Optimal auction design.Mathemat~csof Operations Research,6,58-73. Neuwirth, C. M,, Kaufer, D, S., Chandhok, R. and Morris, J.H. (1990). Issues in the design of computer-suppo~for c~authoringand commenting. In ~ ~ e e dofithe ~ Third s Conference on ComputerSup~ortedCooperative Work (CSCW90) (pp. 183-195). Los Angeles,C A ACM Press. Nii, W,(1986). The blackboard modelof prob~emsolving. A Z M ~ ~ (Spring), ~ n e 38-53. NSE (1991). C~rdinationTheory and Collaboration T e c h n o l ~Workshop summa^. Washington, DC: Available fromNSF Forms & Publications Unit. report by theN S ~ I Review ~ S Panel for Research on Coordination Theory and shington, DC: Available from NSF Forms & Publications Unit. Orlikowski, W. (1992). Learning from Notes: Organizational Issues in Groupware Implementation. In ~ ~ e e dofithe~ Conference s on Computer SupportedC~perativeWork (CSCW 9.2). Toronto, Canada:ACM. Peterson, D. (1986). ~ ~ e e d i n of g sthe Conference on Computer-SupportedCooperativeWork. Austin, m ACM. Peterson, J. L. (1977). Petri nets.ACM Computi~ Surveys, 9(3), 223-252. Peterson, J. L.(1981). P e ~ ~ e t and ~theh~ e~ oe l~i n g ~ SEnglewood ~ t e m s .Cliff,NJ: Prentice-Hall. al Arlington Heights,I L Harlan Davidson. Pfeffer, J. (1978). Q ~ a n ~ a t i o nDesign. Pfeffer, J. and Salancik, G. R.(1978). The ~ ~ eControl ~ aof Ol ~ a n ~ a f ~ A o nResource s; ~ependency Perspective.New York Harper & Row. Reiter, S. (1986). Informational incentive and performancethe in(new)2 welfare economics. In S. Reiter (Ed.), Studies in ath he ma tical Economics (~tudiesin Mathematics, Volume25). Mathematical Assocationof America. Rescher, N. (1977). ~ialectics:A Con~oversy-Oriented A ~ p r ~ to c hthe Theory of Knowled~e.Buffalo, NX State University of New York Press. Rittel, H. and Kunz, W. (1970). Issues as elements of informations y s t e ~ s(Working paper 131). Stuttgart, Germany: Universityof Stuttgart, Institutfur Grundlagen der PlanungI.A.

1. I ~ E ~ I S C I P L I NSTUDY ~ Y OF COO~INATION Roberts, K. H. and Gargano,G. (1989). Managing a high reliability organization:A case forinterdependence. In M. A. V. Glinow and S. Mohrmon (Eds.), Managing c o m p l ~ini ~high technology industries: Systems and people (pp. 147-159). New York Oxford University Press. J.(1994). Decisiondynamics in two high reliability milRoberts, K. H., Stout,S. K and Halpern, J. itary organizations. Management Science,40,614-624. Rockart, J, F. and Short, J.E. (1989). IT and the networked organization: Toward more effective management of interdependence. In M. S. Scott Morton (Ed.), Man~ementin the 1990s Research ~ o g r a m Final Report.Cambrid e, MA:Massachusetts Institute of Technology. Rogers, E. and Agarwala-Rogers,R. (1976). Communication ino ~ a n ~ a t i o nNew s . York:Free Press. Ross, S. (1973). The economic theoryof agency. American EconomicReview, 63,134-139. Rumelhart, D. E., McClelland, J, L. and PDP Research Group.(1986). Parallel Distri6uted ProcessCambridge, MA:MIT Press. ing: ~plorationsin theMjcr~truc~res of C~nition. Schelling, T. C. (1960). Strategy of Conflict.Cambridge, MA: Harvard University Press. Schelling, T.,C. (1978). Micromotives and Macro6ehavior.New York Norton. Schonberger, R. (1982). Japanese M a n u f a c ~ rTechniques. i~ New York Free Press. New York Free Press. Schonberger, R. (1986). World ClassManufacturi~. Schuler, D. and Namioka,A. (Eds.), (1992). Participatory Design.Hillsdale, NJ: Lawrence Erlbaum Associates. e , and Searle, J.R. (1975). A taxonomy of illocutionary acts. In K. Gunderson (Ed.), L a ~ u ~Mind ~ n o w l e (pp. ~ e 344-369). Minneapolis: Universityof Minnesota. Seeley, T. D. (1989). The honey bee colony as a superorganism.American Scientist,77'ovemberDecember), 546-553. Shannon, C. E. and Weaver, W, (1949). The mathematical theoryof communication.Urbana: University of Illinois Press. Shoham, Y. (1993). Agent Oriented Programming.Arti~cialIntelligence, 60, 51-92. Simon, H. A, (1976). Administrative Behavior(3rd ed.). New York FreePress. Simon, H. A. (1981). Sciences of the Arti~ciai(2nd ed.). Cambridge,M A MIT Press. Singh, B. (1992). Interconnected Roles (IR>; A c~rdinationmodel (Technical ReportNo. CT-084-92). (MCC). Austin, Tx:Microelectronics and Computer Technology Corp. A process description formalism (TechSingh, B. and Rein,G. L. (1992). Role Interaction Nets (RIA'>: nical ReportNo. CT-083-92). Austin, 1x: MCC. Smith, A. (1776). The Wealthof Nations (1986 ed.). London: Penguin Books. Smith, R. G, and Davis, R, (1981). Frameworks for cooperation in distributed problem solving. IEEE Transactions on Systems, Man and ~ 6 e ~ e t i c11(1), s , 61-70. Stankovic, J.(1985). An applicaiton of Bayesian decision theoryto decentralized controlof job scheduling,IEEE Transactions on Computers, C-34(2), 117-130. Stefik, M., Foster, G., Bobrow, D. G., Kahn, K.,Lanning, S. and Suchman, L. (1987). Beyond the chalkboard: Computer support for collaboration and problem solving in meetings. C o ~ ~ u nications of the ACM, 30(1), 32-47. Stoddard, D. (1986). OTIS~NE(Case 4186-304). Harvard Business School. Tannenbaum, A. S. (1981). Computer Networks.Englewood Cliffs,NJ: Prentice-Hall. . of the Second Conference on Co~puter-supported C~perative Tatar, D. (Ed.). ( 1 9 ~ ) Proceedi~s Work. Portland, OR ACM Press. Tatar, D. (Ed.). (19%). Pr~eedings of the Third Conference on Computer-supportedC~perative Work.Los Angeles, C A ACM Press. Thompson, J. D. (1967). O ~ a n ~ a t i o ninsAction: Social Science Bases of Adminis~ative Theory. New York McGraw-Hill. Toffler, A. (1970). Future Shock. New York Bantam Books. Toulmin, S. (1958). The UsesofA~ument. Cambridge, England: Cambridge University Press. mar~etplace.Unpublished manuscript. New Turoff, M. (1983). Information, value, and the internal Jersey Institute of Technology.

ON^ AND 1 ROW ST ON

Wal~purger,C. A., Hogg, T., Huberman, B, A., Kephart,J. 0. and Stornetta,S. (1988). Spawn: A dis~i~uted com~ututionu~ ecoIogy ~npublishedworking paper). Palo Alto,C A Xerox PARC. Weick, K. E. (1969). The socialp s y c h o l ~of oqunking. Reading, M A Addison-Wesley. 0.E.(1975). Markets and~~erarchies. New York FreePress. W~~l~amson, New York FreePress. W~lliamson,0.E. (1985). The EconomicInsti~tionsof Cupitu~is~. Winograd, T.(1987). A languagelaction perspective on the design of cooperative work. ~ u m u n Computer Interuction,3,3-30. Winograd, T. and Flores, F. (1986). Unde~tundingcompute^ and cognition:A new Fo~ndutionFor design.Norwood, NJ: Ablex.

C H A P T E R

Case Western Reserve University a ~ ~ r i s ~ V. n aTenkasi n University of Southern California

Our research is concerned with the processes of collaboration that are re~uiredby orga~~zations as they increasingly adopt more network-like or~anizationstructures (Drucker, 1988, Huber, 1984; Malone, Yates, C% Benjamin, 198’7). ~e want to understand the kinds of communication that are needed and design information technologies to s ~ p p o r tthem (Boland Tenkasi, 1995; Boland, Tenkasi,C% Te’eni, 1994). By network-likeorganizatio we mean those that resemble open systems as described by Hewitt (1985, 19$6).O~ensystems are composed of ecentralized, autonomous units, each with different and inconsistent know dge bases. Opensystems are characterized by distributed cognition (Hutchins,1996; Norman, 1993) in which the anization is achieved by individuals and technolo~iesacting mains onparts of the overall problem, but pendenci~sinto accountin their action§. stem organization, coordination emerges ~ i t h i n sharply from coordin

BOLAND AND TENKASI

among themselves. We refer to this communication process as one of perspective makingandperspective ta~ng-meaning the ability to develop one’s ownunderstandingof a situation and the ability to recognize and appreciate others’ distinctive ways of understanding the same situation. Dougherty (1992), for example, found that successful new product development in multidisciplinar~ teams was associated with the creation ofcommunication practices that encouragedappreciat~on of each other’s perspectives and their mutual interdependencies. Unsuccessfu~ teams were those where members failed to take each other into account in their individual decisions. We began our research with the assumption that each autonomous unit in an organization has a complex understanding of its en~ironment,technologies and constraints that determines its interpretations and actions. This understanding of its situation is unique and generally unavailable to other units.Vet the actionsof all the unitsare interdependent, each relying tosomeextentonassumptionsastohowotherunits will respond to changes, threats or opportunitiesin their environment. When the environbuild suitment is placid andrelatively stable, such autonomous units could ably reliable imagesof how others weremaking sense of their situation and taking action within it. The logic of interpretation and action usedby others could be learned from observed behaviors anda tradition of expectations could be built up by working together over time. Brown andDuguid (1991) referred to this process as the development of communities of practice,” As economic, political, and market environments become more turbulent, and as technologies affecting design, produ~tion, and distribution begin to change more rapidly, the diverse elements of a decentralize organizationfaceincreasinglydifferentiated en~ronmentsanddevelop unique local logics that could change rapidly. In circumstances of heightened uncertainty and complexity, it becomes more and more difficult to reliin our independent actions,so that coordinatably take others into account ed outcomes emerge. owle edge intensive firms (Alvesson, 1995; Starbuck,1992) such as those in thecomputer, ~harmaceutical, andbiotechnolo stries,rely onthe highly specialized expertise and kno domains (Purser, Tenkasi, 1992). need The to i ctive knowledge do~ains hasresulted in thedevelopment of nizationalformsbased onnetwork or opensystemmodels,mostnolateral-flesibleform of organization (~albraith,1994; ~albraith& Lawler, 1993). Our research concerns the ay information technologies can mediate and su~port collaborafou~d in knowltion in environments of distributed cognition such as those edge intensive firms, andlateral-fle~bleforms of org~izing. Our findings to date indicate theim~ortanceof reco nizing and supporting the processof inquiry individuals employ to str~ngthentheir ~istinctive

2. COMMUNICATIONIN D I S T ~ ~ U T ECOGNITION D

ways of knowing, which we refer to as perspective making, and the process of communicating about their perspectives with others, which we refer to as perspective taking (Boland, Schwartz, Tenkasi, Maheshwari,8C Te’eni, 1992; Boland, Tenkasi,8C Te’eni, 1994;Boland & Tenkasi, 1995). Creatively managing the dynamic tension between making strong perspectives withina community of knowing, and being open to taking the perspective of another community of knowing, is essential for collaboratively developing organizationalknowledge, We find thattheinterdependent,dynamicprocesses of perspective making and perspective takingare achieved through reflection and conversations that involve the narrativizationof experience as well as explicit rational analytic procedures.

Actors in anorganizationalsettinghavetheirownuniqueperspectives through which they identify and interpret the salient features of their situation, understand the values and goals of the organization, and employ a logic of action. For the indi~dual,Boulding (1956) referred to this perspective as an “image,” Pepper (1942) referred toit as a “world hypothesis,” and Bartlett (1932) referred to it as a “schema.” In parallel with this, a number of scholars have commented on the unique cognitive repertoires that can develop at thegrouplevel.Fleck (1935/19~9) referred to a group’s perspective as a “thought world,” Fish (1980) referred to it as an “interpretive community,’’ and Barnes (1983) called it a “contextof learning.” People in organizations do not have many opportunitiesto actively and openly reflect upon the waysin which they or their group interpret a situation or display a theory-in-use in their organizational practice (Argyris Schon, 1978). Thislack of reflection upon their interpretive practices has ied as a potent reason for organizational failure (Nystrom &StarStarbuck & Milliken, 1988). Achieving significant change in the ng of group members requires them to reflect upon existing and structures. Making representations of those them open for discussion is one way of doing so Our research question was how to enable individuals to make their understandingsof a situation visiblea them withothers. We have approached this~uestionas spective making and perspective taking through the vi with cause maps. of social behavior is predicated upon assumptions ana c t o ~ r n ~ k e ~ knowledge, beliefs, and motives of others. This process of spective taking is fundamental:In coming to know ourselves an

BOLAND AND TENKASI

nicating with others, theknowing of what others know is a necessarycomponent (~akhtin91981; Clark, 19%; Clark & Marshall, 1981; Festinger, 1954; Garfinkel, 1967; Gauss & Fussell, 1991; Mead, 1934). As ~ r o w n(1981) observed, effectivecommunicating requires that the point of view of the other be realistically imagined. This reflects the fundamental distinction between a signal which, by convention, indicatesa specified action or object, anda symbol, which always carries a surplus of metaphorical referent and possi(1980) have ble meanin~s(Giddens, 1979). OtherssuchasRommetveit this point:“An essential componentof communicative competence in a ~luralisticsocial world ..is our capacity to adopt the perspectivesof ifferent o t ~ e ~(p. s ” 126). Individuals utilize a variety of inference heuristicsto estimate what others know. Such heuristics can induce systematicerrors and biases (Kahneman, Slovie, & Teversky, 1982). For example, the a u ~ i l a ~ i l i ~ ~means e~~istic that a person’s own perspective may lead that person to overestimate the will be sharedby others. This false c o ~ s e ~ s ~ s likelihood that the perspective effect, in which people assume that othersare more similar to themselves than is actually the case(Ross, Greene, & House, 1977), is a form of bias partic~larly ~elevant to the perspective taking process. Steedman and Johnson)~roposedthat “the speaker assu~esthat the hearer ~ o w s that the speaker knows about the world and about the converati ion, unless there is some evidence to the contrary” (p. 129). That speakoverestimate the extent to which their knowledge is shared by others is ported in studies by Mead (1934), Garfinkel (1967), and ~ougherty(1992). These biases and heuristics imply that the way humans communicatein organizations is based on implicit understandings of the beliefs, values, and knowledge of others that are for the most part untested and assumed to be the same as one’s o w . Dougherty (1992) provide^ detailed case studies of ’ ion in product deve~opment teams9 and shows how unsuccessful to id en ti^ and reconcile~ualitativedifferences in their perspe~sult, the did not attribute the same importance to elements in the e ~ ~ r o n ~ e n tnot , taketheactivitiesandpriorities of othersinto appreciatethedifferencesbetweentheir“thought account, and did

f~rentfrom the others. As ~o~gher

S

a proble~of: (a)ma kin^ the unique~nowled~e and mean-

2. COMMUNICATION IN DIST~BUTEDCOGNITION

ings of each individua~more accessibleby helping them to represent their perspective visually, and (b) e ~ c h a n g i nand ~ discussing those representations withothers. It is a problem of perspective making and~erspectivetaking in which the ability of indivi~ualsin distinct communitiesof knowin elaborate, differentiate, and complexify their own perspective is balance ty to engage in conversation among themselves about those

the organizational sett

'The eath her head School of a~ementmakes the Spider cause able withou~cost for educational Access it through the research he ~du/spider~ap

BOLAND AND TENKASI

depictinganunderstandingthatanindividualalreadypossessed. We assumed that a well developed perspective was resident in the person’s memory, readyto be portrayed as a map. What we found instead was that individuals startedmaking a map by putting up several factors, along with some causal relations. They then shifted atomode of discovery rather than up, relations would representation. In discovery mode, factors would be put suddenly be seen, and new factors would be suggested. Typical comments made after the first few elements had been added to a map were: “Look what’s emerging here,” or “Oh, this is interesting.” The map maker would be actively constructing an understanding while engaging with the tool. This suggests that individua~sdid not have well formulated causal understandings in the first place, and that the opportunity for reflexivity in the construction of maps was itself an innovationin their practice. That individuals were experiencing a senseof discovery duringmap making andthattheywereactivelyconstructinganunderstandingthrough reflecting upon the map as it was being developed, was a challenge to our original assumptions.We had thought our project was about communication in the senseof a conduit model,in which the problem wasto provide an effiof a mescient, convenient channel for the representation and transmission sage in the form of a map. But we have come to see that it was really about reflexivity and interpretation, bothin the construction andin the readingof maps.We have come to see communication of perspectives in distributed cognition as a seriesof conversations-first with oneself while constructing an understanding in the form of a cause map, and then with another while exchanginganddiscussingthosecausemaps.Thecommunicationis in these conversations of perspective making and perspective taking, not in the map as atransmitted message. A second challengeto our initial assumptions came from listening to what people said when they were constructing a map. They did not only use the abstract categoriesof causal factors and relationsto think through the construction of the map, butalso relied on storytelling. Insteadof identifying a factor to include in a map by generalizing it through inductive reasoning, and instead of linking factors with a paradigmatic statement of their causal relationships (if x increases, theny increases), ourmap makers would puta factor into a map and link it with other factors while recountinga dramatic in~identwith a customer, a competitor, or another manager. This suggests that themap makers way of thinking through a situation was notso much an exploration of a problemspaceusinglogical operators as suggested by on (1977) andothers, asit was a narrationof their experienceof being in situation. Through narrating their experience, they identified important factors and explored how they were linked together. ~~~sequently, we have drawn upon the work of Bruner (1990) who posited ~rrative is a fu~damental mode of human cognition, qual in importance

2. COMMUNICATIO~IN DIST~BUTEDCOGNITION

to the paradigmatic, information processing mode of cognition that dominates cognitive science.We have found his work very helpful in understanding how people coordinate activities not only through exchanging messages about states and actions,but also through narrativizing real and hypothetical events in conversations with self and other (Boland& Tenkasi, 1995). We even found this narrative modeof cognition to be evident when managers analyzed the seemingly abstract and unambiguous representations of an accounting budget record (Boland, 1993). There, we observed that managers interpreted budgetandperformancereports by bringing to lifethepersonsbehindthe accounting numbers, endowing those persons with motivations and intentions, and narratinga sequence of events which produced the numbers.

As an example of how Spider can support perspective making and perspective taking in open system, we review a processof causal mappingby physicians in the neurology department of a large teaching hospital. The objective was to improve thedepartm~nt’sability to collaborate and to increase the quality of medical care. The map makers were neurologists who were long time members of the department’s quality assurance committee. The cause maps of three physicians(A, B, and C) are shown in Figs. 2.1,2.2, and 2.3. The physicians first worked individually with oneof the authors,using the Spider mapping tool to makea personal map of the causes of qualityin medical care. After making their individual maps, they studied the maps made by the other physicians to identify similarities and differences. Finally, they met as a group to discuss the possibility of a synthesis. In constructing a map, each physician would first put downthree to five factors and some of the relations between them. In Fig. 2.1, for example, Physician A started with the factors “Technical Quality ofMedicine”and “Social Quality of Medicine.” These factors were introduced with a story about the physician’s long-standing argument with instituti~nalleaders in neurology about the lack of attention to the social quality of medicine. He then added the factor “Time Spent with Patient” along with a story about a patient who requireda long series of questions and seeminglyo cussions before the importants ~ p t o mand s background know ed for a correct diagnosis emerged. The factor “Pre then added, noting how insurance companies and the hospital to think in terms of patients per hour and to pr cians to reduce the time spent with each patient. f~ctors, Then, PhysicianA began drawing the causal relations among those beginning with the negative relation between “Pressure to CutCosts”and

BOLANDAND TENWI

FIG. 2.1. Map of the causes of quality in medical care prepared by Physician A,

BOLAND

AND TENMI

FIG. 2.3. Map of the causes of quality in medical care prepared by PhysicianC.

Figure 2.4 shows such a cycle, in which cost cutting pressures leads to reduced time with a patient, which hurts the social qualityof interaction and eventually results in the patient migrating throughthe health care system, creating increased costs and furtherpressure to cut costs. This vicious cycle was not in Physician A’s initial descriptionof the situation, nor was it in his unfolding discussion of it. Rather, a self-reflective awareness of the cycle emerged through the process of mapping itself and its discovery was a pleasant surprise, ‘~onfirming”Physician A’s initial insight that social qualiof discovery is not to be takenas a new “truth,” ty wasa key factor. This kind but instead as a stage in the perspective making of Physician A. In his perspective making, his wayof understanding qualityof medical was being syst~maticallyelaborated, refined, andstrengthened. In Fig. 2.2, we see that Physician B sees the problem of quality in healt~care the patient’s somewhat differe~tly,As with P~ysicianA, social interaction and perception of the physician are present in this map, but theydo not play as central a role. Instead,the role of science in driving not only the degree of cure the perceptions of care by society, by physicians and that is possible, but also by patients becomethe focal points for defining quality of care. hysician C raises a different set of issues, in which risks from ,the d~agnosis? and the treatment as well as thea~ilityto adjust

2. C O ~ ~ U N I C A T I OINN D I S T ~ ~ U T ECOGNITION D

T i m Spent With P ~ e n t

l”

FIG. 2.4. Examples of a cycle in cause map of Physician A.

treatment become focal points. The roleof the family in helpingthe patient understand his or her condition and follow through withtreatment (especially as the treatment is adjusted) is highlighted. In addition, Physician C recognizes the workload onthe system andthe attitudes of the care team as important in affecting the way that treatment can be dynamicallyadjusted. The “Abilityto Adjust Treatment’’ then becomes the focal point for defining quality of care, which can beseen from the number of factors leading into it. As each of these physicians is shown the maps of the others, the first response invariably is “this map is very similar to mine.” The physicianwill begin by identifying factors that seem similar, such as “’Social Quality of Interaction’’ (Fig. 2.1) and “Questioning and Listening” (Fig. 2.2), or “’Side Effects” (Fig. 2.2) and “TreatmentRisk (Fig. 2.3). After a while, though, the tone shifts and an awareness of the differencesbetween their map andthat of another begins to emerge. The awarenessof differences grows, and leads to a realization such as: “This person sees things very differentlythan I do!”. This realization of fundamenta~perceptual difference between oneself and others is the beginning of a perspective taking process, in which the individual explores what those differences and their implications are. The recognition of difference andthe process of perspective taking, however, requir~d

BOLAND AND TENKASI

lor episode of perspective making in which one’s own understand in^ constructed and reflected upon. Perspective making in the form of a rovides the condition of possibility for recognizing differences with others through close examination of their maps. n this group of neurologists, the analysis andsynthesis of a dozen maps the department leader to express, “I had been thinking about quality much too simplistically.” The maps were intended to reveal different ways of ut the cause of quality in medical care, but ended up revealing initions of what quality of medical care means. The synthesis group produced contained multiple definitions of quality, domg different areas of the map, includingperceptio~sof relevant groups, versus risks, and technical outcomes on absolute an scale. is the way that collaboration on The important point for our purpose here impro~ngquality of medical care was not primarilya question of communicating e ~ s t i n g understandings witha communication modelof encode, transdecode. Instead, it involveda process of perspe uction of an understandingin the first place, ng, made possible by reflecting upon simil tion of differences was achieved. Only then could they engage in a discussion in the future. might coordinate their group efforts

in this andother ~ ~ o j e c we t s , nowsee the process of comordination in en~ronmentsof distri~utedcognition much

coincide amongm e ~ b e r of s the firm. shared meanin~snow seems much more

2. COMMUNICATIONIN DIST~BUTEDCOGNITION

ture or thatit was somewhat stable. Instead, he used the term to express an active process of interpretation that was continuously being constructed, interpretively extended,and creatively revised.In fact, he refers to the way subjects in his experiments appeared to simultaneously employ multiple schemas andto inventively make associations among themin a constructing a remembrance. With thisin mind, we now think of communication and coordinationin disin tributed cognitionas a skill for reflexively narrativizing ongoing experience a way that constructs and reconstructs understandings of a situation, engaging in conversation with others about these narrativized repr tions. We believe that information technology can support this process by enabling the construction of visual images, such as cause maps resentationsthat“mirror”thenotion of schema, includin~ depictions of a story, pictures, or graphic designs, among 0th Tenkasi,1995).We find Star’s (1989) description of the role of a ‘boundary object’ in doing science work to be useful here. A boundary object is not a message being trans~ittedbetween people with different expertise ona scientific team, but something that can be put out between them and use focus fora conversation among themin which they explore its possible ings and implications, It is an occasion for conversation, nota self~ontaine message € otransmitting r and decodingby a recipient. Using the idea of boundary objects, we are now exploring how cause maps, narrative maps, graphic images, and other forms of representation of can be used in distributed cognition to improve organizational practices communicatingandcoordinating. We are interested in howcommunities develop distinct waysof knowing through their ways of narrativizing experience and engagingin conversation about boundary objects.

We see the taskof information technologyto be one of supporting the subjective process of perspective makingandperspectivetaking.This is an interpretive process of inquiry in which an individual constructs a visual is in, makes that reading available to oth“reading”of the situation he or she ers, and engagesin conversation with them, seeking to extend herown horizon of meaning (Gadamer, 1975). What is required from information technology in distri~utedcognition are facilities of self indication, reflection, and interpretation-an environment for active sense making in which individuals can construct representations of theirchangingunderstandingsandcan explore them in Conversation with others. In addition to specialized software such as the Spider cause mapping tool, groupware systems provide an ideal infrastructure to support perspective

BOLAND AND TENKASI

making and perspective taking. Boland and Tenkasi (1995) identifieda number of ways that the discussion forums provided by group technologies can be used to strengthen knowledge structures within a distinctive community of ~ o w i n gand , also to enable conversation on the perspectivesof others. Such forums can allow for narrative as well as calculative offorms reasoning, can include cause maps as well as maps of narratives,and can structure conin versations about such maps along with their implications for collaboration knowledge work. We see communication and coordination as an ongoing process of interpretation through conversation. We are excited about the prospects for designing toolsof visualization, organizational structures, and group processes that better enable such interpretive conversations.

Support for this research was provided by the National Science Foundation ~rogramon coordination Theory and Collaboration Technology (Grant #IN9015526) by the Digital E~uipment Corporation (Grant#111), by the TRW Fo~ndationand by the Kay Star Foundation. The authors give special thanks to Susan Leigh Star and Ulrike Schultze for helpful suggestions on earlier an version of this manuscript.

Alvesson, M,(1995) ~ a n ~ e m eof n t ~no~ledge-Intensive Companies. New York De Gruyter. Argyris, C. and Schon, D, (1978) O ~ a n ~ a t i o nLearning, al a Theory ofActionPerspective. Reading, M A Addison Wesley. Axelrod,R.(1976) Structure ofDecision: The Cognitive Maps ofPolitical Elites.Princeton, NJ:Princeton University Press. Bakhtin, M. M. (1981) Discourse in the Novel.In M. Holquis(ed.), TheDialogic I~agination. Austin: Universityof Texas Press. Barnes, B.(1983) The Conventional Character of Knowledge and Cognition. InScience Observed. London: Sage. Bartlett, F. C. (1932) f em em be^^^ A Study in ~perimentaland Social psycho lo^. London: Cambridge University Press. Bartunek, J.and Moch,M. K. (1987) First-order, Second-order and Third-order Change in Organization Development Interventions: A Cognitive Approach. Journal of Applied ~ehavioral Science, 23(4), 483-500. Boland, R. J.(1993) Accounting and the Interpretive Act. Accounting, O ~ a n ~ a t i oand n s Society, 18(2/3)}125-146.

Boland, R.J.,Schwartz, D., Tenkasi, R. Maheshwari,A. and Te’eni,D. (1992) Sharing Perspectives in Distributed Decision Making.ACM Conferenceon Computer Supported C ~ p e r a t i v eWork, Toronto, 306-313. Boland, R. J.and Tenkasi,R. V. (1995) Perspective Making and Perspective Taking in Communities of Knowing. O ~ a n ~ a t i Science, on 6(4), 350-372.

2. COMMUNICATION IN DISTRIBUTEDCOGNITION

Boland, R. J., Tenkasi, R. V. and Te'eni, D. (1994) Designing Information Technologyto Support Distributed Cognition.O ~ a n ~ a t i Science, on 5(3),456-475. Boulding, K. E. (1956) The Image: nowl ledge in Life and SocietyAnn Arbor: Universityof Michigan Press. Communities-of-Practice:Toward Brown, J.S. and Duguid,P. (1991) Organizational Learning and a Unified View of Working, Learningand Innovation.O ~ a n ~ a t i Science, on 2(1), 40-57. Brown, R. (1981) Social P s y c h o l ~New . York The Free Press. Bruner, J.S. ( 1 ~ ) A c t sof~eaning. Cambridge, M A Harvard University Press. Clark, H. H. (1985) Language Use and Language Users. In G. Lindzey and E. Aronson (eds.), Handbook ofsocial PsycholoD New York Random House. Clark, H. and Marshall, C. (1981) Definite Reference and Mutual Knowledge. In A. Joshi, I. Sag, and B. Weber (eds.), Elements of Discourse Understandi~.Cambridge, England: Cambridge University Press. Dougherty, D. (1992) Interpretive Barriers to Successful Product Innovation in Large Firms, O ~ a n ~ a t i Science, on 3(2), 179-202. Drucker, P.(1988) The Coming of the New Organization.H a r ~ r Business d Review, Jan-Feb, 45-53. Festinger, L. (1954) A Theory of Social Comparison Processes. Human Relations, 7,117-140, Fish, S. (1980) Is There a Text in ThisClass? Cambridge, M A Harvard University Press. Fleck, L. (1979) Genesis and Developmentof a Scientific Fact.In T. Trenn andR. K. Merton (Eds,), Chicago: Universityof Chicago Press. (Originally published in 1935) Gadamer, H, G. (1975) Tmth andM e t h ~New . York Seabury. Galbraith, J.R, (1994) Competi~with ~ t e r a l ~ ~ ~ i b l e O ~Reading, a n ~ a ~MoAnAddison s. Wesley. Galbraith, J.R. and Lawler, E. E. (1993) O ~ a n ~ i for n gthe Future.San Francisco,C A Jossey-Bass. Englewood Cliffs, NJ: Prentice-Hall. Garfinkel, H. (1967) Studies in Ethnometh~olo~. Giddens, A, (1979) Central Problems in Social Theoy. Berkeley: Universityof California Press. Hewitt, C. (1985) The Challengeof Open Systems.Byte, pp. 223-242. on OfficeInformationSystems, 4(3), Hewitt, C. (1986) Offices are Open Systems.ACM Transac~ions 271-287. Design of PostIndustrial Organizations.~anage~entscience, Huber, G.P, (1984) The Nature and 30, 928-951. Huff, A. S. (1990) map pi^ Strategic T h o ~ h tChichester, , England: Wiley. Hutchins, E. (1996) C~nitionin the Wild. Cambridge, M A MIT Press. P.and Teversky,A. (1982) Judgment Under uncertain^: Heuristics and BiasKahneman, D,, Slovic, es, New York Cambridge University Press. Krauss, R.and Fussell, S. (1991) Perspective Taking in Communication: Representation of Others' Knowledge in Reference.Social Cognition,%l), 2-24. Malone, T., Yates, J. and Benjamin,R. (1987) Electronic Markets and Electronic Hierarchies. Communications of the ACM, 26,430-444. Mead, G. H. (1934) ~ i n dSelf , and Society.Chicago: Universityof Chicago Press. s MakeUs Smart. Reading, M A Addison Wesley. Norman, D. A. (1993) T h j ~that Nystrom, P. C. and Starbuck,W. H. (1984) To Avoid Organizational Crisis, Unlearn. O~an~ationa1Dynamics, 12(4), 53-65. Pepper, S. C. (1942) World Hy~otheses.Berkeley, Universityof California Press. Purser, R. E., Pasmore, W. and Tenkasi, R. (1992) The Influence of Deliberations on Learning -in New Product Development Teams. Journal of Engineering and Technology ~ a n a g e ~ e n9,t , 1-28. Rommetveit, R. (1980) On Meanings of Acts and What is Meant by Whatis Said in a Pluralistic Oxford: Blackwell and Mott. World. In M. Brenner (ed.), The Structure ofAc~on, ROSS,L., Greene, D. and House, P. (1977) The False Consensus Phenomenon: An Attributional Bias in Self-Perception and Social Perception Processes. Journal of Experimental social psycholoD, 13,279-301.

BOLAND AND TENKASI Simon, H.A. (1977) The New Science of Mana~ementDecision (2nd rev.). Englewood Cliffs, NJ: Prentice-Hall, Star, S. L. (1989) The Structureof Ill-Structured Solutions: Boundary objects and heterogeneous ted Distributed Problem Solving.In M. Huhns andL. Gasser (eds.), Readings in ~ ~ ~ i 6 uArt& cia/ ~ntell~ence 2. Menlo Park,C A Morgan Kaufmann. Studies, Starbuck, W. H. (1992) Learning by Knowledge Intensive Firms.Journal o~Manage~ent 29(6), 713-740. Starbuck, W. H. and Milliken, E J, (1988) Executives’ Perceptual Filters: What They Notice and How They Make Sense, In D, Hambrick (ed.), The Executive Effect;Concepts and M e t h ~ sFor Stu~yingTop Managers.Greenwich, CT: JAI Press. Steedman, M. and Johnson-Laird,P. (1980) The Productionof Sentences, Utterances and Speech Acts:HaveComputers ~ ~ h i to n Say? g In B. Butterworth (ed.), L a n g u ~ e ~ ~ u c t i o n s : Speech and Talk.London: Academic Press. Learning: The NarraTenkasi, R. V. and Boland,R. J. (1993) Locating Meaning in Organ~zationa~ tive Basis of Cognition. In R. W. Woodman and W. A. Pasmore (Eds.), Research in Organhat i ~ aChange l and~ e v e l o ~ m e(Vol. n t 7, pp. 77-103). Greenwich, CT: JAI Press. Weick, K, E. and Bougon, M. K.(1986) Organizations as Cognitive Maps: Charting Waysto SUCng San Francisco: cess and Failure. In H. Sims and D. Goia (eds.), The ~ i n ~ i Organhation. Jossey-Bass.

C H A P T E R

University of Michigan

Orincon C o ~ o ~ t i oSan n , Diego, CA University of Texas at Arlington

arcus 1. Huber Intelligent Reasoning Systems, Oceanside, CA Ford Motor Company, Dearborn, MI

S a n ~ Sen i~ University of Tulsa

A variety of tasks and problems become apparent when investigating a broad interdisciplinary field suchas coordination theory and co~laborationtechnol-

ogy. Among them is the problem of characterizing the field in some way. Because our work, andthe work reported in this book as a whole, emerged we can from a programin computer and information science and engineering, begin by characterizing computational tools for sup~orting~oordination.A simple characterizationof tools forsuppo~ing human collaboration has been literature, where seen in the computer-supported cooperative work( C S C ~ different tools are identified with different time and place' characteristics 'For the purposesof this chapter,Uplace" connotes a locusof attention. Thus, different places could be geographically separated, but they could also be conceptually separated. The key issue is that decisions need tobe made in the contextof various physical and/or mental "places."

DURFEE ET AL.

.The result isa lattice asin Fig. 3.1, where we have placed representative entriesin each of the matrix elements (many other entries can and have been madein the literature). Although the question being addressed in the entries of this matrix is implicit in the CSCW endeavor, for our purposes it is important to make it explicit. What the matrix is categorizing are “tools that support collaboration on a particular task among multiple participants who are in the x and y,” where x and y are values along the different matrix indices. Broadly construed, therefore, the kinds of technologies that fit into this type of matrix are collaboration technologies, which assume that participants~~~t to collaborate on a particular task. This perspective abstracts away issues of conflict, in order to pay attention to the detailsof achieving coherent collaboration among participants (see, e.g., the ~TCT-supported work on collaboratories so forth). such as thatof Schatz, Atkins etal., Fischer, Olson, and Conflict, however, can arise whenever there aremultiple agents andmultiple tasks. To get to the point where the question associated with Fig. 3.1 makes sense, the agents must initially establish (negotiate) the goals to ichtheparticipants are committed(e.g.,theCTCT-supportedworkof aus and Wilkenfeld). Moreover,in many applications, agentswill contend over resources (including each other’s attention) as different combinations of agents pursue different tasks. Resource allocation thus becomes a critical concern (see, e.g., the CTCT-supported work of Pasquale). Our work assumes that a sophisticated agent-such as a person-is too dear to be dedicated to only a single task, and thus an agent must be to able pursue multiple tasks at the same time. From this perspective, then, we use the same matrixas in Fig. 3.1 but aska different (although just as important) question. Because mostif not all participantsin a collaborationare also participating in other collaborations, let us use the matrix to categorize “tools that ”support collaborations on multiple tasks by a siqfe participant, when the collaborative tasks are in the x and y” (see Fig. 3.2). Time Same

Merent Times Electronic Bulletin

Electonic ~ ~ b o ~ d s

Electronic Mail FIG. 3.1.

3. COO~INATIONAS D I S T ~ B SEARCH ~~D

Same Times Different Time

Tools for rapid task completion upon arrival among d i s ~ b u ttasks e~ FIG. 3.2.

A person can face tasks at the same place and different times, as visitors (email messages)arrive periodically at the person’s office (mail program) in the course of a day. Or the tasks might be at different places at different times, so that the person must travel (perhaps electronically) from task to task during a day. Either way, computational tools such as advanced interin his or her collabofaces and networking software can support the person rations by allowing the person to move more quickly among(totasks switch physical or mental contexts better) and to complete each task faster. Often, however, tasks for various collaborations should be done at the same time. For such tasks, there are two general approaches to providing computational support. One approachis to use computer processes toprio ~ i t tasks, ~ e which serves to conceptua~ly push the problem back into the “different times” column. These processes can, for example, filter and sort email (Malone, Grant, Turbak, Brobst,& Cohen, 1987), and can solve scheduling problems to help the user navigate among competing tasks so as to attend to them one ata time in the proper order. Hence, this approach performs “triage” on the tasks, but the user should still eventually attend to them all personally. The other general approachis to delegate responsibility for tasksso that they can indeed be handledin parallel. The idea hereis to generate,to some degree, processes for the user that act on the user’sbehalf when he or she is otherwise occupied. Thus,in this approach, the user might never have to attend to some of the tasks. In the case of“same place, same time,” these processes could reside at the user-machine interface, intercept in^ tasks meant for the user and completing them semiautonomously (suchas filing incoming email messages). Much of the recent work in building “agents” into interfaces has been directed toward this problem. In the case of “different places, same time,” the processes could reside (geographicallyor concep tually) remotely from the user, acting on hisor her behalf with little if any supervision on the part of the user. Thus, while interface agents could be monitored and continuously tailored by the user, remote surrogate agents

user employing surrogate agents must therefore have confidence in his or her decision making, and the agents must have the ability to act and interact flexibly and adaptively ascirc~mstanceschange, based on well-founded criteria when at all possible. Our work is to develop theories and mechanisms for creating such surrogate agents, andin particular for giving such agents the ability to coordinate. ~oordination,as we define it, can be in support of collaboration, but need not be. We construe coordination broadly as the processof considering the likely decisionslactionsof others when deciding what to do. If such considerations lead to concerted activity for mutual benefit, then the coordination process supportsa cooperative outcome. But such considerations can also occurin noncooperative situations, where a successful competitor could be one who coordinates its decisions most effectively against those of others. Implicit in the matrixof Fig. 3.2 is that competitionwill exist, at least competition for the attentionof a participant in an interactio Fig. 3.2 capturesaspects of coordinationtheory,while F focused on collaboration technologies. Theworkoncoordination theory that wesummarize in thi ad~ressesonly a small subsetof problems that arisein developi tational mechanisms for coordinating surrogate agents. The problems we particularly concentrateon are: ow to analyze, understand, and represent an application domain in such a way that an agent can make quantitatively based adaptations to itsmultiagentenvironmentwhencarryingouttasksonbehalf of a human user. * Howanagentcangenerate,communicateabout,andmakecommitments to modelsof itself and of others so as toget an adequate appreciation of the multiagent environment with which it must coordinate its decisions. * How protocols for communication and interaction can arisein an arbitrary system of agents, despite the agents having differing objectives and capabilities. In this chapter, we describe some of our results in investigating these problems. We begin with the first problem, and outlinea me tho do lo^ that has allowed us to develop quantitative predictions of performance that can be usedby an agentin the application domainof distributed meeting scheduling. We then turnto the second problem, where, unlike meeting scheduling where mutual agreement on a meeting time is a shared objective, we consider the possibility that agents are not necessarily desirousof collaboration. Instead, coordination might be geared toward avoiding conflicts rather than reaching detailed agreements, and so might require agents to

3. ~ O O ~ ~ N A T I AsODNI § T ~ B ~ SEARCH ED

make fewer commitmentsto each other. Next, we turn to e of protocols in agent populations, and conside ssumptions on the partof a ents might give rise to a rich variety of situation~ependentmodesofinteraction. We conclude by arizing whatwehavedone,andoutliningsome of thechallengesremforthe future.

Researchers who have been developing artificial agents have taken a variety of tacks in realizing them. Some researchers have focused on the notion of building “believable” agents, that act in andreacttotheir enviro~ment (includin~humans) in a manner that appears to mimic how natural agents would behave (Maes, 1994). In a similar vein, some researchers have been investigating ‘‘learning” agents that can,in some way or another, o the behaviorof a human and learn from that behavior toin act a simil ner in similar future situations c itch ell, Carvana, Freitag, Zabowski, 1994). Other researchershavedevelopedmore“instructable” agents, that can be given explicit instr~ctionsabout how to behave under different circumstances ( alone et al., 1987). Our emphasis has been on building agents that can be physically and concep~~ally distant from the humans they represent, and therefore we have shiedawayfromassumingthatanagentcan observe its correspon human or that ahuman can supervise his or her agents. We have thus concentrated more on designing agents that fallcloser to the instructable category, although we have been particularly concerned about the degree to which the instructions for the agents can be well-founded and quantitati~e, rather than heuristic and qualitative. Agents with such detailed models of their application domain and available strategies, even if not “believable” (they might not act like humans-who could well lack or ignore such models) nor “instructable” (improving the modelin a principled way would require moreeffortthansimplyprovidinganinstruction), are expected to be “trustable” (they would make principled decisions based on ~uantitative analyses). Our particular domain of in~uiryhas been the meeting-sche~uling application, and within this domain we have analytically developed and experimentally verified~uantitativepredictions of performance for various strategies involved in scheduling meetings (§en, 1993; Sen & Durfee, 1991, 1992, 1994b), including strategies for deciding how many possible meeting times to propose, strategies for rejecting or counterproposing meeting times, strategies for committing to and canceling meetings, and strategies for deciding which possible meeting timesto (c0unter)propose at any given time.

DURFEE ET AL.

For example, the choice of strategy for deciding which time intervals to consider next for a meeting will have numerous effects, including effects on the density ofmeetings in different parts of the calendar, the likelihood of scheduling future meetingsof different types, the costs of scheduling, and the time needed to schedule meetings. More importantly, given targets for calendar densities and limited costs and time for scheduling, a calendar management agent should adapt its strategy choice based on the larger context of what it expectsto schedulein the future and what it knows of the calendars of the other agents. These adaptations might not be under the constant superembedding domain vision of the user, and thus (we argue) should bebymade knowledge (a rigorous model of the task) into the agent, rather than trying to capture a superficial modelof the user actingin a small sample of cases. The agents in a ~istributedMeeting Scheduling @MS) system exchange relevant information to build local schedules that fit into a globally consistent schedule. To facilitate information exchange, the agents need a common comm~nicationprotocol for negotiating over meeting times. For our agents, we have chosen to adapt the multistage negotiation protocol (Conry, Kuwabara, Lesser, & Meyer, 1991), which is a generalization of the contract net protocol (Smith, 1980). In our protocol, each meeting has a particular agent who is responsible for it, called the host. The host contacts other attendees of the meeting (who are called invitees) to announce the meeting, and collects bids (availability informatio~). This process could be repeated several times before a mutually acceptable time interval is found, or isit recognized that no such time interval exists. Other meetings could be undergoing scheduling concurrently; in general, an agent can simultaneously be involved in scheduling any number of meetings, acting asa host for some and an invitee for others. How well this protocol performsin efficiently converging on good schedules is strongly impactedby heuristic strategies about what informationto exchange and how to model tentatively scheduled meetings. In order to decide what to propose for a meeting, the agents have to search their calendars in a systematic manner using some appropriate search bias. Strategies for communication must balance demands for privacy (which lead to exchanging less information) with demands for quickly converging on meeting times (which can be sped up by exchanging more information). Strategies for modeling tentatively scheduled meetings can range from blocking off tentative time(s) for a meeting unless and until the arrangements fall through,to ignoring tentative commitments abouta meeting when scheduling other meetings.We have embarked on research to develop, analyze, and verify a formal model of DMS to formulate rigorous, quantitative predictions of the perfor~anceof the following typesof heuristic strategies:

3. COO~INATIONAs DIST~B~TED SEARCH

Searc~ ~iases determine the order in which the calendar space is searched to find acceptable time intervals for a meeting. Here, we consider linear early (LE)where agentstry to schedulea meeting as early as possible, and hierarchical (H),where agentsbuild a temporal abstraction hierarchy over the calendar space. At each node in the hierarchy, agents keep a recordof the number of intervals of different lengths free below that node in the hierarchy. The calendar space lends itself to a very natural hierarchyof hours, days, weeks, and so on, and the agentspa~icipatingin a meeting can first identifya good week to meet in, then identifya good day within that week, and fina~ly an actual interval within that day. Given a meeting of some particular length to schedule, the host obtains information about the invitees regarding how many intervals of that lengthare open at each node (e.g., at each week) athighthe est levelof the hierarchy.It multiplies the numbers together for corresponding nodes, ranks the nodes, elaborates the best one, and proceeds to repeat the process for the next level of the hierarchy under the elaborated node? ~ a c ~ r a coccurs ~ n g if a particular portion of the ground level being elaborated contains no solution to the schedulin~problem. Anno~ncement S~ategies determine howa meeting is announced, and usually involve proposing some numberof possible times.We specifically consider the options called best (where only the best meeting time from the host's perspective, rankedby some heuristic like being theearliest, is communicated)and good (whereseveraltimespreferred by thehost, 3 by default, are communicated). idd ding S ~ a ~ e g i edetermine s whatinformationaninviteesendsback based on an anno~ncement. We consider the options called yes-no (where by the host) and alternatives an inviteesays yes or no to each proposal sent (where an invitee proposes other time(s) when it can meet). Commit~ent S~ategies are committed (when a time is proposedby a host or invitee agent, that agent tentatively blocks it offon its calendar so no other meetings can be scheduled there) and n o n c o m ~ i ~ e(times d are not blocked off until full agreement on a meeting time is reached).

Our analysis of various strategy options (Sen& Rurfee, 1992) show that no one strategy combination dominates another over all circumstances. Changing environmental factors like system load, organization size, and so on, can produce a change in the strategy combination that will produce the best If there is a way to predict the best results on any given performance metric. 2Although someof this information could be obtained indirectly, we assume in this chapter that agentswill directly communicate about the densities of various intervals, witha resulting decrease in privacy.

DUWEE ET AL.

strategy combination for a given performance metric and given environmental and system conditions, we would like our automated scheduler to take advantage of that. Such a scheduler would be adaptive to changes in the system and theen~ronment, pro~ding us with the most desirable performance as measured by certain performance metrics.In this chapter, we concentrate on adapting the search bias to adjust to environmental and performance needs. More details, as well as discussion on other aspects of the 1993; design of an adaptive meeting scheduler, can be found elsewhere (Sen, §en & Durfee, 1994a,1994b). Asmentioned earlier, a search bias sorts the available meetingtimes accordingto some metrics;a scheduling agentwill use the results of the bias to order the proposals it makes for meeting times, from best to worst. There are a varietyof factors that could go into ordering proposals. For example, will arrive continbecause a scheduler must consider that meeting requests uously, one factorin deciding where in the calendar to schedule a meeting will be the expectations the scheduler has about the demands of future meetings. For example, someusers might be subjectto receiving high priority meetreq~eststhat arriveon short notice. TheseHPSN meetings demand that calendar have open slots on a relatively frequent (e.g., daily) basis.In turn, then, an appropriate search bias should try to spread meetings out,so as to leave some uncommitte~ time every day. The hierarchical (H) search bias does precisely this: By first concentrating on portionsof the calendar that are most likely to accommodate a meeting, the bias tends to place meets evenly in the calendar. Our simulation studies have verified this behavior (Sen,1993). But the tradeoff with scheduling meetings evenly is that it tends to fragment the calendar.A user who is subject to receiving infrequent requests for ~ong~uration meetings might be better off tryingto pack meetings into the so as to leave larger chunks of time free for, calendar as densely as possible, for example, spendinga day at a remote site. The linear early (LE)search bias does preciselythis, giving top priorityto scheduling meetin~sas early as possible, and thus leaving later portions of the calendar relatively empty. Again, we have empirically verified this behavior (Sen, 1993). Typically, the scheduling climateof a user will vary over time, at times having more HPSN meetings and at other times having morelong~uration meetings. An adaptive meeting scheduler should be sensitive to the changing character of meeting requests, and change appropriatelyso as to maximize the chancesof successfully scheduling current and anticipated future meetings. But other factors, beside the changing meeting environment, caninflualso ence the selection of strategies, such as the expected cost and indelay scheduling a meeting. For example,if agents use theH search bias and have sched-

3. COO~INATIONAS DIST~~UTED SEARCH

ules that are about evenly dense, then the criterion that the H search bias counts on,of differentiating portionsof the calendar based on the densityof meetings, becomes less discriminating. Because all parts of the calendarare equally likely to yield a good meeting time, switching to theLE search bias can actually reduce the cost (number of iterations required) to arrivea at meeting time, as it does not require passing messages about the inner nodes of the H search bias does. temporal abstractionof the calendar space as the But, if the LE search bias is used fora while to schedule meetings,there will be considerable density variations along the length of the calendar; when a meeting is being scheduled for a large number of attendees, such variations can combine to givevery different success probabilities at different partsof the calendar.~d~itionally, if the host is more free than invitees, and the latterare using a yes-no bidding strategy, a savings can be obtained by using the H search bias. Hence, after a while there will be sufficient mismatch between the schedules of agents to warrant the switch back to the hierarchical search bias. The selling point of adaptive search bias is that the scheduling agent can choose the most appropriate bias for any meeting based on the current statesof the attendee calendars. ~ u i l d i n gadaptiveschedulingagentsrequiresthatthesequalitative e~pectationsbemapped to rules of behaviorforstrategyadaptation, to provide quantitative measuresto the which in turn means that we need will agents to helpthemdecide.Thefollowingprobabilisticanalysis assume the availability of the density profile characteristics (or DPCs, which depict the variation of meeting densitiesover thelength of agent calendars) for the desiredmeeting length for the attendeesof the meeting. ~ ~ l y s ~ sConsider . theLE search bias with the invitee agents responding with acceptance or rejection messages (no counterproposals). In this scenario, a host agent is trying to schedule a meeting withA invitees. Let K = L -l + 1 be the number of placesin which a meetings of length l can start on a working day of length L. For invitee x,let nXS K be the number of these intervals open on the dayin question.If the host was using one proposal per iteration, the probability thatwillit take I iterations to schedule the meeting is

proposing N intervals per iteration, the probability that itwill take i iterationstoschedulethemeetingisgiven

by

e,N=

N*i

%l

j=N(i-1)+1

DUWEE ET AL.

Nowwe analyze the H search bias. Let us assume M agents (numbered .,M) are involved in scheduling a meeting of length I, and that they have constructed identical temporal abstraction hierarchies over the base calendar space (the linear orderingof calendar hours). For any internal node in the hierarchy, we will calculate the probabilit~that one or more common intervals are free in the base space of the calendars under that node, for every attendee of the meeting. Let x be the node in question. Because the hierarchies formedby the agentsare identical,every agent hasa(2,x) intervals of length 1 below this nodeof its abstraction hierarchy.Let the number of intervals of length I currently free under nodex for agent i be f#,x). We 1993, for details): can then calculate the following (see Sen, least one interval is freein each attendees calendar under node x} 1,

Given a two-level hierarchy (days and hours), for example, we can use this equation to compute the probabilities for scheduling under each day. The days are then sorted by these probabilities, and the agents negotiateover days in decreasing orderof this probability untila mutually free interval is found. Moreo~er,for a particular situation, we can develop probability mass functions and probability distribution functions of the random variable corresponding to the iteration at which scheduling is complete, using the previous equations for the LE and H search biases.This information can be used to calculate the expected numberof trials to schedule a meeting, and thus give a measure of the expected scheduling cost. If this isa predominant conwill adopt the bias that minimizes expectcern, then the adaptive scheduler ed cost. e r j ~ e ~ t sIn. the following, we consider two different sets of DPCs (see Fig. 3.3). Each set involves a host trying to schedule a meeting of length I = 2 hours with2 invitees. All the agentsare assumed to be managing

calendars divided into10 blocks of 5 hours each.We assume, in the caseof the hierarchical search bias, that the temporal abstraction hierarchy is com prised of calendar hours, whichare grouped into these blocks. The hierarchical egot ti at ion mechanism, used by the host, first gathers information from the invitees about their respective calendar densities in each of the blocks, orders theseblocks by the probabilityof successfully scheduling a meeting in each of the blocks, and negotiatesover one block ata time going down theordered list,As the first iteration involves exchanging information

3. COO~INATION AS DIST~BUTEDSEARCH X

0 P

0

P

E

N

N 0

FIG.3.3, Some example density profile characteristics. Three participantsin

the meeting, with indicated number of OPEN intervals for each of the time blocks indicated.

about the i of the abstraction hierarchy, meetings can only be scheduled second iteration the only. Assuming meetings cannot straddle blocks, there are four intervals in each block that could have accommodated the meeting if the calendars were empty. So, the constant K in the LE equation, and terms a(1,x) Vx E {l,...,lo} in the H equation, are equal to 4. The n, and the fj(1,x) terms are obtained from the DPCs. Considerthefollowingtwocases(formoredetails,seeSen curve 1994a). In the caseof the first setof DPCs, assume that the host has the labeled DPC1, with one interval open per block. In this case, the expected number of iterationsto schedule the new meeting givenLEthe strategy (when announcing the single6est or the three goaf intervals) can be contrasted with In this case, theLE strategy with either announcementstrate gy is expected to be less costly thanHthe strategy (Table3.1). Note, however, that theDPCs in this caseare nice and level-the kind of DPCs that arise from by the H strategy. Thus, whereas these the even meeting distribution attained DPCs indicate theHstrategy has beenin use, the quantitative cost predictions indicate thata shift over to the LE strategy is in order. The second case presents a contrary example. The second setofDPCsis much less uniform,as would be expectedif the schedulers had been using the LE strategy. In this case, if we compare the expected cost (number of iterations) of using the LE strategy with thatof the H strategy, we see that switching to theH strategy is expected to yield a considerable savings (Table 3.1). Finally, let us revisit the notionof privacy. The analysis that we just presented captures a subset of the strategies and performance metrics that are more generally of interest (see Sen8~Durfee, 1994a, fora fuller discussion).

DURFEE ET AL.

TABLE 3.1 ~ x ~ cIterations t e ~ for ~che~uling the

Case1 Case2

1.294 3.836

2.575 10.68

3.575 2.193

For the caseof privacy, the amountof information revealedis related to the numberof iterations and the announce~entstrategyforthelinear announcement cases. Thus, the expected number of intervals that the host E(best) equals the number of iterations, while the intervals for LE(good)isthenumberofiterationstimes the numberof intervals proposedperiteration (in ourEkample, 3). Thus,notsurprisingly,LE(best) retains more privacy than LE(good) because it reveals information more slowly-leading to more time spent iterating. Choosing among these strateends on the relative im~ortanceof privacy versus the delays ociated with more iterations, The hierarchical strategy is less easilycomparable,becausethefirstiteration in H exposesinformation about larger portions of the agents’ calendars, albeit at an abstract level So, thers stillwill not know precisely when another agent is free. in Table 3.1, deciding whether H preserves privacy better than other strategies dependson the degree to which the abstract exposure is of concern. Certainly,H involves many fewer precise intervals being revealed. In this section, we have briefly described how an application domain can be

modeled such that a surrogate agent can make quantitatively based decisions about adapting its strategy, evenif the agent’s associated useris not a~ailableto train it. This kind of analysis can form the basis for developing trustable surrogate agents that can be adaptive over a varietyof environin this area includes mentalandperformanceneeds,Ourongoingwork develop in^ design strategies for such agents, with particular emphasis on the intertwined adaptive design ~ecisionsabout the various strategies that a meeting scheduling agent can employ(see §en& Durfee, 1994a).

omain demonstrated how, during itsactivity on behalf of a user, ight desire to modi^ the strategies (or behaviors) it employs. In main, for example, being able to change between hierarchical (H)

3. COO~INATIONAS D I S T ~ B ~ E SEARCH D

and linearearly (LE) strategies for selecting proposed meeting times allows an agentto m~imize its chancesof success whileminimizing its 'costs as its Y be scheduleevolves. Butsome of thebenefits of astrategy m strategychoices of othexamaccrued if anagentcandependonthe is known ple, if an agentreceives a counterproposal from another agent that to be employing theLE strategy, then it knows that it is useless to~ r o ~ o s e any time intervals earlier than the counter proposal (those intervals must not be free for the other agent, or it would have proposed them). In fact, if a group of agents were to commit to using the LE bias, then they all might take an agent a d ~ a n t aof~ this e to schedule their meetings more efficiently.if But, can unilaterally change its bias, these advanta~esmight not ~bviously,there are some things about another agent that to know,and other things that are less important. In the D strongamount of cooperis e~pected:All participantswanttosuccessfully scheduleameetingeover, a~reementof fairlypre toarrangeameeting in term needed(althoughit is p you in the lab tomorrow afternoon^ such arrangements ciencies in searching and context switch in^ for the agents coordination mi~htrequire much less specific commitme the agents. For e~ample,mobile agents who wishto avoid make commitments, but possibly CO where t they will not be, rather than where they will be, at ~ommitments of a less detaile nature not only avoid overcommitment in

DURFEE ET AL.

time (perhaps forever), or is it more important tovery havespecific predictions about the immediate future? The answer to this question depends on how the agents are interacting. For example, in DMS, the model was very much like the contracting model:An agent knew exactly about the availability of a particular interval for another, but that did not give it insight about other intervals, andtheknowledgeithadcouldbeshort-livedasother agents try to schedule with it as well. Moreover, the specific model in the DMS domain can represent a commitment that the agent generating the model-saying “I am free at this particular time”-might prefer not making.So how an agent presents itselfto othersin the model it projects can impact its predictabi~ityand flexibility. In short, effective coordination usually implies that an agent know just enough aboutothers to achieve its goals (such as avoiding conflict) withou k n o ~ n gany more (which can waste memory, computation, and communication).Similarly,anagentshouldonlymakethemostminimalcommitments it can to othersto achieve the coordination,so as to retain asmuch flexibility for itself as possible.

A fundamental

objectiveof developing a theoryof coordination is that the theory should support the processof determining the right level of modeling that agents should use to maximize performance while minimizing overhead. This determination is typically only found through search: The interacting agents in ’a society must find, often with many failed attempts, the appropriatelevel ofmodelingandcoordinationreasoninghebenefits of coordination without incurring excessive costs. For example, when there were very few automobiles on the road, interactions among them were infrequent enough, and the characteristics of the roads were varied enough, that allowing drivers to develop models of each other and negotiate over the road madesense. As encounters increased and roadways became more uniform, coordinatin~interactionson a case-by-case basis became less efficient compared to enacting traffic laws. In a sense, the traffic laws generalized and codified how past encounters had been negotiated. We turn shortly to an example of coordination search for controlling traffic in a simple robotics environment. First, however, let us consider the com ponents of coordinationsearchmoreabstractly.Theneedtorepresent actions and interactions at different levels of generality leads to the first abstract element of a framework for viewing coordination as distributed search: ~ e I: The ~ ability t to model yourself, others, and collections of agents as a hierarchyof more or less abstracted behaviors.

3. COO~I~ATION AS D I S T ~ B ~ SEARCH ED

What this means is that an agent must be capable of representing itself-its goals, plans, expectations,and so on-in many different ways simultaneously. A courier, for example, represents itself as delivering packages overnight to its customers without giving detailsof the routes that packages will travel. At the same time, a courier also might give a more detailed model of package movement needs when contracting with an airline to carrypackages. Next, in order to specify information about interactions, it is necessary that an agent be able to identify relationships among the activities (goals, plans, etc.)of agents, and the (dis)utilitiesof those related activities: ~ Z e ~ e 2: n tThe abilitytocomparemodelstoidentifyrelationshipsand measure their desirability. Of course, detailed comparisons and measures of desirability are going to be very domain dependent; wesee an example with the robot delivery task shortly. But the quality of the comparisons and measures depends on what level of abstraction they are (or can be) seen. Finding the right level of abstraction tofind, evaluate, and modify relationships requires: ~ Z e ~ e3:n The t ability to search through a space of models, typically usin communicationprotocolsbecausethespace will bedistributedamong agents, tofind relationships at the right level of detail. And once relationships are found, the agents need the ability to modify those relationships,For example, if the anticipated activitiesof two drivers leads to a collision relationship, the drivers generally should search for alternative activitieswith less dire interactions, such as having one driver delayed until the other passesby. This means the agents need: EZe~ent4: The ability to locally search for modifications of models at the right level of detail to change re~ationships. Finally, the search process itself represents actions that the agents are taking, andthey cannot afford to necessarily search for all possible abstractions, relationships, and alternatives. What they need is: ~ Z e ~ e5:n The t ability to apply (often heuristic) control know~edge to avoid unnecessary work/overheadin the local and distributed search process. in a hierElsewhere, we have argued that viewing coordination as a search technique^, archical space allows us to treat previously distinct coordination suchasorganizationalself-design,multiagentplanning,and distrib~te reso~rcescheduling,in a uniform way Durfee, 1993; Durfee

ted search allows us to

DURFEE ET AL.

els of others only to the degree of detail necessaryto coordinate with them successfully. Thus, partof the distributed search that agents must engage in is a search for the right level of detail to model each other. To realize sucha search process, our research has investigated the of use a protocol for coordination when agents have hierarchical representations of their behaviors.The basic idea is simple. When an agent encounters one or more other agents with which it is unfamiliar, the agents should in engage adia~ogue to coordinateadequately.Theybeginbyexchanging very abstract models of their behavior. This might suffice to discover that the agents, in fact, have no needto coordinate at all: They know enough about each other to know that they do not need to coordinate. If, at an abstract level, it appears there could be reason to coordinate, they can exchange selected elaborations of their abstract model, only including information that is expectedto be relevant. They can keep doing this kind of exchange, providing more detailed and narrow modeling information, until they have sufficient models to coordinate their decisions. of hierarc~icalprotocol was seen in the meeting schedu~ing where agents could incrementally narrow down possible meeting times, only concentrating on areas of their calendars that were most owever, notice that the goalof distributed meeting scheduling agents to eventually get down to the level of tal cific time intervals. This need not be the case for all types Our hierarchica~ protocol, in fact, is based on the assumption that it might often be the case that agents break off elaborating their modelsin favor of

3. C O O ~ I ~ A T I ASOD~I S T ~ B ~ E SEARCH D X

FIG. 3.4. Two robotsdeliverincolor-codedobjectsfrom to destination (right room).

source (leftroom)

X

Y

FIG. 3.5. Robots permanently divide area to avoid collisions.

DUWEE ET AL.

X

FIG. 3.6. Spatiotemporal volumes with overlap for first deliveries.

apart spatially (Fig. 3.7) as Robot 2 again uses the distant door,or they could move their behaviorsapart temporally (Fig. 3.8) Robot 2 waits until Robot1 is done before beginning its activity. By communicatingat even deeper levels of detail, the robots could break their behaviors into portions constitutingmoving to the door, thro~ghthe door, and beyondthe door. Doing so will not changethe spatial approach to removing the conflict, but it can improve the temporal solution (Fig. 3.9, where now Robot 2 waits until Robot 1 has made it through the door and then begins. Thus, there is some amount of parallel activity. Finally, they could communicate at a very detailed level about where they will be and when, such that they can both begin moving and Robot 2 waits, at the last second, for Robot 1 to go through the door first. The time to complete both deliveries is much shorter because the robots are essentially workingin part

FIG. 3.7. Conflict removed by changing spatial behavior.

3. COO~INATIONAS DISTRIBUTED SEARCH

FIG. 3.8. Conflict removed by changing temporal aspects of a behavior.

FIG. 3.9. Temporal conflict resolution at an intermediate level of detail.

allel. But is this the best method for coordination? The amount of information exchanged to get to this solution is quite high. Moreover,the sol~tion commits robots to being in certain placesat certain times, whichthey might have trouble living up to. Perhaps coordination at more abstract levels would be more cost effective and flexible! Our simulation experime~tshave investigated these tradeoffs us cons Lesser, 1991), and we do not detail them here. Instead, let anal~icallythe role of abstraction in a distributed search to gai rf (1987) have looked at how t e search within an agent from a large problem into a numbe at ha~penswhen the smaller

DURFEE ET AL.

can be solved by different agents (see ontgomery, & Durfee, 1993, for more ations of the results summari~ed here). problem (such as a p l ~ n i n gproblem) that hasa sol~tionlength (such as the number of plan steps in the solution plan) of n. Let us say that an a~stractionhier~chyof height I, where the first abstraction level the initial problem of length n into n/k subproblems of lengthk. At each abstraction level above, k subproblems are abstracted into one subproblem of at point thereis one problem left length k, until the top level is reached, which of length #. If each problem is solved sequentially, then the overall com) + f ( k ) + + f- t:(n). If we assume that each lexity is: t : ( ~+k+f(k) is solved by a separateagent,thenafterthemostabstract s~~problem ,itssubproblemscanbesolve n parallel,after which oblems can be solvedin parallel,a so on. In effect, all the comple~tyexpression are eliminated, g i ~ n t:~ : is assumed constant fora given problem space, I = log, n, this yields a complexi~ which is O(log, n). Therefore, concurrent *

-=$

* * *

rfee, 1993), but which may not be met in other problems. he logarithmic speed~pis in general beyond reach, the allelismthankstoabstractioncanresult in significant

of coordinating them. For

3. COO~INATIONAS DIST~BUTEDSEARCH

8~Durfee, 1993). This is true despite the fact that many of the assumptionsfor

our more formal analysis do not hold inthis case. Among these are that the delivery tasks (randomly) assigned to different teams are generally not of equal difficulty, and that the “downward refinement property” (~acchus Yang, 1991) generally does not hold because information mustflow both up (a team captain must generate an abstraction of the team based on information from the team members) and down (after negotiating withother captains, a team captain must supply team members with the commitments that have been agreed to).

As we have seen, the use ofabstraction can be criticalfor efficiently searching for coordinated behaviors. If adjacent levels of abstraction in a hierarchical protocol provide nearly the same amount of detail, then the protocol will lead to excessive overhead becausetoo many rounds of communication will be needed.On the other hand, if the levels are too far apart, then agents might not have just the right level of abstraction to optimally model each other. Decisions about how to construct the abstraction hierarchy thus have important ramificationsto the costs and benefitsof coordination search. In the delivery domain, for example, our early experiments were base decomposing delivery paths based on the path generation algorithm. algorithm worked by line of sight, suchthat the level of abstraction between the entire delivery andthe indiv~dualrobot movements was compris series of movement vectors.For many of our experiments, this work because usually we were talking about robots moving through doors such that they approached the door, and left the door, on an angle, result in^ in the three “to the door, through the door, beyond the door” components of the plan. However, when delivery sources and destinations line through the door, or when ts are doingdeliveries in a big the “line-of-sight” methody an abstraction that is no different from that of the full delivery. That is,the “intermediate’, levelof the abstraction hierl to the most a ~etweenthe in

DURFEE ET AL.

couldovercomethepreviouslyseenlimitations of ourapproach.They allowed resolutionof conflicts at more appropriate levels than were possible before, and they provided decomposition knowledge that focused the exchange of more detailed information when that was desired. Thus, an important future avenueof research is to identify methods for automat in^ the constructionof explicit abstraction knowledge; we have shown that the hand-crafted knowledge that we supplied for the indoor delivery application workswell.Our results in thedomainofdistributedmeetingscheduling using a hierarchyof abstractions over time also showed the promise of this approach.But applyingthese techniques to arbitrary domains requires new methods for generating such knowledge. More generally, our researchin the construction and exchange of agent models at varying levelsof abstraction has helped put in concrete, computational terms the conceptual arguments about the costs and benefits of building, using, and committingto different agent models as outlined in the beginning of this section.Much work remains in various facets, including the development of abstraction hierarchies for the models, improving on the metrics fore~aluatingcoordination quality, and extending the repertoire of local search methods and distributed search protocols. Pursuing this last avenue requires a deeper understanding of the building blocks of protocols, and the motivations for constructing different kinds of protocols.

Understanding how protocols emerge among agents which are, by nature, self-interested is an ongoing concern. Articulating the “first principles’’ of protocolconstruction is importantwhendevelopingartificialagents, because the arsenalof protocols that we might provide them with might still not suffice for novel multiagent situations. Rather than falling back on standard, but inappropriate, protocols, we should provide agents with the ability to devisecommunication strategies on thefly, and these strategiesmight lead to new protocols. Our efforts along this front have begunby considering a single, rational nt that is trying to choose its next action to takein a multiagent world. viously, it should choose the action that it expects will be most to its benefit, but the consequencesof its actions (the payoffs) could well depend on the actions that other agents might be taking at the same time. Thus, to choose its best action, an agent should try to predict the actionsof others. Of course, if the others are also supposedly rational, thenwill they be going throu~hthe same process. An agent thus needs to think about what the ents expect the other agents to do. In fact, this nesting could con-

3. C O O ~ I N A ~ O ASND I S T ~ B ~ ESURCH D

tinue on, into deeper recursively nested modelsof the decision situations that agents believe the others face. We have developed methods for representing and solving such nested models, called the Recursive Modeling Method (R”; Gm~rasiewicz& Durof Rl”here, we focus on fee, 1993,1995). Rather than go through the details , RMM’s abilityto allow an agent to compute the one aspectof ~ Mnamely, expected utilityof sending a message. A message contains information, and the information can change the stateof knowledge of a recipient.In considering whether to send a message, therefore, a speaker should model the expected effects that the message will have on the hearer. In amounts to the expected changesin the nested modelsof agent turn could affect the likely choice of actionsof the hearer.A speaker will consider how the change in the hearer’s likely actions would impact its own (the speaker’s) expected payoffs. If the expected payoffs rise, the message could besent;theexpectedutility of themessageispreciselythedifference between the expected payoff after sending the message and the expected payoff before sending the message. Thisworksfineforindividualmessages,butcanitexplaindialogues between agents? For example, consider the simple dialogue of one agent asking a question, and another agent answering the question. The seemingly simple protocol rule of “respond to a question with an answer” is actually not all that simple when agent autonomy is considered. my should an agent answer a question? What isin it for the agent? Using R ” , we have shown how assigning cost to communication and computation can cause an agent that would otherwise share par mation to refrain from doing so if it believes that a hearer m know the information (Durfee, Gm~rasiewicz,& Rosenschein, asking a question amounts to an “admission of ignorance” directed toward an agent that the question-asker believes will see utility in supplyi~gthe desired information. This not only requires the question-asker to develop a nested model of the potential answerer, but also means that the questionasker must assign utility to the outcome of the complete dialogue (in this case the two messages) rather than to a single message. Of course, such reasoning about nested models of agents’ decision-~a~” ing situations, especially across a series of messages, can be quite costly. This is why it should only be done when the pr~stablishedprotocols no longer seem suitable. After performing the recursive reasoning about communication, moreover, an agent could cache the result and use feedback from the environmentto determine whether,in fact, it wasa good decision. By genera~izingthecircumstancesthattriggeredthemessagedial (S) could over time construct a new protocol. An se ~reliminaryideas and investigating whether, i they can lead to the emergenceof protocols.

DURFEE ET AL.

In this chapter, we have outlined the conceptual framework for our work on

collaboration technology and, more centrally, coordination theory, and we in terms of describing computational mechhave grounded that framework anisms that embody aspects of it. For the most part, we have kept to a fairly high level of discussion,pointingto(ratherthanpresenting)themore etailed technical treatments in favor of trying to illuminate the broader ~roblemsand solution strategies in this niche of coordination theory. Our in dynamic systems understanding ofhowcoordinationisaccomplished comprised of complex agents is still far from complete, and the computaso far seemsto raise more tional embodiment of the understanding we have q~estionsthan it answers. Nonetheless, our efforts have shown that it is pos sible to develop quantitative models of multiagent application domains that allow~principled decisions about strategies to be by made surrogate agents. As surrogate agents need to interact in more diverse ways with different kindsof otheragents,moreover,thekinds of informationtheyshould exchange, and the different commitments they should make, will require to bemuchmoreflexibleaboutmodelingthemselvesandothers. action becomes crucial, therefore, when agents engage in a distributed search for sufficiently coordinated behaviors. The rules for conducting this in terms of how abstractions are formulated and how information is ,are currently rather primitive, but the foundationis in place for a riety of mechanisms to be explored, including exploring how new rotocols and other rulesof interactio~might emerge for new applications.

t, inc~uding input

C h a p m ~D. , (Sept. 19 .What are plansfor?Technical Report AI ftp://publications.ai.mit.edu/ai-publications/1000.1499/AIM-1050A,ps Bacchus, E , & Yang, Q. (August 1991). The downward refinement property. In ~ ~ e of e the ~ i ~ s ~welfth int~rnational Joint Conference on Artificial Intell~ence (pp. 286-292), Sydney, Australia.

3. COO~I~ATION AS DIST~BUTEDSEARCH

Conry, S. E., Kuwabara, IC, Lesser, V. R, & Meyer, R A. (Dec. 1991). Multistage negotiation for distributed constraint satisfaction. IEEE Trans on Systems, Man, and Qbernetics, 21(6), 162-1477. Corkill, D. D., & Lesser, V. R. (August 1983). The use of meta-level control for coordinationin a distributed problem solving network. In Proceedings of the Eighth International Joint Conference on Artificial Intelligence(pp. 748-7543), Karlsrugh, Germany. Dijkstra, E. W. (Sept. 1965). Solution of a problem in concurrent programming control.Communications of the ACM, 8(5), 569. Durfee, E.H., & Lesser V. R. (Sept. 1991). Partial global planning: A coordination framework for distributed hypothesis formation. IEiZ Trans. on Systems,Man, and Qbernetics, 21(5), 1167-1183, Durfee, E. H.,& Montgomery, T. A, (Dec. 1991). Coordination as distributed search in a hierarchical behavior space.IEEE Transactions on Systems, Man, andQbernetics,21(6), 1363-1378. Durfee, E. H. (1993). Organisations, plans, and schedules: An interd~sciplinary perspective on coordinating AI systems. Journal ofIntelligent Systems(Special Issue onthe Social Contextof Intelligent Systems), 3(2-4). Durfee, E. H., Damouth, D., Huber, M., Montgome~,T. A., & Sen, S. (1994). The search for coordination: ~ o w l e d g ~ u i d eabstraction d and search in a hierarchical behavior space. In C. Castelfranchi & E, Werner (Eds.),Artificial social systems(pp. 187-206). New York SpringerVerlag. Durfee, E. H., Gmytrasiewicz,P. J., & Rosenschein, J.S. (July 1994). The utilityof embedded communications: Toward the emergence of protocols. In Proceedings of the T~jrteenthInternational Dis~ibutedArtificial Intelligence Workshop (pp. 85-93), Seattle, WA. Ephrati, E.,& Rosenschein, J. S. (July 1994). Divide and conquerin multi-agent planning. In Proceedings of the T w e l National ~~ Conference on Artificiai Intelligence (pp. 375-380), Seattle,WA. Gmytrasiewicz, P. J., & Durfee, E. H. (1993). Toward a theory of honesty and trust among communicating autonomous agents.Group Decision and Negotiation,2,237-258 (Special issue on Distributed Artificial Intelligence). of recurGmytrasiewicz, P. J., & Durfee, E. H,(June 1995). A rigorous, operational formalization sive modeling. In Proceedings of theInternationalConferenceonMulti-AgentSystems (pp. 125-132), San Francisco, CA. Hoare, C. A. R.(Au~ust1978). Commun~cating sequential processes. Comm~n~carions of the ACM, 21(8), 666-667. Huber, M. J., & Durfee, E. H.(June 1995). Deciding whento commit to action during observationbased coordination.In Proceedings ofthe International Conference on Multi-Agent Systems (pp. 163-169), San Francisco, CA. Johansen, R. (1988). Groupware: Com~utersupport for business teams. New York Free Press. Knoblock, C. A. (July 1991). Search reduction in hierarchical problem solving.In Pr~eedings of the Ninth NationalConferenceon Artificial Intelligence,Anaheim, CA. Korf, R. E. (1987). Planningas search: A qualitative approach.A~ificial Intelligence, 33(1), 65-88. Co~~unicationstheofACM, Maes, P. (1994). Agents that reduce work and information overload. 37(7), 30-40. Maione, T.W., Grant, K. R., Turbak, F. A.,Brobst, S. A., & Cohen, M. D.(1987). Intelligent information- har ring systems. Communications of the ACM, 30(5), 390-402. Mitchell, T., Caruana, R., Freitag, D., McDermott, J., & Zabowski, D. (1994). hperience with a learning personalassistant. Communications of the A C . , 37(7), 80-91. Montgomery,T. A., & Durfee, E. H.(1993). Search reductionin hierarchical distributed problem solving. Group Decision and Negotiation, 2, 301-317 (Special issue on Distributed Artificial Intelligence). Nii, H.P. (Summer 1986). Blackboard systems: The blackboard model of problem so~vingand the n 38-53, e , evolution of blackboard architectures.~ M ~ ~ i7(2), Sen, S., & Durfee, E. H.(November 1991). A formal studyof distributed meeting scheduling: Preliminary results, In ACM conference on O ~ a n ~ a t i o n aComputing l Systems (pp. 55-68), Atlanta, GA.

DURFEE ET AL. Sen, S., & Durfee, E. H.(February 1992). A formal analysis of communication and commitment in distributed meeting scheduling.In ork king Papers of the 11th International~orkshopon Dis~ibuted A~i~cial lntell~ence (pp. 333-342), Glen Arbor,MI. Sen, S. (October 1993). ~ e d i c t i ~ ~ aindcon~act-based eo~s d ~ ~ i b u t e d s c h e d Doctoral u l i ~ . thesis, University of Michig~, Sen, S., & Durfee, E.H.(March 1994a). On the design of an adaptive meeting scheduler. In Proceedings of the TenthIEEE Conference on~ A p p l i c a ~ o nSan s , Antonio, TX. Sen, S., & Durfee, E. H.(1994b). The role of commitmentin cooperative negotiation. I ~ t e r ~ a t i o ~ a1Journal ofIntell~entand C~perative Information Systems,3(1), 67-81. Shoham, Y., & Tennenholtz, M. (July 1992). On the synthesis of useful social laws for artificial In ~ ~ e e dof ithe~ Tenth s ~ational Confere~ce on ArtiFient societies (preliminary report). cia1 Intell~ence,San Jose, CA. Smith, R. G. (Dec. 1980). The contract net protocol. IEEE Transactions on Computers, 29(12), 1104-1113. Tambe, M.(June 1995). Recursive agent and agent-group tracking in a real-time dynamic environment. In ~ ~ e e d i nof g the s Inte~ationalConference on~ u l t i - ~ eSystems nt (pp. 368-375), San Francisco,CA.

C H A P T E R

Sarit Kraus

Bar llan University, Rarnat Gan University of Maryland, College Park Jonathan Wilkenfeld

University of Maryland, College Park This chapter reports on ongoing work in the development of a strategic model of negotiation. The model is applied to a range of environments in which negotiation can be utilized to resolve conflicts in two domains: access to resources, and cooperative task performance. These situations include both intelligent agentsin multiagent environments, and human applications in multilateral crisis situations. OurfocusonDistributedArtificialIntelligence (DM) hasfacilitatedthejoining of two usually distinct areas: the development of autonomousagents in computer science, and crisis bargaining in decision science/political science.In both cases, negotiationsare employed in situationswheretimeiscriticaltosomeorallparties, resources are scarce, and tasks may require cooperative behavior. The significance of this work lies in its ability to provide a method whereby time spent on negotiation among the agents is minimized, and agreements involving mutual benefit are assured. The work reported here represents progress made over 5 years by a computer scientist and a political scientist, whose collaborative work has been very much in the traditionof the initiative on Coordination Theory and Collaborativetechno lo^ of the ~ationalScience Foundation. This interdisciplinary collaboration has resultedin new insights in both realms,in ~ a y s which would not have been possible within the confines of single disciplin~s, The discussion begins witha presentation of the strategic modelof negotiation. Thisis followed bya discussion of several applicationsof the model

KRAUS AND ~ L ~ E N F E L D

in both robotics and human environments. We conclude witha review of the relevant literature and compareit withour approach.

The strategic model of negotiation is a modification of Rubinstein’s Alternative Offers model (Rubinstein, 1982, 1985). We utilize modified definitions aus, Wilkenfeld, and Zlotkin (1995).’We assume here that there is a set of n a ents denoted by that negotiates the division of M units of a present a formal definitionof an agreement.

~~~. An agreement is a tuple (S,, ...,S,), here si E M. si is agent i’s portion of the resource or task. The negotiationprocedure is as follows. Theagents can take actions only :one at certain times in the ordered set 3= (0, 1,2 ...).In each period t E 3 of the agents, say i, proposes an agreementto one of the other agents. That agent ( j ) either accepts the offer(chooses Yes) or rejects it (chooses No), or opts out of the negotiation (chooses Opt). Theother agents which neither received nor made an offer may optofout the negotiation (chooses Opt), or they can choose not to do anything (chooses Nop).We require that the ents always make offers in the same order, and an agent’s negotiation n general is any function fromthe history of the negotiations to its e. assume that each agent has preferences or utility functions (denoted by U)over agreements reachedat various pointsin time, and for opting out at ~ariouspoints in time. Thet i ~preferences e and the preferences between agreements and opting outare the driving forceof the model. In situations ion, we will assume that there is a finite set of agent their ~apa~ilities (e.g., their disk space, computagreements), These characteri~ationsproduce a difeach typeof agent. We also assumethat in such sitagent has some probabilistic beliefs about the other agents’ e set of possible types of agents is known, andthe pe is also known, the agent has incomplete infortype of the other agent(s).

‘See Osborne and Rubinstein (1990) for a detailed reviewof the bar~aininggame of Alternative Offers.

4. STRATEGIC NEGOTIATION

equilibriumz at any stage of the negotiation. That is, in each stage of the negotiation, assuming that an agent follows P.E. the strategy, the other agent has nostrategy which is better than to follow its own P.E. strategy. §ubgame perfect equilibrium is essentially a backward induction argument, using the rationality of agents at each stage of the gameto decide whata good choice & Werlang, 1988).So, if there is a (unique) is and then rolling backward (Tan perfect equilibrium, and if it is known that an agent is designed to use this strategy, no agent will prefer to use a strategy other than this onein each stage of the negotiations. When there is incomplete information, there is no proper subgame. In such situations we use the concept of seque~tiffl equili~riu~ (§.E.) instead (Kreps & Wilson, 1982).A sequential equilibrium includesa system of beliefs (Kraus et al., 1995),in addition to a profile of strategies (as in P.E.). At each negotiation stept, the strategy for agenti is optimal given its current belief (at step t) and its opponent’s possible strategies in the S.E. Each agent’s belief (about its opponent’s type) is consistent with the history of the negotiation. That is, the agents’ belief may change over time, but only consistent in a negQtiation interaction has with the history.We assume that each agent an initial probability belief about its opponent’s type. The strategic modelof negotiation has been developedas away of reaching mutual benefit, while avoiding costly and time~onsuming interactions that might increase the overhead of coordination. That is, we have provided a model in which agents can avoid spending too much time negotiating an agreement and thereforeare better able to stick to a timetable~atisfying for their goals. In the processof developing and specifying the strategic model of negotiation, we have examined bilateral negotiation as well as multiagent negotiation, single encounters and multiple encounters, situations characterized by complete as wellas incomplete information, and the differing impact of time on the payoffs of the participants (Kraus, 1997; Kraus & ~llkenfeld,1993b; Kraus et al., 1995; Kraus& Zlotkin, 1992; Schwartz & Kraus, 1997). Although some combinations of these factors can result in minor delays, the model nevertheless revealsan important capacity for reaching agreementin early periods of the strategic negotiation. In all the situations that we consider the strategic-negotiation protocols that we suggest satisfy the following criteria: symmetrical distribution (no centralunit or agent), simplicity (process simple and efficient), stability (distinguishable equilibrium point) and satisfiaor task completed).If there is bility or accessibility (access to the resource complete information, conflict is always avoided. ’A pair of strategies (G, T) is a Nash Equilibrium if, given T, no strategy of Agent 1 results in an outcome that Agent 1 prefers to the outcome generated by (G, T) and similarly for Agent 2 given G.

KRAUS AND ~LKENFELD

We examinetheproblems of resourceallocationandtaskdistribution amongautonomousagentswhichcanbenefitfromsharing a common resource or distributing a set of common tasks. The first situation we examine consists of agents that must share a joint resource; the resource can only be usedby one agent ata time. Underthese circumstances,anagreementis aschedulethatdividesusage of the resource among the agents? It is desirable that all agents that need the resource will gain access to it. Typical examples of joint resources include: communication lines, printers, disks, bridges, road junctions, fresh water, clean air,so and forth. Other work in the DAI community dealing with the resource allocation problem includes a multistage negotiation protocol that is useful for cooperatively resolving resource allocation conflicts arising in distributed networks of semi-autonomousproblemsolvingnodes(Conry,Meyer, & Lesser, 1988; Kuwabara & Lesser, 1989); tradeoffs in resource allocation and real-time performance, with mechanisms for resource allocation based on the criticality of tasks (Lesser, Pavlin,& Durfee, 1988); resource allocation using specialist “sponsor”agents(Kornfeld & Hewitt, 1981); and resourceallocationvia resource pricing (Chandrasekn,1981). particu~arlygood example of a shared resource is a communication satellite, due to the high costs associated with its launching and continued maintenance. The only waya company may be able to afford access to such a resourceis by sharing it withother companies. Even competing companies may find it mutually beneficialto participate in such a joint project. When a common resource is to be shared, a coordination mechanism will be required to manageusageof the resource. Discussion of such a coordination mechanism should begin (and may even conclude) prior to discussion of other technical aspectsof the joint project. There is always a cost associated with the elapsed time between when the agent needs the resource and when it actually gains access, and the will cost depend onthe internal stateof the agent (i.e., task load, disk space, so andon). In the following sections, we demonstrate the applicationof the strategic modelof negotiationto the resource allocation problem in several different settings. 3 0 ~ model r is also applicable in the case where the resource itself can actually be divided ents. This case does notdiffersignificantly from the case where only the resource usage time can be divided.

4. STRATEGIC NEGOT~TION

We assume that there is one agent currently using a resource(A symbol-

izing “access”), anda second agent that also wants to use(Ws~bolizing it ‘~aiting”), W wishes to gain access to the resource during the next M time

periods. The agents begin a negotiation process on the re-division of the resource between them.A continues to use the resource as the negotiation proceeds. We now present a number of assumptions concerning the utility functions of the agents. We denote A’s utility functionby UAand W’s by Uw.First we assume that the least preferred outcome for both agents is disagreement isag agreement)^

.

x 3}:Ui(x)>U~~isagreement).

The next two conditions (Al),(A2) concern the behavior of the utility function U’on agreements reachedin different time periods.~ondition(AI) requires that among agreements reached in the same period, Agent i prefers larger portionsof the resource. U‘((r, t)) >U’@, ?)f

e: For all t E 3 f r , S E Sand i E

For agreements thatare reached within the same time period, each agent prefers to geta larger portionof the resource. A prefers to The next assumptionstates that although time is valuableW, to prolong the negotiation.

e: Each Agent i E {W,A} has a number cisuch that:Vt,, t2 E .3:S, S E S, U’((s, ti)) 2Ui((S,ti)) iff (si +- c&,)2(i$ +- cif2), where c, 0. We assume that agentA gains over time (cA>0) and that Agent W

loses over time(c, 0), that is, Agent Wprefers to obtain any given number of units sooner rather than later, while AgentA prefers to obtain any given number of units later rather than sooner.

41n Kraus et al. (1995) we assume thatfor Agent A disagreement is the most preferred outcome. If negotiat~ngis costly to A also, disagreement is the worst outcome also for A. The results of the negotiations when assumptions Al-A5 hold do not dependon A’s attitude toward disagreement. 5For all S E S and i E A, S, is Agent i’s portion of the resource. Throughoutthe rest of the is written first. chapter, A’s portion in an agreement

KRAUS AND ~ I L ~ N F E L D

e: for any t E .t3:Uw((Opt,t)) >Uw((opt, and UA((Opt, t)) UA((Opt,t + 1)).

ost of o p ~ out n ~over

t + 1))

W prefers opting out sooner rather than later, and A always prefers opting out later rather than sooner. This is because A gains over time while Wloses A would neuer opt out.In the worst case, A would over time. For this reason, prefer for AgentW to opt outin the next period. Even though Agent A is gaining over time, an agreementwill be reached (after a finite number of periods). The reason for this is that Agent W can threaten to opt out at any given time. Thisthreat is the driving forceof the negotiation process toward an agreement. If there is some agreementS that A prefers at time t over W’s opting out in the next period t + 1, then it may agree to S. So the main factor that plays a role in reaching an agreement is the worst agreement for AgentWin a given period t which is still preferableto Wthan opting out in time periodt. We will denote this agreementby SKt E S. If Agent A will not agreeto such an agreement, its opponent has no other choice but to opt out. Agent A’s loss from opting out is greater than thatof W. This is because A’s session (of using the resource) is interrupted while in progress. This is stated in the following assumption.

.

e for reer ern en^ For every t E 3U~((SK~, t)) > Uw((SK~’,t + l)), Uw((Opt,t)) >Uw(($K~l, t + l)), and if 2: 0 then UA(($Kt, t)) > UA((Opt,t + 1)) and UA((SK~’, t + 1)) >UA((sKt, t)).

$F

If there are some agreements that Agent W prefers over opting out, then Agent A also prefers at least oneof those agreements over W’s opting out even in the next period. An additional assumption is necessary to ensure that an agreement is possible at least in the first period. That is,there is an agreement that both agents prefer over opting out.

.

ement: For all i,j E R, Ui((Sjpo,0)) 2 Ui(($i~o, 0)) the worst agreement for Agent i in Period 0 which is stillbetter than opting out. Silois

We consider two cases. In the first case, an agent loses less per period In the secwhile waiting than it can gain per period while usingresource. the ond situation, an agent loses more while waiting for the resource than it can gain while using the resource. For this second agent, sharing a resourcewith others is not efficient. Therefore, it prefers to have its own private resource if possible. However,in some cases the agents don’t have any choice to but share a resource (like a road junction or another expensive resource).

4. STRATEGIC NEGOTIATION

We first consider the case where Wloses less over time than A can gain over if it is big enough, it is possibleto find an time. In such a case for any offer, offer in the future that will be better for both sides (Le, both agents have positive total gain). Although it might appear that such an assumption will cause long delaysin reaching an agreement, we have proven that, in fact, the delay will be at most one period since W may opt out. However, because better agreements for both parties can be foundin the future, the agreement that is reached is notPareto Optimalover time. The intuition behind this proof isas follows. First, an agreement won't be achieved when it isW's turn to make an offer andthere is still the possibiliA is gainingover time and would ty of an agreementin the next period, since like to delay reaching an agreement. Second, A will always offer WatTime T an agreement that will prevent Wfrom opting out (i.e., Because Wwodt get anythingbetter in the future and because it loses over time, itwill accept such an offer.So, if Wis the first agent to make an offer (thisa isreasonable assumption becauseA is using the resource and does not have a motive to start thenegotiations), the agreementwill be reached in the second period with S%'. If A is the first one, agreementwill be reached in the first period with Agent W's losses ouer time The second case considers the situation where in period t are greater than Agent A 'Sgains: In this model, for any agreement E 3there is no other agreementin the future that both agents will prefer over this agreement. On the other hand, if an agreementS in period t i s small enough, one can find an agreement in a period earlier than t which both agents prefer overS in Period t. According to our assumptions, this property will cause the agents to reach an agreement in the first period. We demonstrate the case whereW loses less over time thanA can gain over time,in the following example.'

Ie I. The United States anderm many have embarked on a joint seientific mission to Mars involving separate mobile labs launched from a single shu~lein orbit around the planet. Euch country has contracts with a number of companies for the conduct of ~periments.These ~perimentswere preprop rammed prior to launch. A~angementswere made prior to launch for the sharing of some equipment to avoid duplication and excess weght on the mission. Instructions to begin each ~perimentmust be sent from Earth. The US.antenna was damaged during landing, and itis expected that communications between the United States and its lab on Mars will be down for repairs forone day (4440 minutes^ of the planned5day duration of the mission. The United States can use a weaker and less reliable backup line, but this 6Additional conditions consider the borderlinecases: S^y+IcwlI:M. 7Examples 1,2, and 3 also appear in Kraus et al.(1995).

KRAUS AND W I L K E N ~ L ~

inuolues diuerting this line from other costly space experiments, and thus the expense of using this line is uery h&h to the United States. The United States would like to share use of the German line during the l-day period so that it can conduct its planned research p r ~ a mOnly . one research group can use the line at a time, and that line will be in use for the entire duration of the particular ~periment. A negotiation ensues between the two labs ouer diuision of use of the German line, during which time the Germans haue sole access to the line, and the United States cannot conduct anyof its ~periments(except by use of the uery expensive backup). By prearrangement, the Germans are usingofsome the US. e q u i ~ ment for their ~periments,and are~aining$5,000 per minute. Whereas the Germans cannot conduct any of their ~perimentswithout someUS. equipment, the unite^ States could conduct some of its experi~entswithout German equipment. The United States is losing $3,000 per minute during the period in which they must rely on their backup communications line. An ~ e e m e nbetween t theUnited States and Germany to share the communications line will result$l,in 000a gain per period minute) for each group. If an agreement on sharing the line is not reached, the United States can threaten to opt out of the a~angement.In this case, the United States will be able to conduct a small portion of its experiments by using ofall its equipment and no German equip~ent,and by using the backupcom~unicationsline. The overall US. gain will be $550,000, but it will lose$l 000 per any minuteof the negotiation. If the United States opts out, theerm mans will not be able to continue their ~periments ~w~thout the US. equip men^ and their gain will be res~ictedto whateuer they had gained at the point United the States opted out. If the Germans to pay theUnited States~100,000 for use of the US.equipopt out, they will need ment up to that~oint.Note that the Germans play the role ofA (a~achedto the communication line) and the United States plays the roleof W (waiting for the line). A dollar is the smallest unit of currency in this example.

or mall^ U'((S, t))= 1 ~ s+ 5000t, g Ug((Optu,t))= 5 ~ tUg((Opt,, , t)) 5ooOt -1OOOOO U"((S,t)) = 1000~~ -3000t, Uu((Optu, t)) = 550000 -lOOOt, U'((0pt~ t)) = -1ooot * M = 1440 *

=L:

*

The Germansprefer any agreement ouer o p t i out. ~

32'= 551 + 2t

l ) with An ~ e e m e nwill t be reached in the second period (Period It should be noted that there ~are e e m e n t in s the re that both fer ouer reaching the agreement($ 7,553) in the second period, This is because the Germans gain more over time than the United States loses ouer time. For example, the agreement ($879,561) in the fourth time period (Period 3) is better

4, STRATEGIC NEGOTIATION

for both agents than ( $ ~ 7 , 5 5in ~ the ) second time period. The problem is that there isno way that the United States can be sure that when the fourth time period arrives, the German will offer the^ ($879,561). In that time period the Gerthe States fkm optmans need to offeronly ($885,555)in order to prevent ~nited ing out, andthey don’t haveany motivation to offermore.

In some automated agent interaction situations, the agents do not have complete information about each other and about the en~ronment.The incompleteness of information may be the result of different factors. For example, an agent may hide its actions from the other agents; an agent may not be able to explore the environment and may be missing information about theen~ronment; the resources that are available to one agent are not known to the others; or one agent is not familiar with its opponent’s utility function. We have considered situations when the agents have incomplete information about eachother’sutility functions (Kraus& Wilkenfeld, 1993b; Kraus et al., 1995).Thesituationofincompleteinformationbecomesevenmore when we can expect recurring encounters between the same two agents can use information obtainedin one encounterin a subsequent one.So, we will assume that there is a setof agents whose members negotiate with each other, from time to time, on sharing a resource.However, we still assume that in a given period of time no more than two agents need the same resource. When there is an overlap between the time segments in which two agents need the same resource, these agents will be involved in a negotiation process. We assume that there isa finite setof types of agents in the environment (Type = {l,.. &}),and each has a different utility function that depends on its resource usage.The different distributionsof resource usage among the agents can be due to different tasks that the agentsare executing, or differif alltheagents are communications entconfigurations.Forexample, servers that share a common communication line then an agent that has smaller disk spacewill use the resource more frequently than an agent that has larger disk space. M e n j E Type plays the roleof W (waiting for the resource) we denote it by 4,and when it plays the role of A (attached to the resource) we denote it byAr We also assume that each agent maintains some probability belief concerning its opponent’s type.We denote by , !@ where i E {A, W}, j E Type, i’s probability of its opponent being of Type j. We assume thatV i {A, W} Z;=, @I = 1. This probability belief changes over time. We denote by the set of all possible configurations of agents (i.e., R = {W,, W2, ...,Wk, A,,

,A&

’

KRAUS AND ~ L K E ~ F E L D

The assumptions of the previous section are still valid,but we add additional requirements to (A2). We assume that each agent has utility with a Aii E constant costof Time c,..We will concentrate on the cases where Agent ~ y gains ~ eover time (cA,. 0) and Agent rY; loses over time(c5 c 0). Agent W prefers any given portion of the resource sooner rather than later, while Agent A prefers any given portionof the resource later rather than sooner. The exact values, cAiand cy are private information. That is, Agent A, knows of k its private gaincAi,but may not know cwi,although it knows that it is one values. Furthermore, we will consider the situation where it is common knowl1 c .. c I cwlI c 1 CA, I c I CA, I. That is, Agent W, edgethat Icw,I c loses less than agentW; loses while waiting for the resource. Agent A, also gains less thanA, while using the resource. Both agents lose less while waiting than they can gain while using the resource. We also assume that for any timeperiod t, Syk' SF-*tf , . S?*'. That is, W, is more willing to optout (compared with reaching an agreement) than 1w;. In addition, we assume that S F + Icwl S M.We show that in situations that satisfy our conditions, thereexistsasequence of strategiesthatare in sequentialequilibrium which have the following properties. +

general, will notoffer Ai anythingbetterthanitspossibleutility out in the next iteration, with the additionof W;.lsloss over time (i.e.,STf+' + Icy I), since it can always wait until the next iteration and opt out. 2. If Ai receivessuchanoffer,itmayrealizethatitsopponentisno -1strongerthanType i. Thisisbecause astrongeragent will notoffer Icy 1. If it realizes that its opponent is at most of Type i, it can wait until the Since Aigains more thanW loses Icy I c cA,, next iteration and offerWSw*f+3. Aiwill prefer it over W's current offer. 3. It is better forH( to "pretend"'t0 be the strongest type by offering only gy+3+ I c 1. This offer will be rejected, but in this way itwill not convey any information about W's type toAi. 4. In the next iteration, if W;.gets an offer that is worth less than opting out (lessthan ) ,thenit will reallyoptout.Thatis, if Aioffers andWis I it will opt out. This is because Wiknows thatin any of a stronger type than future time periodt : it will not reach an agreement better than $wnf+3, and would prefer to opt out over that possibility. Therefore, Aj computes its for any i E Type, according to its beliefs, expected utility for offeringSwt1f+3, and offers the one where its expected utility is maximal. After receiving the offer, Wwill either accept it or opt out according to its type. 1. In

'Men we say that an Agent B pretends to be Agent C in a given situation, we mean that Agent B will take the same actionas Agent C, regardless of its expected utility in this situation.

4. STRATEGICNEGOTIATION

This means that, if the agents use sequential equilibrium strategies, the ne~otiationwill end in the second iteration and the agents will reach an agreement in this period with high probability. The exact pro~abilityand the details of the a~reementdepend on AgentAis initial belief.

We returnto the exampleof the missionto Mars. Suppose that ne6s{agents) onMarsknowstheexactdetails of thecontractsthe other has with companies. There are two possibilities for the contracts: high (h) 'Shold is h, then their u t i l i ~funcand low{l). I f the typeof contracts the German l. They gain$5)000 per minute during the tions are similar to those of ~xample negotiation and gain $l)OOOper minute when they share the line with the United States. I f the ~nitedStates also holds contractsof Type h, then their utility $~,OOO ~nctionsare also similar to those of ample 1. The United States loses per minute during the negotiation p e r i ~and gains $1,000 per ~ i n u t ewhen with the er mans. I f the United States opts out the overall US. 0)OOO) but theywill also lose $l)OOOper minuted u r i the ~ nego tiation. ~ o ~ e ifvthee German ~ con~actsare of Type l, then they only gain $4,000per minute while using the line by themselves. The United States losses while negotiating if their con~actsare Type l are only $2,000 per minute. But if the United States opts out, their overall gain is only $450)000. They still negotiatefor the usage of the er man line in thenext 24 hours {i.e.,M = 1440) &om the time the negotiation ends. ~ o r m a let l ~S E S, t E 3

l l

Ugh((s,t)) = 1000sg+ 5O0Ot Ugh((OptU, t)) = 5000t Ugh((Optg,t)) = 5000t 1000

-

LIg/((s,t)) = 1 0 0 0 + ~ 4000t ~ Ugt((OptU,t)) = 4000t UBI(cOpt,,t)) = 4ot -l o 0 0

Let us assume that erm many playing the roleofA ) is of v p e h and the ~ ~ i t e d States { p l a y i ~the role of W) is of Type 1. We denote them byg,, and ut. We will consider two cases.S u p p ~ g,, e believes that with probability0.5 its opponent is of Type h and with probability0.5 its opponent isof Type l (Le., tplf=0.5 and 4,"= 0.5). According to our results (Krauset al., 1995)) and in the First p e r i ~U, ) will pretend to beof Type h and will offerg,, (g?' + I c,,, 1, S:h9' -Ic",, 1) = ($890,550). g,, will reject theoffer,In the second time period, g,, compares betweenof~e~ing W ( $ ~ ~ , 5 5which ~ ) , will be accepted by both types, and offering($988,452) which l, but aftersuch anoffer,if W is of Type h, it will opt out. will be accepted by Type

I

KRAUS AND ~ L K E N ~ L D

Because its ~ p e c t e d u t ifrom l i ~ o f f e r i($~~,553) ~ is h ~ h e rit, makes this offer, which is accepted by U,. ow ever, suppose gh believes only withprobubili~0.1 its opponent isof Type h, and withprobabili~0.9 its opponent isof Type 1. The behavior of U,in the first ~ e ris i similar ~ to theprev~ouscase, It pretends to beh. ow ever, in the second p e r i ~gh’s , ~ p e c t utili~ ~ d from ($988,452) is h ~ h e than r (~~~,55 and 3 )therefore it makes this offerto which is accepted by U,.

Our second situation involves a set of autonomous agents which needs to satisfy a common goal. In order to satisfy any goal, costly actions must be taken. In addition,anagentcannotsatisfythegoalwithoutreachingan a~reementwith the other agents. Each of the agents wants to minimize its costs (i.e., prefersto do as little as possible). We note that even though the agents have the same goal (under our simplified assumptions), there is actuallya conflict of interests. The agents try to reach an agreement over the division of labor. We.assume that eachstep of the negotiation takes time, and the agents have preferences for reaching agreements in differenttimeperiods.Researchonthetaskdistribution problem in the area of Distributed Problem Solving systems includes Davis and Smith’s work on the Contract Net (Smith & Davis, 1983); Cammarata et al.’s work on strategies of cooperation that are needed for groups to solve shared tasks effectively in the context of collision avoidance in air traffic ~Cammarata,~ ~ r t h u&rSteeb, , 1983); Lesser and Erman’s model of a distributedinterpretationsystemthatisabletofunctioneffectivelyeven thoughprocessingnodeshaveinconsistentandincompleteinformation (Lesser & Erman, 1980); and Carver et al.’swork on agents with sophisticatedmodelsthatsupportcomplexanddynamicinteractionsbetweenthe agents (Carver, Cvetanovic,& Lesser, 1991). The “delivery domain” is a good example of the task distribution problem (Fischer & Kuhn, 1993; Sandholm, 1993;Wellman, 1992;Zlotkin & Rosenschein, 1993). A group of delivery companies can reduce their overall and individualdelivery costs by coordinating their deliveries. Each delivery requirement is a single task. Delivery coordination is actually the exchanging of tasks. One company, for example, that needs to make a deliveryfrom A to B and a delivery fromC to D can execute other deliveries from A to E3 with no extra cost. Therefore,it may agree to exchange itsC to D delivery for another A to B delivery. The application of the strategic model to the delivery domain allows multiple delivery companies to reach an efficient agreement on task distribution without delay, that will be mutually beneficial.

4. STRATEGIC NEGOTIATION

.

There are three agents that are responsible for the delivery of electronic newsle~ersof two different companies. The delivery is done by phone (either by famachines or electronic mail). The expensesof the agents depend only on the number of phone calls. heref fore, if there is a subscriber who subscribes to all companies'n e w s l e ~ ethe ~ , three newsletters maybe delivered to it by one of the agents for the priceofonly one phone call. The agents negotiate over the distribution of the commonsubsc~ptions.Each ofthe agents can opt out ofthe ne~o~ations and deliver allof its own newsle~ersby itself We assume that there are three electronic newsletters (Nl, N2, and N3) that aredeliveredbyseparatedeliveryservices (Dl,D2, and D3). The payment a ~ a ~ e m e ntot sthe delivery services are as follows. The publisherof NI pays Dl $200per d e l i v e ~ of one edition; similar^ the publisherofN2 pays D2 $225, and the publisher ofN3 pays D3 $250. Each delivery to a given subscriber (i.e., a hone call to this subscriber's server) also cost $1, and each agent loses $1 for each time period that it is late in m a k i a~ delivery. There areM subscribers with subscriptions to all newsletters (i.e., NI, N2, and N3), and there are substantial savings to a delivery service if one ofthe agents can deliver all newsle~ersto the same subscribers.If an agreement amongDl, D2, and D3 for joint~eliveriesto the M joint subscribe^ is reached, then the publisher ofN3 will pay D3 only $215 per deliveryof an edition, and in such an event the publisherofNI will pay Dl $170, and the p~blisher N2 of will pay D2 $200 (the lower prices reflect the fact that there are c o m p e t i ~ advertisers in the two newsle~ers,and consequentl~ their jointd e l i v e ~ may detractfiom the sales impactof each newsletter). They must pay $1 per phone call to the server, and will lose $2 for any time spentin ne~otiation.Notice that in this example, only the number of phone calls to the subscribers made by ~elivery a agent plays a role in its payments and not the distribution of the restof the subscribersb e ~ e e the n other two agents.

orm mall^

- - -

U'((Opt, t)) = 200 M t and U'((s, t)) = 170 -S, -2t U2((0pt, t)) = 225 M -t and fl((s, t)) = 200 -s2-2t U3((Opt, t)) = 250 M t and U3((s, t)) = 215 -s3-2t Suppose M = 100. Then S: = 69 -t, $f.' = 74 -t and = 64 -t. Note that forall i E {l,2,3}, 3'" is not unique in this case.

$2'

T=36 and it is D3's turn to make an offerin the timep e r i prior ~ to i": In this period, Dl is w i l l i ~to deliver up to 34 newsle~ers,if anagreementwill be reached (i.e., S)% = 34)and D2 is w i l l i ~ to deliver up to 39 n e w s l e ~ (i.e., e~ = S^$" = 39). So, = (34,39,253. It is easy to compute, that whenever it is Dl 's~ r n to makean offer(t is divided by31, x'= (31,38,31), when it is D2's turn to make an offer,x' = (3536,291 and when it isD3's turn to make anoffer(prior to time period 351, x' = (33,40, 27). Therefore, in the first time period (01, Dl will o#er (31,38,31), and its opponents will accept its offer.

KRAUS AND ~

L

~

N

~

L

D

Given the enormous volumeof data storedin modern information retrieval systems, searching a database of documents requires vast computational reso.wces.To meet these computational demands, various researchers have developed parallel information retrieval systems (e.g., Aalbersberg & Sijstermans, 1991; Frieder & Siegelmann, 1991; Stone, 1987)). As efficient exploitation of parallelism demands fast access to documents, data organization and la cement significantly affect the total processing time. We assume a distributed database environment,in which sites are owned by different independent agents. Thus,a good response time for queries and, in particular, a good allocation,are measured subjectivelyby the different sites, and no one a fixed database global costis defined.In addition, we do not want to assume but rather a dynamic one.At one end, new documents are arriving all the time, and at the other end, the subjectsof the queries to a specific siteare also changing dynamically; thus the agents may benefit from dynamic allocation of documents. The strategic model can be applied asa dynamic data placement stratefor distributed memory and distributed l/O multicomputers, where the sites are owned by different agents. Both,existing and new documents will be swapped among the agents avianegot~ation, where each agent measures its priorities using locai estimation functions and local learning/prediction algorithms. An example of an information center where our techniques may be useful is the Earth Science Data and Information System (ESDIS) which a segment is of the Earth Observing System.

TheEarthScienceDataandInformationSystemiscomprised of several earth science data centers, eachof which includes large amounts of earth observation data from NASA and other correlative research missions in upper atmosphere, atmospheric dynamics, and global biosphere (Hughes SIX Corporation, 1994). Each data center provides access to, manipulation of, or distributionof datasets (including supporting information and expertise) for a widecommuni^ of users. Data centers also provide selection and replication of data and needed documentation and, often, the generation of user tailored data products. The search for data can be done using the ESDIS V0 Information Management System (IMS).In this system, from one user interface, a scientist is able tosearchtheGlobalChangeMasterDirectoryfordirectoryinformation NASA and ~~A-affiliatabout datasets, search and locate datasets at seven

4. STRATEGIC NEGOTIATION

edDataArchiveCenters(DAACs),viewsamplebrowsegranulesgfrom by the DAACs. datasets, and order data products offered The ESDIS V0 IMS system consists ofan IMS Server at each DAAC that responds to requests for services from theV0 IMS Client, the user interface (Hong, 1994). EachESDIS V0 DAAC isindependentandimplementsits dataset holdings independentlyof any otherDAAC. However, the messages that are passed between the V0 CIientandthe V0 IMS Servers describe datasets as various parameter valid values. Each dataset is definedin the V0 IMS system with different values for the parametersof source, sensor,geophysical, and dataset name. The minimal field that must be specified to be a valid V0 IMS search is one of the following fields: sensor, geophysical, or dataset name, and geographical area. A prototype version of the V0 IMS Client was releasedin August, 1994 to the scientific community. Currently the allocationof documents to theESDIS centers is done statically. Each center storesand maintains databy specific subjects orby specific collection sources. When a user in a given data center needs a document that is stored at a different data center, the document is transferred to the user's local machine. Becausethesystemwasreleasedtothescientific co~mun~ty quite recently, there is still not enough information on its performance. But it is already clear that one of the bottlenecksof ESDIS is the network and the volume of large docum~ntsthat need to be transferred (mainlyusing F"). We have developed a dynamic strategic-negotiation protocol for document allocation.We assume that each data center will have some estimation of which scientists intend to use services its in the near future and what their main interests are.Using this information it is usefulstore to the related documents as close as possible to the location where will theybe needed most frequently, and thus whena user asks for specific documents, the data tenter will be able to provide them much faster than in the static approach. However, these preferences may conflict with preferences of other datatenters. We propose to resolve these conflicts through strategic negotiation, and show that using this method, theservers have simple and stable negotiationstrategiesthatresult in efficientagreementswithoutdelays (Schwart~& Kraus, 199'7). browse Is a representation of a dataset or data granule used to prescreen data as an aid to selection priorto ordering. For example:a browse image might bea reduced resolution version of a single channel from a multichannel instrument. A granule is the smallest aggregation of data which is independently managed (i.e., described, inventoried, retrievable). Granules may be managed as logical granulesor physical granules.

KRAUS AND ~ L ~ N F ~ L D

In pa~icular,we prove that in the case of simultaneous responses to an offer, there are a large number of possible stable negotiation strategy profiles. These strategies depend on the specific settings of the environment and cannot be providedto the agentsin advance. We propose that the agents choose strategy profileswhich are optimal with respectto agiven social criterion. We proved that the problem of finding the optimal solution which maximizes the social criterion is itself NP complete, and therefore propose using a heuristic method for obtaining a near-optimal solution. We prove that such methods yield better results than the static allocation policy currently used for data allocationfor servers in distributedsystems. We develop threetypesof heuristic a~gorithms for obtaining the near-optimal solution: a backtracking orithm, a genetic algorithm, and a r ~ d o m - r e s t hill~limbin a~ We im~lementthese algorithms, and compare them for several environments. ~e found that the hill~limbingalgorithm provides,in general, the best results. In situationswhereagentshaveincompleteinformation, arevelation process was added to the protocol, after which a negotiation takesplace, as in the complete information case. Using negotiation efficiency by self-motivated servers demonstrates the benefitsof using a negotiation mechanism in real world,multia~entenvironments.

The theoretical work on a strategic model of negotiation has also been applied to a typical problemin multilateral international negotiation.In the following link it back to the sections we describe a hostage crisis scenario (Section 3.1), basic assumptionsofthestrategicmodel(Section3.2),andprovidean overviewof a decisionsupportsystemdesignedtosupport e~perimental work in negotiation theory and the training of crisis negotiators (Section 3.3).

In the field of international relations, the hostage crisis situation was chosen as a typical case of multiparty negotiation. Thescenario is based on the hypotheticalhijackingof a commercialairlinerenroutefromEurope to India and its forced landing in Pakistan. The passengers are predominantly Indian and the hijackers areknown to be Sikhs. The hijackers demand the release from Indian security prisonsof up to 800 Sikh prisoners (see & ~ilkenfeld,1990a)).*'The three parties must consider several possible ''The original specification of the model was based on a Middle East setting involvin Egypt, and Palestinian hijackers. The experimental results reported in this chapter used the India-Pakistan-Sikhmodel in order to minimize participantsbiasduringthe course ofthe experiments (see Section 3.3.1).

4. STRATEGIC NEGOTIATION

outcomes: India or Pakistan launch military operations to free the hostages; the Sikhs the hijackersblow up the plane with themselves aboard; India and negotiate a deal involving the release of security prisoners in exchange for the hostages; Pakistan and the Sikhs negotiate a safe passage agreement; and the hijackers give up. The specific issues to be negotiated are the following:

*

Number of security prisoners to be released by India in exchange for release of all of the hostages.

Indian request for logistical information to enhance the probability of success of an Indian military operation. * Indian request for assistance (or at least neutrality) during an Indian operation, to enhance the probability of success. * Indian request that Pakistan deny the hijackers press access in order to prevent them from publicizing their message. * Pakistani request for Indian assistance during a Pakistani operation,to enhance probabilityof success. * Pakistani request that India accept a Sikh offer.

*

Hijackers’ request for pressaccess to publicize their cause. * Pakistani request that the hijackers give up or reach an agreement for safe passage. * Pakistani request that the hijackers accept an Indian offer.

*

In the simulation setting, actors negotiate these issuesuntil an agreement is reached or a player opts out of the negotiations by launching a military operation (India or Pakistan) orby blowing up the plane (Sikhs). Each party to the negotiation has a set of objectives, and a certain number of utility points is associated with each (see Kraus & Wilkenfeld, 199Oa). Utility points were assignedin order to express a complex set of preferences in such a way that subtle distinctions can be made among them. In combining the range of utility points associated with each objective with the possible outcomes, a matrix is generated which yields a point output total for each outcome. We note that thesepayoff points are not utility functions (in the decision theory sense), but rather our description of the crisis. Each

KRAUS AND ~LKENFELD

player will develop hisor her setof preferences for the outcomes based on these utility points (see Doyle, 1989). Time is incorporated into the model both asa reference point for the calculation of utilities and probabilities, and as a differential factor for the three parties. In general, time works in favor of the hijackers, and against India and Pakistan. Time impacts on the probability of success and failureof an operation to free the hostages (whether it is day or night, whether there is time to train a rescue team, and so on), on publicity for the Sikh’s cause (regardlessof whether direct press accessis granted), and deterioration over time in India and Pakistan’s internal and external images. For more detail on the Hostage Crisis Simulation, see Kraus and Wilkenfeld (1990b).

We assume that there are three players: the “Initiator” of the crisis-the Sikh hijackers (Sik); the “Participant” (against itswill) in the crisis-India (Ind); and a “Third Party”-Pakistan (Pak). There exists a set of possible agreements between all possible pairs of players.” The negotiation procedure that follows Ss as described in Section 1. In the Hostage Crisis, the hijackers opt outby blowing up the plane. Indiaor Pakistan opt outby launching military operations.In this case the setof agents is defined to be: pak’j.We assume that the set Si,i, i, j E ,i ;tj includes the po ments between players i and j. We also assume that Si,]= and denote the set of all possible agreementsby S. In analyzing the Hostage Crisis case, we identified special conditions on the sets of possible agreements. In our case S,,,, = 8, that is, there is no possibleagreementbetweenIndiaandPakistanthatcanendthecrisis. There is only one possible agreement between the hijackers and Pakistan (hostages are freed and hijackersare granted free passage). We assume that India holds MSikh security prisoners and an agreement between the hijackers and Indiais a pair (ssik, sin,) where ssik,S,, E IN,s,ik 2 1 and S,, + S,, = M. That is, an agreement between the hijackers and India is the divisionof the M prisoners between them.In the hostage crisis situation M = 800 (Sikh prisonersin Indian jails). We also assume that playeri = sik, ind, pak has a utility functionU’:[[SU opt}} x 3) U isag agreement) -3I - .

We have identified several conditions that the utility function of the players in the Hostage Crisis satisfy. We assume that these conditions are known to all players. That is, we have developed a model of complete information. “In the formal model, we have concentrated on the negotiation process. We haven’t incorporated the possible actionsof Pakistan (e.g., providing information) into the model.

4. STRATEGICNEGOTIATION

'

For en ui((rsjk,

('S&,

'

sjnd),(rs&! rind)

rind),

t,

E Ssjk,jnd

and

t).

n and India prefer reaching a

orIndia/Sikhagreement) sooner rather than later. In particular, India prefers to release any givennumberofprisonerssoonerratherthan later,whilethe hijackers, through Period 10, prefer to obtain any given numberof prisoners later rather than sooner. India has a number Cind 0 such that:' For any sjnd), (rsjk, rind) E SSik,, and t2 E 2 uind((ssik, 1 ') uind((rsik,

rind), '2)

iff

(sind + cjnd

*

(rind +

'

* 4).

The hijackers, through Period10, prefer any agreement later rather than sooner. In particular, the hijackers have two constants csik >0 and 1, the tth memberof the sequence belongsto Mj if and only if it is j's turn to speak at stept, i.e. if and only if j = t mod n. We now let m'denote the message sent at step t; if i = t mod n, then we definem: "= m '.Thus m: E di[(ml,...,m'-'), e,],

(7)

where dl[.]is a subsetof M, or of 2.The initial announcementm' = m: is chosen (by Person 1) from a set dl(e,).The (set-valued) functionsdj satisfy the requirement that forevery e = (el,.. ,en)in E,a completed conversation is obtained no matter what choicesare made from thesets di[.].Call thetriple IT = ((Ml, ..,Mn), d,, ...,d,, 2)a p r o t ~ o on l E. Call the conversation m = (m',..,m', t )a ~~onuersation for e if it starts with m' E dl(e,),continues with choices that obey(v), and ends witha member of 2. Let C(n) denote the setof all nconversations for all e in E.Since the only component of e that entersi's function d, is the componente,, the set +

6. ONECONOMIES OF SCOPE IN COMMUNICATION

is a rectangle. For the neonversationm = (m',.. ,mi,z), the rectangle's proEi is jection with respect to G;(R)

= {e,E E,1 m' E d,(el);m: E di[(m',.. ,mt-'),e,] fori=tmodnandl1, C, contains two elements a;*, a;* * , each a subsetof S, such thatC,-1 = (Cl\{o*, a;* *}) U {a;* U a;* 1 ' . Given k sets Ul, ...,Uk, we shall call a t-element partitioning C of U,x x Uka rectffn@e bisection pffrtitioni~~ (rbp) if C is the tth member of a bisection sequence of rectangleswith andeachpartitioning in thatsequenceiscomposed respect to (Ul, ...,Uk)." We next prove a general lemma. Its statement uses the term ~ormal' defined in Section 5. *

Lemmff. Su~posethe ~ i e m e n set t C isanrbp of Ulx x U,.If {I,J)is a ~ a r ~ t i o of n i(1~, ...,k}, and V x JYP, W = x t,JUP, then X isnormal ~ i t h respect to (V,W). Proof The proof Q = 1 and Q = 2 .

is by induction on Q. The proposition holds trivially for

If we examine the first two members of the bisection sequence that yields a given rbp I:of V x Mi/, we see that either there is a partitioning of V, say 7) and a partition in^ of C, say {E,8,such thatE,E are, respectively, rbp's of and x or@ere is a ~artitioning{R F} of W and a par {E,C} of X such that2,X are, respectively, rbp's of x Wand x W loss of generality, we may suppose that the former statement hol

v

v

{v,

v

v

131.e.,each setin the pa~itioningis the Cartesian product of its projections withrespect to

v,,(I.,U,. #

Now suppose that the assertion of the Lemma holds for all(3with 2 (3 .Since I 2 I Q and I g I S Q, the induction hypothesissays that2 and E are

(v,

(v,

normal with respect to W) and W) respectively. That means there exists a coveringTVof whose setsare v-projectionsof sets in E,and a covering Twof Wwhose setsare W-projectionsof sets in E,such that

v

Similarly, we have coveringsTF,Tw, ofand

W , respectively, such that

~ithout loss of generality we may suppose that ~ a covering of Vwith 1 7 ' ~I + I Tg I sets, Since 2 nE= a,the collectionTFU 7 ' is each a Vrprojection of a setin C. The collection Twis a covering of W, each of whose sets is a projection of a set in C. In view of (1)-(3), we therefore have

(v

Thus C is normal with respectto W) and the induction is complete. NOW suppose that the coveringCAof EA= E' x is (i) an r 11 coverings with those twop se that the coveringC minimal among allcoverin~swith those twop r o ~ ~ r tThen i~s.

141tis not known whether the conclusion of the Lemma holds if the rectangl~overingC is a partition in^ but does not have the bisection property. No counterexample has been found.

6. ON E C O N O ~OF I ~SCOPE IN C O M M U N I C A ~ O ~

cation is best for the deterministic~onversationscenario,witheachannouncement being a binary string, when we take m ~ i m u mconversationJ,let p(J) denote completion timeas our cost measure. For a positive integer the smallest integer d such that Z d 2 J.Then our cost measures for three mechanisms with brp's EA,C,, E become p( I EAl), p ( I E ,,:)l and p (I E: I). Since the functionp is nondecreasing, wesee thatno (FA, FJ-realizing deterministic protocol for the merged organization can have a shorterm ~ i m u m completion time than a protocol which corresponds to the product partitioning obtained from the rectangles that the two organizations were using before the merger.

It is naturalto seek a regularity condition on finite mechanisms-a condition that is,in some sense, analogousto our regularity condition for continuum mechanisms-and that rules out counterexamples to the Finite Conjecture. The bisection re~uirementis not such a condition, since for a continuum mechanism the coveringt: is generally not a partitioning and yet the continuum version of the conjecture is true. Instead, the regularity condition we ~~l consider is suggestedby the situationin Fig. 6.6,where a n o n - ~ o rrectangle covering is portrayed. Oneof the rectangles in that co~ering-the rectangle indicated by broken lines-is not a 'proper' rectangle but rather consists of disjoint pieces with partsof other rectangles intruding betwee pieces. If we required the rectangles to be 'proper,' then would non-no e ruled out?If so, then what condition on the mechanisms use and the merged organization would yield proper rectangles, and hence a ,and therefore the validity of the Finite Conjecture? ing, we note that no generality is lostif we henceforth let = E3x E4be finite sets.' setx nite mechanism(in rectangle form) on the finite EA , w i t h E : = {E~ ~},we ~ ~define a t ~ ~ i ~ e n s i o n f f ~ ~ o r ~ f '?he Finite Conjecture only deals with functions 4,G which are not only finite valued but which can be tiled witha finite number of rectangles. (That rules out, e.g., the casein which El and E, are each the real line and FA equals zero whenel2 e, and one otherwise).Now suppose El and E, are continua and suppose the smallest numberof rectangles on each of which l'$ is constan: is, say, L.We can select a finite numberof points from each of those L rectangles to in such a way that (i) the obtain EA,a finite subset of EA.We can select the points, moreover, points in a given oneof the original continuum rectangles comprise: new rect_angle (a subset bf EA), and (ii) the L new Lectangles comprise a minimal tiling of FA where FA, denotes the irestriction of FA to the setEA.

a L ~ D of P C. ~ Let vAbe a n u m b e r i ~ ~ n c tfor i o nEA= Elx Ez,i.e., a one-to-one function froml? onto {l,2, ..,. },J where J A = Il? 1 .Define vs to be aJE-valued numbering function for E’, where JE = 1E’ I .Let A EE {l,2, .,,,JA},B EE {l, 2, .. .}’J For each m in M,define T, EE {(a, b)

E A x B I a = v,(eA),b = v,(es),

(eA, e’)E G,}.

Then the triple[v,, vb,{T, Im E M ) ] is a TDP of C. It is easy to check that (7, 1m E M }covers A x B, and thatT, is a rectangle in A x B: it is the product of (its A-projection) and T,”, (its B-projection). Notice, moreover, that the projectionT$ itself definesa rectanglein EA = E, x E2.If we decode the points a in, ;c‘ we obtain precisely the rectangleG: x dmin EA.Thus a QA-element covering of A defines a Q,-rectangle covering of EA=Elx Ez.Figure 6.7 provides an illustration, Now define an outcome function hAwhich assigns to each of those Q A rectangles-say to the rectangle comprising the decoded elements of rt-the action hA(m)in ZA. That outcome function together with those Q A rectangles (suitably indexed) comprise a mechanism on EA = Elx E;. If, moreover, our original (four-person) mechanism (M,C, (hA, hJ)realizes (FA,F’) on EA x then the (two-person) FA on EA. mechanism we have just constructed realizes Now suppose the setsin a covering of A x B are not only rectangles but proper recta~Lesas well (Le., each projection of every rectangle is a set of a ~ a c e npoints). t Then we can use the following correct ‘two-space’ version of the claim denoted(* *>in Section 5.

, c :

(* * *) ~onsiderA x B in !R2, where A is a ~ n i t set e ofpoints on a hor~ontaL axis and B is a ~ n i t set e of points on a verticaL axis. LetS be an !-set c o v e r i ~ofA x each of whose sets is a proper r e c t a ~ ~Then e . S is nor~aLwith respect to (A,

Before proving(* * *)we note its implications. Suppose we know that (i)

FA can be realized (tiled) withQ A rectangles in EA but not with fewer,(ii) FE

can be realized withQ’ rectangles in EBbut not with fewer. Suppose we nevertheless had a mechanism which realizes(FA,F’) on EA x EEwith Q < rectang~esand has a TDP cons is ti^ of proper recta~Les.~onsiderthe Q rectangles in the coveringof A x B given by the TDP. (* * *)tells U construct a (&-set covering of A from the A-projections anda ing of B fromtheB-projecti Q; Q: S Q (Le.,it tells usthat a situation like the example in Fi has an ‘improper’ rect h e mechanism on E’ re ruled out). Then we ,contradicting the assumedminimalities,

TOP of a ~~chanism on 4 x € , x €a x E, The mechanism:

1

S

S

FIG. 6.7. The points shown as dots are a portrayal of q . The points show as asterisks are a portrayal of aT (Note that each of those portrayals is a rectangle but nota proper rectangle). The portrayals of the other rectanglesin C are not shown.

MARSCHAK

To summarize: EAx E5has Let ussay that a finite (four-agent) mechanism on the set finite r mechanism ~ has a TDP consisting entirely of the a~acency~ r o ~if ethe proper rectangles. Suppose (* * *) is correct. Then the finite conjecture holds for four-agent mechanisms with the adjacency property. Proof“ of (* * *).Let P 1st.Suppose that the smallest coveringof A whose sets are A-projections of the rectangles in S has 4 elements. Suppose that the smallest covering of B whose sets are B-projections has4elements. Suppose S is not normal with respect(A, toB), i.e., P 1, proceed to the points H and if H = 1 proceed to the line b = H + 1. On that line againfind the tallest rectangle among the rectangles-there are again at least 4 of them-whose numbers are attached to the points on the line. Note that none of these rectangles are among those encountered in the first step (along the bottom line). Let the B-projectionof the new tallest rectangle be the second inset our cov.Proceed in this manner untilB has been covered. At each step at least 4 rectangles have been accounted for and those rectangles are never encountered at a later step. Consequently, there can be no more than P/4 steps, each steppro~dinga setin our coveringof B. By assumption,P/4 4, so we have indeed constructeda covering of B with fewer than4 sets. That concludes the proof.

have found two waysto ‘salvage’ the finite conjecture. We can impose the all the mechanisms used. That is, perhaps, analogous S conditionsin the continuum case (where we do in it uggling of many numbers into one number). Or we n property for the coverings that the mechanisms f we confine attent~on to the deterministic~onver‘‘A slightly different version of this proof appearedin Marschak and Vazirani(1991).

6, ON ECONOMIES OF SCOPE IN COM~UNICATION

sation scenario, then the merged organization can indeed do no better than to replicate what the two organizations were doing before the merger. The ’proposal’ scenario, however, seems quite plausible and is a standar scenario in the literature on continuum mechanisms. It is quite surprising that in a finite example, where the goal functionsFA,FB are not particularly ‘exotic7, the merged organization doesnot replicate the two separate ones even though it wantsto achieve exactly the same tasks as the two separate organizations. One may react to this discovery, of course, by saying ‘This just shows that number of possible proposals is not a good cost measure,’ or ‘It shows that additional elements of a mechanism’s information-processing costs (computation for example) need to be modelledas well.’ Those are natural reactions and it would certainly be useful if such comments le new modellings, for which the unexpected phenomenon disappears. On the other hand, the mechanism/messag~space model of informationboth economics and processing is the only one to have been studied inwidely by computerscience. If weseek toresolvethedecreasing-returnle incorporatingprecisemodels ofinformationprocessingintotheanager-with-privat~information’ explanation, then the mechanism/message space model has to be a serious candidate. Finite mechanisms are more realistic than continuum mechanisms. The failure of the finite replication-is-best conjecture is therefore quite disruptive, since it suggests that for certain elements of information-processing cost there may even ibe ~ c r e a sreturns. i~ To clear the decks for the T o ~ M ~ a gexplanation er it would certainly be helpful to rule out such disturb in^ counterexamples. Should we accept the somewhat artificial adjacency constraint?Or should we insist on a scenario in which mema bers conversein strict accordance witha shared protocol that never gives member more than one choice when it is that member’s turn to speak?

SupportfromNationalScience ~oundationgrant 1~9120074is gratefully acknowledged. Much of the chapter suggestions from U. Vazirani, who is absolved from any responsibility for errors orunusual interpretations.l am particularly gratefulto C. B. McGuire, who wrote a computer program that discovered the critical 16-point example in Section 5.

Calsamiglia, X., 1977, Decentralized resource allocation and increasing returns, Journalof Economic Theory. Calsamiglia, X. and A. Kirman, 1993, A unique informationally efficient and decentralized mechanism with fairoutcomes, Econometriea.

Hurwicz, L., 1977, On the dimensional requirementsof informationally decentralized Pareto-satisfactory adjustment processes, In: K. J. Arrow andL.Hurwicz, eds., Studiesin resource allocation processes (Cambridge University Press, Cambridge). Hurwicz, L., 1986, On informational decentralization and efficiencyin resource allocation mechanisms, In: S. Reiter, ed., Studiesin mathematical economics (Mathematical Association of America, Washington,DC), Hurwicz, L.and H. Weinberger, 1990, A necessary condition for decentralization and an application to intertempora~ allocation, Journal of Economic Theory. Jordan, J., 1982, The competitive process is informationally efficient uniquely, Journal of Economic Theory. Jordan, J., 1987, The informational requirements of local stability in decentralized allocation mechanisms, In: T, Groves, R. Radner and S. Reiter, eds., Information, incentives, and ecoof Minnesota Press, Minneapolis, MN). nomic mechanisms (University Lovasz, L,, 1990, Communication complexity:A survey, In: B.H, Korte, ed., Paths, flows, and VLSI layout (Springer Verlag, Berlin). Kaldor, N., 1934, The equilibrium of the firm, Economic Journal. Mount, K. and S. Reiter, 1974, The informational size of message spaces, Journal of Economic Theory. Marschak, T. and S. Reichelstein, 1993,Network mechanisms, in~ormationa~ requirements, and hierarchies,Working paper. Marschak, T.and S. Reichelstein, 1995,Communication requirements for individual agents in networks and hierarchies, In: J, Ledyard, ed., The economics of informational decentralization: Complexity, efficiency, and stability (Kluwer Publishing Company, Dordrecht). Marschak, T. andU. Vazirani, 1991, Communication costs in the performance of unrelated tasks: Continuum models and finite models, Journal of Organizational Computing. Rather, R.,1993, The economicsof decentralized information processing, Econometrica. Radner, R. and T. Van Zandt, 1992, Information processing and returns to scale, Annales d’Economie et de Statistique. Reichelstein,S. and S. Reiter, 1988, Game forms with minimal strategy spaces, Econometrica. Sato, 1981, On the informational size of messagespaces for economies with public goods, Journal of Econometric Theory.

C H A P T E R

Stanley Reiter

~ort~western University

The aim of the work reported in this chapter is to build a formal model of technology in which technological change can be studied. Technological change takes place by invention, or discovery. It proceeds by finding new ways of doing things, by finding new things to do, or new uses for things already known. These things typically involve a creative act, Something previously unknown or nonexistent becomes known or comes into existence. A formal model of the growth oftechno lo^ or knowledge musttherefore accommodate creative acts.It must do so without pretendingto be a model that predicts specific creations, inventions, or discoveries, because,to do that would bein effect to make them. That is,a model that predicts the fruits of creative actswould itself be an engine for making creative acts-for short, a Promethean machine,' A model of knowledge, including technological knowledge, and its growth is potentially useful in connection with a number of different problems in economics.Among these is the theory of economic development. There appears to be wide agreement among economists, and others, that technological change is perhaps the most important driver of economic development? This chapter does not contain a theory of economic development. The model presented isa partial model intended, likea modular component in a stereosound system, to be embedded in or linked to a more general economic model. The present model contains functions that can serve as con-

‘

nectors to a broader economic model. In that model they would be endogeare exogenous nouslydetermined,whereas in thepresentmodelthey par~eters. In economics technology is modeled in a number of different ways. In almost allof them the conceptof a commodity is used. Technology is modeled, for example, as a relation between input commodities and output com modities expressed by a function, the production function. In some cases processes of production are explicitlyintroduced in terms of variables called activity levels or process intensities? Then production possibilities are expressed by a mapping, which may be linear or nonlinear, between the space of activity levelsand the commodity space. Although there are many different models of technology appropriate to specific purposes, the simplest and perhaps most general is the production set model. The production set is a subsetof the commodity space consisting of all input-o~tput combi“pronations that economic agents know how to produce. “Commodity’’ and duction set”are primitive conceptsin economics. The commodities and the production setare regarded as exogenously given. Part 2 presents a model of technology in which technology is, as is conventional, viewed as knowledge ofhow to produce things. However, in contrast to conventional modelsof technology in economic theory, knowledge of how to produce things is represented explicitly. Technology is therefore a specialcategory ofknowledge.Commoditiesandproduction sets are derived in the present chapter from technological knowledge. Because technological knowledge depends on other knowledge, such as scienti~icknowledge, technology is embedded in a more general model of knowledge. Although that model, presented in Part 3, should logically come before the modelof technology, the intuitive, informal discussion in Part 2 of modeling technologyserves to motivate and clarify the model of knowledge presented in Part 3. owledge resides in the mind of a person. Itmay have several representations.It is modeled in this chapter as a finite subset of the setof all possible statementsin a natural language, say, English. The English language is based on a finite alphabet. That alphabet may be augmented by appending a finite number of other symbols. The collection of sentences of finite length, (i.e., finite strings of symbols from the alphabet, well-formed accordingto the rulesor usage of the language) is countably infinite.It contains a representation of every item of knowledge that can conceivably be expressed in En~lish.In this way we avoid on the one hand, models that impl~citly bound other, what might be known in the future, butis not now known, and on the introducing aggregates thatare not well-defined sets. e ~ e i at Time t is represented asa finite subsetof the The ~ n o ~ iofperson set of possible English sentences. Formally, that set is a primitive of the model; its interpretationis the knowledgein the mind of a Person i at Time t.

7. ~ O ~ E D G DISCOVERY, E , AND GROWTH

Thus, in an applicationof the model the representation of what is in the mind of the person whose knowledge is being modeled is the oftask the applier.It cannot be expected that therewill be a unique representation.In these circumstances, andin view of the use to be made of the modelin this chapter, it is desirable to have an economical, or parsimonious representation. Therefore, an equivalence relation on’ the set of sentences is introduced in Part 3, so that a knowledge set consists of equivalence classesof sentences, or of representative sentences, each representing the classof sentences equivalent to it.A representative sentence may be interpreted as an “idea.” A knowledge set may also be given additional structure. For instance, different areasof knowledge may be recognized formally. This model of knowledge is presented and discussed furtherin Part 3.4 The main focus of this investigation is onthe growth of knowledge. This takes placeby two processes,by learning fromothers, andby the discovery (creation or invention), of new knowledge. Knowledge has theproperty that its acquisitionby one person does not diminish what is available for others to acquire. This is sometimes extendedto the idea that knowledge oncein existence is a public good, freely available to all. The latter proposition is clearly not true. Although one can borrow Von Neuman’s book on quantum mechanics froma library without charge, that does not automatically transfer the knowledge represented in it into themind of the borrower. Acquiring of knowledge, evenif it is already known to others, entails the expenditure effort and resources, and also entails private acts of discovery. It is not so different in this respect from creating new knowledge. This matter is discussed furtherin Part 4. Discovery takes place when new ideas (representative sentences) are added to someone’s stock of knowledge. Whether discovery is of socially new knowledgeor of privately new knowledge,discovery typically involves a creative acton the partof the discoverer. Hence,a model of the growthof ~ o ~ l e dmust g e rest ultimately on a theory of the process(es) of creation. Part 4 presents a view of acts of creation that provides the basis for the modelof the growth of knowledge presented in the following Part Briefly summarized, that view is: i. Discovery results from interaction between an individual mind and a

body of knowledge;

ii. More knowledge facilitates discovery, but does not guarantee it; (Discoveries are almost always made by people with knowledge of the

area in which the discovery is made, for example, discoveries in game theory are usually madeby people who knowa lot of game theory-in chemistry,by chemists); iii. There is no penaltyin the formof inhibition of discovery from having more (adequately structured) knowledge;

iv. There are individual differencesin cognitive and other skills and abilities relevantto making discoveries; v. Discovery is a purposeful activity-intensity of effort and resources devoted to that activity have an effect; vi. Creative acts typically involve bringing together disparate ideas, often from different framesof reference. It seems clear thatunderstan~ing the ideas that an inventor worked with

in the process of making a particular discovery or creationis necessary to

understanding the processof coming to that discovery. Descriptions of that process for particular inventions make up muchof the content of books dealing with the history of science or technology. That sort of understanding is, of course, retrospective. But, to recapitulate a point made earlier, a theoryof discovery or creation cannot reasonably attempt to predict froma specific set of ideas the particular creation(s) theywill give rise to when subjectedto the effortof a particular person, (who may not be the theorist).If there were such a theory, and if it werea good one, thetheory would itself make the discoveries.6 Furthermore it would doso without the abilities,skills, effort, and resourcesof any person other than the theorist. Fortunately,to analyze the economic role of discovery it is not necessary to have amodel that predictsor explains specific discoveries.It is enoughto explain: i. a measure of the flow of discoveries, ii. the dependenceof that flow on economic and social parameters, and, iii. the economic effectsof a flow of discoveries, measuredin some aver-

age sense,

In addition, a model can provide ex ante for discoveries whose specific nature is unknown, for example, new processesof production or new commodities, which are to be included ex post, after their specific nature is known. Sucha model can permitexante analysis of economic consequences of discoveries. To model the growth of knowledge without attempting to make what I have called a Promethean machine, it is useful to have some concept of a~o~nt o ~ ~Innthis o ~model l e ~ ethe. amount of knowledge is a measure of the sizeof the setof sentences that constitutes knowledge. Such a measure would generally be multidimensional, correspondingto distinctions among different kinds of knowledge. The simplest case that is of a on~imensional measure. Although this is from some viewpoints evidently a gross oversimplification, it is one that connects naturally with existing models of produc-

7, ~ O ~ E D G DISCOVERY, E , AND GROWTH

tion in which technological change is considered.In those models, technological knowledge is represented by a (real) parameter, as, for example, a coefficient multipl~nga Cobb-Douglas or CES production function, and technological change is expressed as a change, usually an increase,in the value of that parameter. Theon~imensionalmeasure we use here is the number of sentences in the knowledge set being measured (after due allowance for nonuniquenessof representation), and its growth is the change over time of the number of sentences that embody knowledge. This is analogous to the on~imensionalmeasure of a heterogeneous collection of objects used in production called “amount of capital.” Although these are clearly oversimplifications, sucha measure can be useful as a starting pointin a some questions. Part 5 presents a mathematical modelof the growthof knowledge based on the view summa~ized earlier. A “preparedmind” is stimulated by a combination of ideas to conceive a new idea.In terms of sets of sentences, several subsets of knowledge come together to stimulate the conjecture and exploration of a candidate new sentence. Thus, the possibilities of cross-fertilization are given by the power setof the knowledge set being considered. But, becausea subsetof sentences can be considered to be a compound sentence, andas such an element of the knowledge set under consideration, the size of the knowledge set itself measures the number of potential cross-fertilizings. These ideas can be expressed in different mathematical structures. have at present no clear basis for choosing among them. In order to avo1 merely mathematical complesities, I give these ideas a very simple mathematical expression. Starting with an isolated person, this leads a firsttoorder in the sizeof the set representing the knowledge of linear difference equation that person at Time t. For mathematical convenience this difference equatio is replaced by the analogous differential equation, in continuous time an with the amount of knowledge treated as a continuous variable-a realnumber.Becausethesize ofknowledge sets istypically very large,andthe changes relatively small, this does not appear to be a heinous simplificati Nest I consider a community of persons engaged in research or R activities (i.e.,in the attemptto produce new knowled~e).Such a community consists of persons in communication with one another through various means. The model of an isolated person is extendedto one of an interactin community. The result is a system of differential equations that characterizes the simultaneous growthof the knowledgeof each person in the community. This system of equationsis the basis for analyzing the growth of the subset of knowledge that is technology, The growth of owle edge, and of technology, can be exponential, depending on the values of the parameters that represent ability, skill, resources, and effort applied to discovery and learning. This means that there can be

REITER

(exponentially)increasingreturns in theproduction of knowledge,and exponential growthin the amount of knowledge. Is this consistent with what can be observed? Part 5 presents some data on the growth of knowledge in computer science from 1958 to 1990. The size of the body of knowledge in that field is measured in two ways, first, the number of pages published each year from 1958 to 1990 in the fieldof computer science, and second, the number of artiin those years.Of course, theunderl~ngassumption is that cles pub~ished the number of ideas per paper, or per sentence, is on average, not too far from constant. The resultsare that knowledgeso measured grew exponenThe rate of growth is probably tially at about10%per year over that period. somewhat understated, because papers that properly belong in computer science were not found, because they were published in journals not primarily devotedto computer science, and this is more likely to be the casein later thanin earlier years. Through the introduction of a (time varying) commodity space and a (time varying) attainable production set, both definedin terms of the underlying technology, the growth of the attainable production set and its dependin Part 6, ence on the underlying parameters can be analyzed.? This is done in the context of a Leontief model of production. Exponentia~ly increasing returns in knowledge and technology translate into exponentially increasing ret~rnsin the production set.In examples analyzedin Part 6 afixed amount of the primary resource yields an exponentially increasing amountof oututs over time as knowledge grows, because the output coefficients grow exponentially~Can this result be consistent with the physical laws governing relations between matter and energy? First, growthof knowledge may lead notto growthof a given set of output subcoefficients, butto new waysof satis~ingeconomic wants via different stances and processes not now known in such a way as to remain well inside the fundamental constraints imposedby physical laws? Learning curves in manufacturingare a well-observed phenomenon usually attributed to increase of knowledge by those engagedin production. So, learning curves oughtto be derivable froma model of the growthof knowledge. Part 7 presents derivations of two forms of learning curves from the model of the growthof knowledge. As noted before, the model presented in this chapter is a partial one. Insofar as application to economic development is concerned, it is intended that it will ultimately be linked withor embedded in a model in which a general dynamic economic analysis of its implications for +economic growth or development can be carried out, andin which the effectsof instru~entsof social and economic policy can be studied. The paper by Arrow (1962) on learningbydoing and that of Lucas (1988) contain analyses to which this mode1 might be linked.In connection with development, this chapter can be viewed as an

7. ~ O ~ E D G DISCOERY, E , AND GROWTH

attempt to provide a model of the growthof knowledge and technolo can be connectedto and support such analyses.”

*

The effects of discovery or inventionon economic life come mainly through changes in technology and ultimately changes in production of goods an services. Among the many ways of representing technology used in eco-. nomic models, the production set model is perhaps the simplest and most general, In that representation, we postulatea commodity space, usuallya (finite dimensional) Euclidean space, anda subset of it, called the production set, which contains all theinput-o~tput vectors that are technologically feasible (i.e., for which there is knowledge of a process or proces§es which if carried out would produce the specified outputs from the specifie inputs).l*An act of production isa choice of a point fromthe productionse I propose to use a somewhat different modelof technology, in which the commodity space and the production set are constructed from technolo cal ~ o w l e d ~rather e , than given exogenously. To motivate this model an to clarify its interpretation it is useful to have an example of a techno1 mind. There isa wide choiceof examples available, for instance, agric steelmaking, aircraft manufacturing, pharmaceuticals. Unfortunately, these examplesare complicated. Knowledge about any one of them is described in an extensive and complex literature not readily accessible to nonspecial~sts. An example drawn from any of these unfamiliar areas of technolo be more confusing than illuminating. likely to be familiar to most p ~ o p l ~ t h e However, there is an example isthat to produce ediart of coo~ng-which,in other words, is the knowledge of how ble dishes from raw materials. For the sake of the exampleI take this body of owle edge, as of today, to be what is recorded in cookbooks nowin existence, Typically, a cookbook contains(i) recipes, and(ii) discussions, that may or may not be part ofa recipe,of matters relevant to the preparation of food. A recipe can be viewed as consisting of four parts: *

First, a description of what will be produced if the recipe is carried out, including not only the name of the product (e.g.,Cheese ~ ~ ~ e but s ) ,perhaps also other aspects of the useof the product-in this case, that it is “interesting, attractive, and not messy,” and therefore goodserve to atcocktail parties; Second, a listof ingredient§ and outputs, with the quantities specified; Third, a list of statements describing the actions to be takenin order to execute the recipe (i.e.,to carry out the actof production specifiedby the recipe);

WITER

Fourth, a statementto the effect that the recipe works. A statement to this effect in the form of the claim that "All the recipes in this book have been tested under a variety of relevant conditions," is often found in the Introduction to a cookbook. This may be thought of as a statement attachedto each recipe. Otherwise,in the absenceof an explicit claim,I take it thattestimony to the reliabilityof the contents is implic~t in publication. in economicsfocusonthelistof Thestandardmodelsoftechnology ingredients and products, whichare modeled as input-output vectors in the commodity space, and abstract from everything else." The model I use incorporates all the elements of a recipe. ~ o w l e ~of ge how to produce a product or service, not restricted to preparation of food, is thought of as embodied in recipes, Each recipe is a list of statements in a natural language, say English, that describe what is to be done,including at each step how much of each substance, equipment, or labor is involvedin thatstep.'Theingredients, substances, andproceduresinvolvedare described and identified-named-in the natural language. at ever details are necessaryin order to identify the elements required are given. The use ofnaturallanguagedescriptionspermitsdenumerablymanydistinctions among entities(substances, objects, laboriousprocedures) to bemade, without specifying them once and for allin advance. For a list of statements to constitutearecipeitmustalsocontain a description of whatuses are fulfilledby it,14 anda statement certifying that it ~ o r ~Thus, s . all four parts of a cookbook recipe are modeledas a listof English sentences,where the list ofin~redientsand products mayor may not be displayed separately, but may appear dispersedin the sentences that specify the actions to be taken.' This model of technology does notso far use the conceptof a commodity. The objects, substances, or other entities that appear in a recipe can be anything that can be given a name in the natural language. Althoughthe collection of names of entitiesin the language is a given finite set, a new name can be coined at any time to refer to a substance or entity hitherto unknown, or to makenewdistinctionsamong substances or objects, ~urthermore, there canbemorethanonenameforthesamething,and there is no assumption that the equivalence of these names is recognized. For each namedsubstance or object there is also given some way of representing its quantity, for instance by elements of an additive group, like the inte~ers,or the rational numbers. ~ifferentcookbooks may each contain a recipe for the same dish, and these recipes may have different descriptions. For example, one can find recipes for beef stew in many cookbooks. Furthermore, these recipes may not prescribe exactly the same steps in preparation, or may not specify exactly the same ingredients, or the same quantities. They may give differ-

7. ~ O ~ E D G DISCOWRY, E , AND GROWTH

ent names to the product, or may describe slightly different uses for the product. In any of these cases, the recipeswould be different. And yet the differences may be small enough that they may be consideredto be minor variations of the “same recipe.”It is natural in situations like thisto regard all these descriptions as equivalent. As already mentioned,a cookbook, and more generally a body of technological knowledge, mayalso contain statements of a general nature embodying knowledge. For example, some cookbooks discuss the chemical compositionandthemodesofactionofvariousbakingpowders.16 Althou~h statements about the action of baking powders may be common to cooking and to chemistry, it would not be hard tofind statements about chemistry that are never found in cookbooks, and to find a graduated collection of statements that are ‘between’ them in the sense that they go from statements more relevantto cooking to statements increasingly remote from it. This illustrates a situationin which the line separating technological knowledge from other areas of knowledge is somewhat arbitrary.In the present model, the certification of tested recipes makes an unambiguous distinction between technology and other knowledge. I propose to embed the model of technology in a larger modelof a body of ~ ~ o ~ and l e where ~ ~ necessary e , make distinctions among fields of knowledge as needed. The larger model would include tested recipes, but could also include partially specified recipes. For example, it might include recipes that have the first three parts, but lack the fourth. That is, they may be d as conjectured, but as yet untested recipes.It could also include nts about propertiesof entities involvedin recipes, or about entities related to such entities,or about statements about such statements, and so on. And it could include the contentsof books and journal articles on food human physiology,and’more remote areas of knowledge chosen to the useto be made of the model. Thus, technological knowledge is a special case of knowledge. The discussion of modeling technology presented in this Part can serve as a motiin Part 3. vation for a model.of more general knowledge that follows L

To begin with let theset of persons atany time, t, be a given finiteset, denoted = p,2, . ,N}.” of Q ~ e ~ M o a~t a.person “knows” can be described in natural language, as a finite collectionof sentences in that languag~English,for example.’ Let E denote the set ofEnglishsen~~~

tence (i.e., finite strings of symbols from the English alphabet), possibly augmented with a finite number of other symbols.’ all,”and “for some,” which are, The availability of quantifiers, such as “for of course, English words, makes the language rich enough to include mathematics. It also includes self-referential sentences. Thus the set undecidable propositions, becauseof Godel’s theorem. Because the use to be made of this model does not require the set of sentences to include only those thatare either me or false, or to be freeof the problemsof self-reference, thereis no need to restrict it further.% of For i E A let K‘(i,t) be a subset (i,t) is interpretedas Person i’s knowledge at TimeIt t.includes any recipes that Person i knows at t, together with any other knowledge that i has. Thus, K’( i,t) r> K‘*( i,t),

where K’*(i,t) is the setof recipes known by i at t. The set K‘(i,t) may include the rulesof logic, the rulesof the calculusof probability, or of other systems of thought. However, it is possible for Person i to know some sentences, and to know the rules of logic, andyet not to or one that know a sentence, whichis implied by what he or she does know, is logically equivalentto what he knows. Z e to~ e Here, in contrast to other possible usage, the word~ ~ o ~ refers sentences (or structures of them) that may or may not be true, or whose truth may beu n k n o ~as, well asto sentences that Personi is aware of, and to which he may attach some degree of belief. Because the credence that Person i attaches to a sentence can be expressed by a sentence, thereis no difficulty in expressing thisin the model? Let I(i,t) be an equivalence relation on the set K‘(i,t). Sentences are equivalent according to I(i,t)if Person i at t regards them as expressing the same thing. In particular, the relation I(i,t) expresses the equivalence of recipes that would beso considered in the discussionof equivalent recipesin Part 2. The relation I(i,t) can be extended from K’(i,t) to all of .One such extension results from letting the equivalence classes be singletonsE \onK’(i,t), but others might also be possible. Now, let K(i,t) = K‘(i,t)/I(i,t), namely, the quotient setof K‘(i,t) with respectto the relation I(i,t). The elements of K(i,t) are representatives of the equivalence classesof I(i,t), and

7. ~ O ~ E D G DISCOVERY, E , AND GROWTH

may be called ideas. (Here the term ideas must be understood broadly, for instance, to include the canonical or basic recipes discussed in Part 2.) Each of these may be represented by a canonical sentence that expresses the idea. From now on, an element of the knowledge set K(i,t) is understood to be a canonical sentence (i.e., a representativeanof equivalence classof sentences that express the same thing). As in the case of technology in Part 2, it is useful to have an example of knowledge in mind. Examples that have the virtue of familiarity are Economics, orGame Theory (or a subfield of specialization, such as Economic Theory). In these fields existing knowledge is for the most part written in the form of books, journalarticles, or preprints.However, the set K(i,t) is interpreted asincluding only whati has in his or her head at Time t. Thus, if Person i bought ~yerson'sbook on game theory at Time t-l and has it on his shelf unread, its contents are not included in K(i,t), unless i has acquired this knowledge throughother sources. The sets K(i,t) for i E A= have the possibilityof being organized further with many different structures. In the example of economic theory, general equilibrium theory might be distinguished from principle-agent theory. Or, a sentence that states, say, that one theorem is a special case of another, might be included in a structure consistingof relations defined on a subset of sentences in Kfi,t), expressing the depth of knowledgein an area. We introduce a measure of the amount of i's knowledge at t, denoted,

called thesize o ~ ~ ~ ~ , ~ . A natural candidate is the measure of K(i,t), which, since K(i,t) is finite, is the number of elementsin K(i,t).23*24 Then, k(i,t) = 1 K(i,t 1 .% Distinctionsof substance amongdifferentideas introduced formally. Let = {Lq,q = 42,

or knowledgecanbe

*l

be apartition of E,/I(i,t)26and let Kq(i,t) = LqnK(i,t), and let kq(i,t)= IKq(i,t)l,

for q = 42, ....

FEITER

The partitionL classifies knowledge, the elements in Lq being considered of the same kind. The sets Lq can be called areas of k n o w l e ~ eor , su6~el~s. Since K(i,t) is finite, only a finite number of the sets K,(i,t) can be nonempty. If the partition L is the finest possible, then to know that a statement (idea, or recipe)is in the set Lq' for some q' identifies that statementup to the equivalence relation I(i,t) (i.e., the statement is uniquely identified as a particular idea). In this model new knowledge consists of new statements appendedto an existing knowledge set. This involves a creative act.The next Part, Part 4, discusses the creative process,

Person i makes a d i s c o v e ~(or an ~nvention)when he addsa new sentenceor list of sentences to his knowledge, perhaps describing a new product or process of production, or anew theoremF7 Discoveryor invention involves a creative act-something that was not before a certain timeis afterwards. The aim of this Part is to develop an understanding of the act of discovery in Part 5. That model is sufficient to justify and support the model presented intended notas amodel of the creative act itself, abut model of the ~ o ~OF t h k n o w l e ~ ~ase a result of discovery, a model that can be used to analyze the economic significanceof discovery. Plato, through the mouth ofSocrates, speaking about the creative powers of poets, states that the poets create by ' !.divine power. . Since then people have soughtin different ways to understand this power. Some have studied the livesof "obviously" giftedcreators, such as Einstein, or Picasso, with the aim of detecting how they differ from people who have not made such remarkable discoveriesor creations. Although it is clear that there are indi~dualdifferences in intellectual abilities, andin creative powers, whatever they may be, it is also clear thathuman beings have created, discovered, and invented in every time and place where human beings lived. The divine power seemsto have been distributed generally to humanity. Whateverabilitiesfacilitatecreativeachievements,originality,discoveryand invention, theyare part of the human genetic endowment.^^^ It is evident that discoveries are almost always madeby people who have owle edge of the field in which the discovery is made. Most discoveries in mathematics are made by mathematicians,in chemistry by chemists, andso on. In earlier times, when technology was further removed from science by people than it appears to be now, discoveries and inventions were made without academic or professional credentials, butrarely by those who had

7.

OWLE EDGE, DISCOVERY, AND GROWTH

little knowledgeof the field to which the inventions belong. Even today many “small” discoveries, leading to improved technology, are made by those who carry out production.Of course, it is not the credentials of the person that matter, but the fact of working intensively in that field. At one time many medical discoveries were made by practicing physicians. Nowadays medical discoveries are overwhelmingly made by researchers who may be biologists, biochemists, and the like, or MDs, who are focused on research. They are more rarely made by physicians whoare exclusivelyin clinical practice, or whose activity and experience are removed from the areas which now spawn discoveries. ~ ~ a~ention e to a field play important This suggests that the~ n o ~ lofeand roles in discovery in that field. It is useful at this point to have in mind a more concrete situation. Consider the performance of a graduate student and an experienced faculty member (in economics) in reading and understandinga new paper. The stuto follow the contentsof the paper, and dent typicallywill have to work hard is likely to do so in a limited way? He can be expected to be able to summarize the contents and perhaps to supply some missing steps in the arguments, but not usually to be ableseetoand evaluate the contents in the context of the broader literature. The faculty member typicallywill absorb the contents more quickly, and will relate them to other work in the field.He will be ableto form a jud~ment about thesi~nificance of the paperin light of his knowledge of the literature. He is also more likely see to other applications of the methods used, or thin of other methods that could be used eitherorwith in place of the onesin the paper. A few years later when the student has become an experienced faculty member himself, his performancewill be very like that of the facultymember now. On the other hand,if the faculty member should chooseto read a paper in a field he is not familiar with, suchas computer science, he isl i k e Iy to have the same difficulties as the student, perhaps miti~atedby his knowledge and experiencein those aspectsof his field thatare related to the material he is reading. One important difference between the student and the faculty member is the differencein their knowledge. The knowledge ofa novice most typically consists of relatively isolated pieces rather thanof a richly interconnected inte~ratedstructure of t h o u ~ ~ t ~ 2 The task of understanding the paper involves making what are subjectively discoveries.It is notessentially different from that of making other disc0veries.3~ It is important to note that commandinga large bodyof knowle~gedoes not in itself lead to difficulties that offset the benefits of nowl ledge, provided that knowledge about the structureof knowledge is also large.% This is

RE?TER

typically the case when the knowledge under consideration is that ofan experienced specialist. Another aspect of the facility that goes with expertise and experience involves the distinction between knowing and knowing how, mentioned in Part 3, footnote 16. One aspect of knowing how to do somethingis captured by the description of the process that must be carried out to produce the specified something, as in the representation of technology by recipes. A aspect of knowing how is not so convincingly captured by verbal ntation of knowledge. Such knowledge, which might better be called skill, may not involve language at all.% The student may also have lesser skills than the more experienced faculty member. The model presentedin Part 5 deals with both aspectsof knowledge. h kno~ledgeexists in the mind ofan individual, it is rare that a single person operatesin isolation. It is typically the case that an individual is part of a community of those with overlapping interests in a subfield of knowledge. The community usesa number of means of formal and informal communication to share knowledge. It seems clear that knowledge of persons in a conmunity grows more rapidly than doesofthat persons working in isolation. oreo over, the product of persons who function in isolation often shows signsof that isolation?6

~adamardstates, “Indeed, it is obvious that invention or discovery, beinit mathematics or a n ~ h e relse, e takes placeby combining ideas.”37 This conception is not unique to Hadamard, but is frequently,if not universally, mentioned by students of creativity. Koestler,in his book The Act of Creatio~~ makes it the central element in his understanding of creation. In his view, normal thought takes place within a frame of reference, or an associative context, a type of logic, universeof discourse, or a particular code or matrix. In ordinary life we may use many different frames of reference, usually one at atine, switching from oneto another accordingto the situation. According to Koestler, creating involves bringing together otherwiseinde~~ndent frames of reference. Koestler gives this process the name 6is~ia~o~~’ The term may be understoodto i ~ e n teither i ~ a propertyof the process of creation, or a property of the creative p r ~ ~ cIt tis. possible, as Perkins argued, to hold the view that the creative product does involve the joining of different frames of reference, while maintaining that ordinary mental processes are capable of and do accompli§h the work of bisociation (Le., there is no special processof bisociation)~’ ringing together ideas from different dom~ns, or combining different frames of reference, suggests that the potential for discovery or for creative acts affordedby a given knowledge base is related to the setof c o ~ 6 i ~ a ~ o ~

7. ~ O ~ E D G DISCOVERY, E , AND GROWTH

of ideas orof frames of reference in that knowledge base.In this way a state of knowledge can be saidto generate a setof potential discoveries-discoveries waiting to happen. Indeed, the history of science providesmany examples of roughlysimultaneousandindependentdiscovery of thesame thing?l It is not necessary togo deeply here into the processesby which an act of creation or discovery comes about, (even if, as is doubtless not the case, to model those processthey werefully known), because the aim here is not es as such, but rather to focus on certain salient characteristics of them that providethebasisforaquantitativemodel of thegrowth of including the knowledge we call techno lo^, and to enable us to assess its economic significance. Oneof these salient ideas is that the knowledge base itself contains the seeds of discoveries “waiting to happen.” These seeds are in the form of combinations of ideas thatwill sooner or later stimulate~ar e ~ a r e ~to conceive the new idea and make the new discovery. In this context a prepared mind has four properties: ~~~~

it is one in command of the relevant knowledge base; it is one that has sufficient command of the skills relevant to the task at hand; it is one that is intensely focused on the specific knowledgein question; and, it is one thatis suitably disposed to the discovery to be made?2 The first three properties donot require further comment, but perhaps the fourth does. The actof discovery is an interaction between elementsof knowledge of different kinds. Someare of the kind that make up in the field of specialization, and othersare rather more general ideas that can be seenas integral and perhaps unchangeable aspects of the personality of the in~estigator?~ “Thus, discovery or creation of new know just a matterof “ripeness,”of discoveries waiting to happen, bu involve a contribution from the discoverer(s) that goes beyond and effort? re cases of important discoveries made by a mind to seefrom a d~ff~rent p

that it is the person

REITER

things from oneor another kind of unconventional viewpoint. And this is not necessarily due to ignorance. In most cases the product of the unique perspective is dismissed, and does not contribute to knowledgein a significant way. In some cases the investigators are dismissed as cranks, usually correctly. But in rare cases discoveriesof value are made. How should this be modeled? In this chapter, thesephenomena are part of what is meant by a ”prepared mind” (i.e., a property of the individual distinct from, but interacting with, the body of knowledge he she or commands). Because the focus here is not on the individual discoverer, these considerations can enter the modelin Part 5 in a way that does not require going in detail into the mental processes involved, A related consideration is the roleof chance in discovery. It is evident, and has been much discussed both in general terms:5 and with reference to specific examples of serendipitous discoveries, such as Fleming’s discovery ofpenicillin,that chance plays a role in individual discoveries. However, the model presented in Part 5 is not stochastic. It could be modified, as indicatedin Part 5, so as to make it stochastic. This has not been done, mainly because the use made of the model here avoids technical complexities associatedwith a stochasticmodel, so in the interestsof clarity and simplicity stochastic elements areomitted. Beyond this, although chance no doubt plays a role in discovery,it is not so clear what that role is and how it should enter in a formal analysis, Is the role of chance more significantthannoise in determining whether aparticulardiscoveryis made, orwhen it is made, orwho makes it?* The model presentedin Part 5 doesnotaddress these questions, but ratherisintended to facilitate analysis of the economic consequences of an ongoing process of discovery, In any case, that model of discovery can be modified to include stochastic elements. Discoveries differin importance. Some, such as the theory of relativity, or the discovery of the mode of action of the genetic code, appear to change existing knowledge in a fundamental way, while others appear to be small additions or modifications. The model presented in Part 5 does not make a formal distinction between “big” and “small” discoveries. The need to make this distinction depends on the use to be made of the model. For some purposes the model will have to be given more structure, but for present purposes one can assume that the flow of discoveries consists a “typical” of mixture of big and small ones:’ The model presented in Part 5 attempts to formalize the considerations iscussed in this Part. To summarizethese considerations, they are: 1. Discovery grows out of the existing structure of knowledge; 2. It does so as the resultof the application of effort,resources, skill,and

talent by individualsto the existing structure of knowledge;

7. ~ O ~ E D G DISCOVERY, E , AND GROWTH

3. Individuals do not operatein isolation, but existin communities. They

learnfromoneanother;theirinteractionfacilitatesthegrowth knowledge of all of them; 4. Discovery proceeds by combining ideas.

of

dynamThe main aim of Part5 is to make these considerations the basisa for ic model of the growthof knowledge, of technology, and of production possibilities.

We begin with a classof models of the g r o ~ ofh knowledge in the caseof an isolated person,i, as follows. Recall from Part 3 the partition

I(i,t), and recall that K9(i,t ) = L9nK(i, t )

and in Lqbeing of the same classifies knowledge, the elements kind. The sets Lqcan be calledareas of ~ R o ~ i eor~se~, ~ ~ Since e l ~ sK(i,t) . is

finite, only a finite number of the sets Kq(i,t) can be nonem~ty. Then, kq(i,t+l)-kq(i,t)

gives the changein the (amount of) kno~ledgeof person i in the subject area in the Periodt to + t 1. A model that at Time t determined the values of the variables kq(i,t~~) would t~ereby predict the numberof discoveries in the subfieldg. If the par/I(t) were the finest possible, then the model woul discoveries up to I~quiva~ence. If the sets Lq were fine enou~h,say in the extreme case sin~letons,the ~eterminationof kq(i,t+l)amounts to predict in^ a specific discovery. There

is a gradation of specificity from this Promethean theory at one extreme, to the other extreme at which Q = 1 and discoveries are counted without making any distinctions basedon substance. For the usesto which the model is putin this chapter, the model with Q = 1, or 2 is sufficient. Furthermore, discoveries are treated as deterministic. The situation may be likenedto modeling the occurrenceof fires. Thereare fires of many kinds, It maybe for certain purposes necessaryto disting~ish chemical fires from forest fires, and both from house fires. Nevertheless, for other purposes it may be desirable to ignore distinctions among types of fires, and to treat the totalnumber of undifferentiated fires per year as if it were deterministic, perhaps interpreting that number as the expectednumber of fires that would come out aof stochasticmodel. And it may be possible to relate thatnumber usefully to certain characteristicsof the preexisting situations that tend to give rise to fires, in spite of the impossibility of predicting individual fires. The motivation for the choices made here is twofold. First, to explore a model from which a one~imensiona~ measure of the growth of owle edge can arise, and second, to keep the technicalitiesof the model as simple as pos§i~le, w~ile a l l o ~ the n ~ effect of the g r o ~ h of knowled~eon technolouction possibilities to be analyzed, turn now to the formal model of the growth of knowledge.

in Part 4 seesdiscoveryastheresult

of an interaction n, the agent of discovery, and a bodyof knowledge she or

made by the agent himself, such as decisions to invest in the

a(i,t) = (a~(i,t),a~(i,t))for i

E;

ere a i,t) is to be i~terpretedas a measure of ilities. Theseinclude the

7.

OWLE EDGE, DISCOKRY,AND GROWTH

predispose someoneto make a particular kind of discovery as discussed in Part 3.2. The parameter a,(i,t) represents the level of skill that Person i has attained at Time t. Creative abilities may,course, of be no different from cognitive abilities generally?'

of effort put forth by Person i at Time t; to be interpreted as the intensity

to be interpreted as the resources i has available to use att; and finally, d( i,t) = f(e( i,t), r( i,t), a( i,t))for t = 1,2,...

(4)

where d( i,t) is to be interpreted as a measurei'soffertili~ or ~ ~ ~ ~ in c ~ i discovering. It is assumed that the functionf is nonnegative and isincreasin~in all its ar~uments.That is, moreeffortincreases i's fertility in discovery,more resources also does,and more skill and ability do too.50 Choice ofthe functionse( i,t)and r (i,t) would in general depend on incentives of Person i to expend effort and resourcesin the attempt increase his These incentiveswould in turn depend at le urns to suchknowledge,such as rep orking conditions in the case of an academ In general,thesefunctions thecase ofaninvestor in gies,ordecisionfunctiout in order to keep the m h n o ~ othe ~ , demand side i fore, the elements necessary make to decisions about howto a and resources to discovery and learning are not endogenous.

(i,t) = f(e(i,t), a(i,t)).

WITER

Hadamard's observation previously referredto that"itis obvious that invention or discovery .. takes place by combining ideas,'' is one among many expressions of thesamenotion,Indeed, if there is an~hinguniversally agreed upon by students of creative activity it is this idea. Valery's statement of the idea seems particularly apt.52 According to Valery, the discoverer plays two roles, one as knower of his bodyof knowledge, in which capacity he makes up combinations of ideas which he presents to himself in his second role, thatof the one who recognizes the value of what is before him. This idea suggests that the products of discovery or creative activity are the resultof an interaction between the potentialdiscoverer and his knowledge, in which potential discoveries are generated in the mind of the discoverer by combinations of existing ideas, In this view the body of knowledge can be seen as a population (of ideas) that breeds new ideas, and requiring the interventionof a discovererto bring them to awareness. This idea suggests a model in which every combination of elements of K(i,t) is a potential stimulant for, or seed of, a new idea in the mind of i. Therefore the numberof subsets of K(i,t) is the numberof opportunities for discovery presentedto Person i at Time t. Then, Personi's fertility parameter, determined by the intensity of effort and resources devotedto discovery, and by i's skill and ability, determine the yieldof discoveries from the potential ones. Modeling this would seemto require thatco~binationsof subsets make up the breeding population (i.e., subsets of the powerset of K(i,t)), However, the power set of K(i,t) is determined by K(i,t) itself and so is its size. Indeed, as waspointedout earlier,asubset of K(i,t)canbeuniquely described by a compound sentence (i.e., an element of K(i,t)). Thus, theset (i,t) itself can be used in place of its power set, and considered to be the "breeding population.^ Theuseof the size of K(i,t) instead of that of its power set amountsto a (nonlinear) changeof units. The largerthe the more combinations of statements it permits, and therefore th the opportunity for a combination to stimulate a new ideain the mi son i,% The more person i's mi more the willhis be pro~ucti~ity in discovery from Because, as discussedin Pa ery from more owle edge, ship between knowledge and discovery is increasing in knowleausemoreeffort, or moreresources, skill or ability produces more discovery afrom given body of owle edge, the relations~ipbetween discovery and the parameters that represent these characteristicsis also increasingin the parameterd (i,t)." Perhaps the simplest, but not the only, mathematica~ expression of these considerations is the following.

7. ~ O ~ D G DISCOVERY, E , AND GROWTH

Let K(t) denote some body of knowledge and d(t) the parameter measur-

ing the fertility or productivity of a potential discoverer who commands the

knowledge K(t) att, Let k(t) be the sizeof K(t). Then, the equation k(tt-l) -k(t) = d(t)k(t), t

= 42,

.,

(5)

describes the growthof knowledge over time that the process of discovery described before would generate. Thus, the growthof knowledge is the result of twofactors, one is the number of "breeding parent" statements in the populationof statements constituting the relevant body of knowledgeat t, and the other ais measure, namely d(t)) of the discoverer's fertility or producti~tyin finding something new and int~restingamong the potentially new statements or ideas "waitingto be discovered."55 Equation (5) is a difference equation in discrete variables, and as such somewhat inconvenient to work with. Theset K(t) would typically be very large, and the differencein the size over a short intervalof time comparativelysmall. Itis a convenientandperhapsacceptableidealization to replace (5) by its continuous analog, in which the variables k(t) and t are continuous.a Here k(t) isa continuous measureof the sizeof K(t). Then, the differential equation analogousto (5) is h(t) = ~(t)k(t), t

20.

(6)

where

This is an ordinary different~al equation with variable coefficient. Since the focus here is on the basic ideas underlying the model it is desirable to carry out the analysisin the simplest mathematical setting in which it makes sense. For this reason, suppose that fort,all d(t) = d(0) = d. Then equation(6) has the form i(t) = ~ ~ ( t ) . his familiar differential equation for k(t) has the solution k(t) = C exp(5t).

REITER

In the continuous model withQ >1, the time derivatlve

replaces the time difference k,(i,t+l) -kq(i,t), and givesthe change in the knowledge (in isolation)of person i in the qfhsubject area in the period tto t + 1. It will nowbe assumedthat for eachi E A and allt, thepartition of E/I~i,t)contains only a finite number,Q, of subsets. Then

K( i,t) = K,( i,t) uK,( i,t) U ...U$ ( i,t), and, because the setsKq(i,t) are pairwise disjoint, k(i,t) = Ckq( i,t). The counterpart of equation (6) applied to these knowledge sets is,

where the coefficient dqr(I$) represents, for r= g, i's fertility indiscovery arisfrom a~~ocating effort, resources andskills relevant to the subfield g, and ate ability, and, for # r g, d, (i,t) represents serendipitous discovery arisin from effort devoted to r that results in findin something in g. The same t) a p ~ e a r on s the right in every row, because any ideasthat i ~ o w ats t combine to suggest a discoveryin any subfield. The effect of itsinterests cation of effort and resources that i makes among the fields of 1, . .,Q are expressed by the matrix of coefficients in e purposes it would be desirable to preserve the among subfields whenthere is more than one person. But, because of (g), equation (9) implies

R(i, t) = 6t), k(i, (i, t) where

ing taken over all g and r. Therefore, the distinction^ am pear, and, when is it as the coefficients are c ame as (6') with k(t) =

(10)

7. ~ O ~ E D G DISCOVERY, E , AND GROWTH

For simplicity, unless otherwise stated, it is assumed that Q = 1. This amountstomakingnodistinctionsamong thediscoveriesthe model attempts to predict. If Q > 1 is assumed, then the functions k(i,t) must be replaced by the vectorof functions (kl(i,t),...,kQ(i,t)), and the coefficients d (i,t) by d,x i,t). When d(i,t) is a step function, representing one or more changes in the allocation of effort and resources to discovery, perhaps as a result of policy decisions, the solutionof (10) is a piecing together of exponentials at rates determined by the effort and resources i devotes to discovery, and on i’s skill and ability. A change in the values of these parameters results in a change in the valueof d (i,t>,and therefore those changes have the effect of changing therate atwhich i’s knowledge grows.If, for example,i decided to reduce to zero the effort devoted to exploring K( i,t), then i’s knowledge would cease to grow alt~gether?~ If any of the parameters were to change from time to time, thenso would d(i) and the solution would bea concatenation of exponentials with different ratesof growth, corresponding to the changes in d (i,t). Although knowledge is always knowledge heldby some person, discovery in isolation from the knowledgein the and invention generally do not occur minds of others. There is typically a community of persons who communicate their ideas and results among themselvesby various means.The existence of this community enriches knowledge and accelerates the process of discovery.58 be the set of per ns in the community under consideration. Each person in the community has access to the knowledge of others, through publications or various forms of direct communication. But knowledge in the minds of others, or written down in papers, does not enter into the mind of Person i without effort. Person i therefore must allocate his effort between acquiring knowledge from others and working to discover new knowledg~ between reading papers and writing them. Let e’( ii) denote the efforti devotes to discovery, and e2(ij) the effort given toacquirin~nowl ledge from j. Then, for i,j in d(ij) = f(e1(i,j),e2(ij),r1(ij),

r2(ij),a(i)),

REITER

is the parameter measuringi’s productivity in discovery, when j = i, and in acquiring knowledge fromj otherwise. It is assumed, similarly to what was assumed in the case ofan isolated i and j in { l , .., person, that5’ for all where not alld(ii) = 0. It is sometimes convenientto assume that for alli, d(ii) 0. When Q > 1 (Le., when more than one subfield of knowledge is distinuished), the coefficientsd(ij) are replaced by Equation (lo), characterizing the growth of i’s knowledge in isolation, must be modified to take accountof the fact thati’s knowledge can growin two ways, the first by discovery and the second by learning from others. Thus, for i = 4 2 , ...,N,where N is the number of peoplein A,60 d(l,t)=~(ll)k(l,t)+.*.+~(l~)k(~,t)

d(2, t ) =:&2l)k(l, t ) +

+~

( 2 ~ )t )~ ( ~ ,

d(~,t)=~(~l)k(l,t)+‘.~+~(~~k(~,t).

Equation (12) can be written more compactly in matrix form as’ k ( t ) =M ( t ) ,

where

There are several different processesof interaction among peoplein the community thatare represented by equations (12) or (13). These include the follo~in special ~ cases.

7. KNOWEDGE, DISCOVERY, AND GROWTH

First, suppose that each member of the community publishes or otherwise communicatesto the others knowledge that she or he regards as new. ~e may suppose that i?k(i,t)/at is the amo~ntof new knowledge Person i chooses to add to his knowledge set at TimeIf t.Person i should encounter at t some piece of knowledge already in K(i,t), then it would not be added to K(i,t). If a person's judgmentof what is new is a competent judgment, then, at leastto a first appro~imation, these time derivatives measure what is put into the communication network among the agents at Time t. Each agent then extracts from those inputs knowledge added to hisor her own knowledge setat Time t. This leadsto the following equations. k(1,t) = Sf(ll)k(l,t)+S*(lZ)k(Z,t)~+~'(lN)R(N,t) k(2,t) = 8(21)R(l,t>+ Sf(22)~(2,t) + S'(2N)k(N, t ) d(N,t)= S'(Nl)k(l,t)+ + SI(NN -l)R(N -1,t) +~I(NN)~(N,t).

This is a linear system thatin vector form is AfR(t)= Sk(t),

where

and

In general, D' is invertible. Hence, the system reduces to R(t) = P k ( t ) ,

where

MITER

is the. inverse of D‘,with the ith row multiplied by d(ii), i = 1, ...,N. The entries in k ’ are functions of the d(ij) defined at thebeginning of this Part, andhence are functions of theparametersrepresentingability,skill, resources and effort. Anotherspecialcase of equationssystem (12) or (13) of interest is obtained if for eachi and j # i, d(ij) = d(i).

Still another special case particularly relevant to economic development is that in which the knowledge of the agents is nested. That is, everything that is known to the least owle edge able agent is also known to the next least knowledgeable one and so on to the agent whose knowledge atset time t includes all the others. Without loss of generality let them be

In this case, one view of the processof transfer of knowledge leads to the conclusion that the matrixL) in equation (13) is a (lower) triangular matrix, all of whose entriesare nonnegative, and whose diagonal elements are d( ii), i = 1, ...,N, which may be assumed to be positive and distinct.In this viewi’s access to j’s knowledge leads to discoveries just as i’s access to i’s knowledge set does,but perhaps with different yields. A refinement of this view is that person i only works on that part of K(j,t) that is not already in K(@). Then, under the assumption that know~edge is nested, R(1,t) = S ( l l ) k ( l , t ) ) ~ ( 2 , t )+) S(22)k(Z,t) d(2,t) = S ( Z l ) ~ ( l , tR ( ~ , t=) ~ ( ~ l ) k ( l, k(2,t) t) + ~ ( 2 2 ~ ~ (Zk(3,t)) , t ) +* + S ( ~ ~-1,t) ~ kk ( ~( , t~) ) , *

ich can also be writtenas, k ( t ) =A ~ k ( t ) ,

here

7. ~ O ~ E D GDISCOVERY, E, AND GROWTH

A"

fall)

S(21)

S(31)

0 0 S(22) -S(21) 0 S(32) -S(31) S(33)-S(32)

\S(Nl) S(N2) -S(N1)

i

0

i i

o o

)

.

S(NN)-S ( N ~-1))

**

Here there is a boundary condition, namely that k(j+l,t)-k(j,t) 20. The set K(i,t)\K(i+l,t) contains what i knows att, but i+l does not know at t. If we assume that effort, and resources of i devoted to learning from Person 1 is always at leastas productive as the same effort and resources allocated to learning from any other person, where i f: 1, then

fy

=1:

... 1) 0 S(21) S(22)- S(21) 0 ~~

S(N1)

0i

0i

0

0

~ ( N-S(N1) ~ )

~

In this case i does not allocate different amountsof resources and effort to acquir~ng knowledge from different people. Because everything that i can learn from othersi can learn from Person1, it is equivalent under the state assumption for i to learn everything from Person 1 or to learn separately items that Personsi-l, i-2, ...,1, know thati does not.As before, these equations are valid for k(1,t)-k(j,t) 20 for all j = 2, . ,N. Note that if

then Agent j can use (j 1) on the entire set K(l,t),including pose therefore that for all j = 2, ...,N, If the dyna~ics are escribed by equation (1 ,and if d(l1) >0, then there is at least one eigenvalue that is real and positive.

MITER

We now examine the solutions of (13) more explicitly, under theass^^^ tion thatD has N linearly independent eigenuectors. Under that assumption, the solutions(13), to or (14), or (14) when (15) holds, consist of linear combinationsof pure exponentials. Let the (N linearly independent) eigenvectorsof D be Let S be the matrix whose columnsare the eigenvectors ofD. Then the matrix L,given by. where,

is the diagonal matrix whose (diagonal) elements are the eigenvaluesof D.62 Under these conditions (i.e., when (16) is valid), the vector ~ifferential equation (13) has the following general solution.

and Thus, the general solutionaislinear combinationof pure ex~onentials,~ the constantsci are determined by the initial conditions, that is, ci = S"k(0).

Note thatwhendistinctionsamongsubfields takes the form givenin equation (18)

are madeequation (12)

7, ~ O ~ E D GDISCOVERY, E , ANDGROWTH

which can be written k(t) = Ak(t), are NQ vectors. Assumptions where K is anNQ c1 NQ matrix and k(t) and k(t) about the structure of can be imposed in order to address specific questions about the interaction among different subfields. r~~ation of

nowl led^^

Consider an economyin which the research sector consists of a setA of people partitio~edinto two groups, the first denoted A,, consisting of N, people, and a second denoted4,consisting of NZpeople. Forthe entire group cons~stingof N=N,+N,

people, when no distinctions are made among subfields, equation (13) of Part 5 is the system k(l,t>=~ ( l l ) ~ ( l , t ) + . * . + ~ ( l N ) k ( N , t ) k(Nl,t)=~(Nll)k(l,t)+***+~(N~N)~(N,t) k(NI+18)

=W

I + ,

w w ) + '+ ~(N~+lN)k(~,t)

~(N,t)= ~(~l)~(l,t)+.**+~(NN)k(N,t) of differential equations governing the growth of knowledge. Under certain assumptions stated below, this system can be put in the form

REITER

where

That is, equations(20) describe the behaviorof the first sectorand the second sector as aggregates.This equation system is derived from(19) as follows. Suppose that fori,j E A , , d$"fori~A~,j~A~, d$t

6(ij)=

d$"fori~A,, jEAl, slfori, jeA2.

Then, N

Nl

j=l

N J=Nl +l

j=l

N j=1

j = N I +l

That is to say, each person learns equally effectively from the inothers each roup, though perhaps differently from thosein the other group than from those in his own group. It follows that, 4

i=Nl +l

7.

OWLE EDGE, DISCOVERY, AND GROWTH

Let

Then, for

the system of equations(21) can be written (in vector form) as

where

The effort, resources and perhaps also skill and ability that an individual devotes to learning and discovery are likely to vary over his/her lifetime. Most likely the resulting time path of d(i j,t) is one that is concave, perhaps an parabola opening downward. However, it is plausible thatthe average fertility of a community of researchers is constant over time as young ones enter and old ones retire, leaving the total number constant. Therefore, the aggregate equations (25) lend themselves to a more meanin~ulanalysis of the in equation (24), namely: implications of policies that effect the parameters, (i) policies that effect the skills with which persons enter the research

community;

(ii) policies effecting the allocation of resources and effort to learning

and discovery;and (iii) policies effecting the number of peoplein the researchcommunity.

KITER

Policies (i) and (iii) are likely to involve long-run considerations. They wouldincludethesizeandquality of educationalinstitutions,and of research institutions. Policies(ii) are likely to have shorter horizons, involving current support for research efforts. Itis clear from (24) that analysis of policy instruments can proceed via the effects on the average levelof productivity in the research community on the one hand and the total number of researcher on the other. itis also clear from (10, orthecorrespondingsolution to (20), that resources devoted to increasing the elements of D yield increasing returns to scale in terms of knowledge. In Part 6, it will be clear that this implies increasing returns to scalein production.63

1 and 2 Considertwofields,say,biologyandpharmaceuticals,denoted respectively.M Peoplein biology do basic research; people in pharmace~ticals develop new drugs. Suppose thereare N, people engaged in biological research and N, in developing drugs, whereN = N, -tN,. On the one hand, suppose that the pharmacists use knowledge generatedby the biologists, but that the biologists do not learn from the pharmacists. We will compare this with the situation in which neither the biologists nor the pharmacists by equalearn from the other group. Then the aggregation procedure given tions (20) through (25) applied to(26) yields

where

where

7. ~ O ~ E D G DISCOVERY, E , AND GROWTH

with

and dl,-d,,>0, and dl, 0. Then,

has eigenvalues 1, = dl,, 1, = d,.

The corresponding eigenvectorsare

The matrix S of eigenvectors is S=

and hence

Therefore, the solution is givenby

REITER

where

Therefore

These equations can be used to compare the effects on the growth of pharmaceutical knowledge of varying the parameters dij.It is interesting that theparameterd,enterslinearly,while d,, and enterexponentially.Itis evident that increasingd,, or d22increases thegrowth rates of 6,(t)and x2(t), respectively. Sincedzlenters linearly into the equation E2(t), for it may not be so evident that increasing d,, from the value0 has the effect of increasing the growth ofg2(t)exponentially. The secondcase holds when d,, = 0. In the second case, where the pharmacists do not learn from the biologists, d'a =

= 0;

here primes denote the corresponding variables immediate~y evident that

If we assume thatdl,,= d,, for i = l,2, then

and,

in the second case. It is

7. ~ O ~ ~ E D DISCOVERY, G E , AND GROWTH

Because the relevant quantities in this expressionare positive, the ratio grows exponentially at the rate d,, -d,. Thus, while both E,(t) and &(l) increase e~ponentiallyin isolation,&(l) increases much faster when transin isolation. fer of knowledge from 1 to 2 is possible than it does

The model just presented predicts that in fields in which positive effort, resources, skill,and ability are applied, knowledge grows exponentially. Is this at all consistent with what can be observed? Thefollowingdatawerecollectedforthefield of computerscience, broadly defined. The numberof pages published annually in a sample of 21 journals of computer science, and separately, the number of articles published annuallyin those journals were collected for the period 19581991. to If the number of ideas perarticle, or the number of journal pages per idea, is not too variable, these quantities would be good approximations to the measure E&!), where q labels the fieldof computer science. 7.2 exhibits the regressionof Table 7.1 shows the data collected, and Table the natural logarithm ofnumberofpagesandnumberof articles, respectively, onTheseregressionsshowthatthenumber of pagesgrew of articles ex~onentiallyat the rateof about 9%per annum, and the number at about 11% per annum. These regressions seem to be very close fits.

The technological significance of the growth of knowledgederives ultimately the analysis is to use the dynamfrom its effect on production. The next in step ics of discovery and growthof knowledge to derive the growth of production possibilities. This might be carried out without introducing the concept of commodity into the model of technology. As the number of recipes grows, the number of different products, or useful substances would in general grow,or new recipes might give more efficient ways of producing existing things. In order to connect the model with standard models of production the commodity and production setare introduced, and the analysis carried out

REITER

TAB"!

7.1

The Number of Pages and the Number of &des Published Annually in ei Sampleof Computer Science Journals From 1954 to 1991 ~. " .

Year (time) 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 198 1 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991

Pages (P) 380 558 750 1038 1080 1490 1298 1992 l 692 1960 1628 1956 2844 3158 3440 2862 3414 3094 5014 4700 5678 7302 7302 8630 8560 10036 11400 12338 15126 18376 23042 21982 24572 249% 24498 27234 272 18 29686

. . . ~

Articles (A)

48 42 72 96 154 186 182 204 190 242 220 236 316 322 352 274 324 356 428 426 498 688 632 720 722 1188 880 994 1I08 I 260 1508 1636 1732 1672 1644 1768 1888 1958

in that framework. In the interest of clarity and simplicity, this is done in a

simple Leontief model of production. Research and Development@&D)refers to the step that converts knowledge, in~ludingconjecturedrecipes,intotestedrecipes. In thepresent model there is so far nothing that distinguishes this step from any other act of discovery.However,it is plausible in thisconnection to thinkofan attempt by an individualto develop a new recipe (which includes the invention of a new product, or a new use for an existing productas well as a new

7, ~ O ~ E D G DISCOVERY, E , AND GROWTH TABM7.2 Regressions of the Natural Logarithmsof the Number of Pages (LP) and the Number of Articles (LA) on Time (TIW) 1.

~ G ~ S I OF O THE N LOG OF THE N SCIENCE (LP)ON TIME LP = 6.361

~ OF PAGM ~ IN R COMP

+ 0.1 l TIME

(107.65

(41.64)

R SQUARE = .g79RBARSQUAREl=.g79 2.

~ ~ S I OF OTHE N LOG OF THE N~~ SCIENCE (LA) ON TIME LA = 4.301

OF ARTICXJSIN

COMPU~R

+ 0.09 T I m

(56.21) (29.24) R SQUARE = .g54 RBAR SQUARE

= .g52

~ ~ f The e . numbers in parenthesis are T-statistics

way of making something from existing materials), as primarily motivated by the anticipation of an economic return. Therefore, analysis of decisions to investin R W projects would requirethat theproblem be embeddedin a model in which anticipated returns could be expressed. This remainsto be done. For the present, themodel incorporates effort and resources devoted to R&D as parameters.It therefore allows analysis of the effect ofR W decisions on the technologyK*,This is donein Part 6.3,

General knowledge leadsto new technology through the application of effort, and resources toR W .In this processa potential innovator findsa project that seems sufficiently promising in technical and economic terms, and commits resources to its development intoa tested recipe and ultimately into production. The model as it stands does not include the economic structures, such as, marketsandpricesandconsumers'demand,orotherinstitutions,on which decisions to invest in the development of a particular area or project in an turn. Hence the analysis of expected or potential returns to investment R W prospect is not endogenous. Suppose instead that this economic analysis is done in a different model (ultimatelyto be integrated with this one) and leads to the selection of a certain fraction of the prospects for in~estmentand to the allocation of effort and resources to their development. The part of this process dealtwith in the present model is as follows.

WITER

Suppose to begin with that every person in is a potential innovator or pari at Time t consists of two parts, ticipant in R&D. The knowledge K(i,t) of Person rest, so that one contain in^ tested recipes and the other containing the K( i,t) = K*( i,t) U K( i,t)\K*( i,t).

(1)

Then k(i,t) = k*(i,t)+ k**(i,t),

(3

where k*( i,t) is the sizeof K*( i,t) andk**( i,t) is the sizeof K( i,t)\K*( i,t). Suppose that the yield of new recipes from the knowledge of Person i at t depends on the size of the knowledge base, and the effort and resources put into developmentby Person i. Then, let w(i)a(i)), = x(e*(i),

(3)

where, for all values of a(i), x(O,a(i)) = 0. Then, the yieldof new recipes from the knowledge baseK(i,t) given(3) is k*(i,t) = H(w(i),k(i,t)) The simplest form of H is that given in (4), where the coefficientw(i) is the ~ r ~ ~ coft iiinvR&L). i ~ (If Person i is not engaged in R M , then e*(i) = 0.) Then, for alli and t,@ k*(i,t)= w(i)k(i,t).

(4)

Let W = I W, where W = (w(i), ...,WO), and I is the identity matrix.It follows from(l?) of Part 5 that k*(t) = Wk(t) = W ~ ~ t ) S - ' k ( O ) = Wexp(At)S" k(0)

w(l)(cle%ll +c2e@q1+-+ cNeAN'sNl)

k*(t) = w ( ~ x c l e ~ '+c2e4'szw slN +"+cNeAN's,,)

7. ~ O ~ E D G DISCOVERY, E , AND GROWTH

economic model of production (the production set in the commodity space) can be derived from the model of technology presentedin Part 2 as follows. Given a collectionof recipes, a commodity space can be defined. Let K* = K*(t) = uK*(i,t)

denote the recipes known by the personsi in A at time t, and let denote the listof names of entities that appearin K*. I.e., a name is in MO(")

if and onlyif there is some recipein K* in which that namea~pears.6~ The list M(K*) is well-ordered in some arbitrary way, from 1 to m(K*), the index of

the last item on the list. An i ~ p u ~ o uaway ~ u tfor a recipe r is a function from M(K*) to Z, where 2 is the setof possible measurementsof substances on the list M(K*), say Z is an additive group. Let F(r) be a function whose value at the reciper is the set of input-output arrays for r. F(r) = {f E ZM6*)I f is an input-output array for r). This function can be given the following representation. Eachname n in M(K*) can be identified with its position on the list. A function f

E

F(r)

can be representedas the array =

cl,...Zm(K*)),

where if the name in position j is not mentioned in recipe r. Let Z* denote the spaceof m(K*)-tuples,z.

Commodities can now be defined.

,

MITER

~ ~ ~ o ~ j tA jcommodity e s . is a set of names of objects from the list M@*) that are regarded as equivalent in the model. Thus, let

be a partition of M@*). Each subset in the partition is a c o ~and ~is given ~ a iname. ~ For present purposes, the name of the commodity MC can be identified with c. For example, a set of names of objects consisting of “Taurus,” “Oldsmobile,” “Buick”...might be given the name “full-sized American-made car,’’ defined to be a commodity and identified as the third commodity. If Ford Motor Co. should introduce a newly designed full-sized car next year, it could be included as an instance of the commodity “full-sized Americanmade car” next year. ~ ~ yc ~is an eleFor each commodity, c, let ycdenote the 9 ~ f offc, where G,” Then, the commodity space is ment of an additive group, C

The elementsof Sp

where the cthcomponent of y is a quantity of the cthcommodity Also,there is a mapping, Y,

The function Y translates quantities of objects into quantities of commodities, Thus, letM, consist of the objects 1, 2, ...,p with measurements z,, z2, . ,z., Then, Y(z,), Y(z,), .. .,Y(z,) are the corresponding quantities of commodity 1, and the total amount of commodity 1 correspond in^ to z,, z, . ,z, of the objects 1, .. ,p is C Y(zj). The mappingY therefore determines an aggregation rule for the partition For example, suppose the objects M, are sea-goingvessels of different peeds. Theyappear on the list M@*) as different objects, but the defining commoditiesputs them in one class, M,, called “ship.” The measurements associated with the original objects might be linear dimensions of the vessel, parameters that characterize the shape, and the speed. These constitute a multidimensional quantity and would differ from per vessel to vessel. The quantity of the commodityship might be “ton-miles hour of transport capacity,” computed from the measurements zj of the individual vessels via the function Y.

7. ~ O ~ E D G DISCOVERY, E , AND GROWTH

Theprodion setdetermined by K* is a subommodityspace S .defined by theconditionthat

ion Set,

Y(K*) = {y E Sp(P) I 3 r E K* such thatz E F(r) and Y(z) = {y

E

= y}

Sp(P) I 3 r E K* such thatY"(y) nF(r) # E?}.

That is, the production set determined by K* consists of all commodity vectors y such thaty corresponds to an input-outputarray produced by some recipe in K*.69 Assumptions can be imposed on K* that imply the familiar properties in models (e.g., convexity, or linearity), assumed about the production set where it isa primitive. If there is technological change,so that K*(t+l)...K*(t);' then the commodity space may change and the production set determined by K*(t+l) may be different from that determinedby K*(t). If a new recipe produces a new product, then either that product cambe classified as an instance of an existing commodity,or as anew commodity. Thecommodityspaceassociatedwith a giventechnology K*, isnot unique. Different choicesofandof the mappings F and Y lead to different commodity spaces. The distinctions among substances and objects are that made for the purposes of knowledge mayor may not bepreserved in the disecotinctions that economic institutions make, and the distinctionsby made by economicmodels,each nomicinstitutionsmaynotbepreserved designed to serve a different purpose. However, in both the case of economic institutions and models, the classification expressing the distinctions that are recognized is made with knowledge of the technology K*(t). Therefore, thecommodity space can change after a new recipe entersK*. Although the description, properties, and name of a new product cannot beknown in advance of its invention or discovery, it can in advance be assigned the numberm@?) + 1 in the listof names M(K*(t+l)) and,if the new product isto be classifiedas anew commodity, assigned the number C+ 1 in the list of commodities. A similar convention would apply to the case in which a new use for some object or substance is discovered, a use that leads to a distinction being drawn between objectsor substances that were hitherto regarded asin~istinguishable.~ 6,2.4. ~~e~ ~ a i n a b l e ~ r o ~ u Set c tin i othe n ~ o ~ ~ S~ace. o ~ i tTurny

ing to the production set, (and suppressing the time index,t, temporarily), in 6.2.3) let the commodity space be (as Sp(p) =G = I;fGc *l

WTER

and suppose that Y, a subset of G, is the production set determined bythe technology K*(t) = K*(l,t)x ...xK*(N,t), andthe partition element y U G ,where

can be writtenin the form

where y*is the vector of p r ~ u c e d~ o ~ ~ and ~ iy2~ is the e vector s , of unprodzzced or p r i ~ fc fo ~~ ~ ~ j whose t i e s amountsare given from naturd2Suppose the produced commodities are 1, ...,Cland the unproduced ones are Cl+1, . ..,c. Let

denote the endowment vector. Only the components of v that are unpro0. Then the set duced commodities can be different from

consists of commodity vectors that do not use more than the available amounts of unproduced commodities.Let ? = Y n { y E G 1 ykv}, denote the setof ff~ffinff6le pr~uctions, where Y is the production The focus of this model is on the relationship between discovery and invention and economic growth. Therefore,isitappropriate to assume that other possible causes of growth are absent. These include growth deriving from increasingreturns and externalities,from trade, and from growth in the supply of unproduced resources. Therefore,the model of production considered is one with constant returns and no substitution possibilities, and the endowment of primary comm~ditiesis held constant. In that case, in the absence of technological chacflge,Y would beconstant over time. Because Y, and therefore Y, depend on the functions F and Y, analysis of the effect of new knowledge on production sets must involve properties of those functions. However, instead of specifyingFand Y directly and deriving the production set through them,I consider a familiar simple modelof production, and letF and Y be defined implicitly, determined by "reverse engineering."

7. ~ O ~ E D G DISCOVERY, E , AND GROWTH

Assume that Y is given by a Leontief input-output model with no joint production. Then, Y = { ~ E IGy=Ax,x>O,

where A is an L x L,matrix and x = (x,, ...,xcI)). It is assumed that each produced commodity has one activity (industry) that produces only that commodity? A the efficient frontierof 9 is the Under standard conditions on the matrix intersection of a hyperplane determined by the column vectors of A, an therefore by the coefficientsof A, with the non-negative orthant of the commodity space. In that case, the points at which the efficient frontier intersect the coordinate axes_characterize the efficient frontier. Moreover, they also Y, since it is the convex hull of those points and the origin, characterize the set Thecoefficientsof A dependontherecipes in K*, andhenceas K* changesover time,so does the matrix A. The effect ofdiscovery or invention onproductionpossibilitie_sisexpressed in the change over time of the K* changes by attainable productionset Y and its efficient frontier. Because including new recipes, we first model the relation between recipes and the coefficientsof A. Because of the underlying linearity it is plausible to suppose that the relation between recipesin ,K"(t) and the coefficients of Y is homogeneous linear. The coefficients of Y are the same as those of Y, and are denoted ape Thus, forp = 1, ...,L1, and c = 1, ....,L, The coefficient b;, i = 1, ...,N, measures the effectof technological knowledge of person i on the coefficient ape. Sincesets the K*(i,t) consist of tested recipes, the effectof knowledge in any one of them on the technical coefficientsisself-contained.Thisjustifiestheassumptionthattherelation between knowledge of agents and technical coefficients is,to a first approximation, linear. Recipes that effect any particular coefficient can be in the of knowledge setof any person. Accordingto (2.1), the total effect is the sum the effects from each person. Interactions among the knowledge sets of different personsare already capturedin the underlying equations that determine the sets K*(i,t). Substituting from (5) the solutions for the functions k*(i,t) in the case where D has N linearly independent eigenvectors,= gives the following equation for a,,(t).

EITER

a~(t)=b~k*(l,t)+...+b~k*(N,t)

= b;w(l)k(i,t)+-+ b~w(N)k(N,t)

=b ~ w ( l ~ C l e a l+C2ea2's2, ts,l +-+CNeaNtsNN)

+

+e**

b~w(~~c~+ e C2ea21S2N ~"SI+ N i-cNeaN'SNN ) =Cleal'(b~w(l)sll +~*.+b~W(N)SNl)

(2.1')

+*e*+

cNeaNt(b~W(l)SN~b ~ w ( ~ ) s ~ ) +***+

N

N

j=l

i=1

If the impact of recipes on source, then

a technical coefficient is independent of the

It follows that N

N

/=l

i=l

If further, the productivityin R W of agents is the same, then for ail i,

w(i) = W, and N

N

j=1

i=l

Let

i=l

Then, (2.4) can be written

7, ~ O ~ E D G DISCOVERY, E : AND GROWTH

Notice that the same conditions yield the corresponding expression for the sum of the componentsof k*(t), namely, N i=1

j=1

The next step in the analysis is to find the rateof growth of the attainable production set9,and its efficient frontier. This is done in the contextof a low dimensional example. ~ u ~Set.i o Consider ~ anexample in which there are 2 produced commodities and1 primary commodity (Le., C= 3, and C = 2). Let the initial endowment vector be

and the matrixA be,

v

Figure 7.1 shows the projection of the set into the plane givenby yj = -1, and the pointsj i l and j i 2 at which the line determined by the two columns of A intersect they1 and y2axes respectively? Then the equations are Y1= a11x1 -QaX2 Y2

= -a123

+

a2s2

-1 =-x1 -x2

From 3.3), x1= 1 -3. ~ubstitutingin (3.1) and (3.2) gives

FIG. 7.1.

and

Su~stitutingin (3.6) gives,

-a1da12 +a22)-a12(a11 _ .

Y2

Let

a11 +Q21

+a211

a12 -Y1-*

+a22

a11 + a21

7. ~ O ~ E D G DISCOVERY, E , AND GROWTH

and

Then equation (3.8) can be writtenas y2=B-Cy1.

To find the pointsF,and F2where the efficient frontierof 1 intersects theyland y2-axes respectively, set yz= 0 and solve (3.8)for Y1,and then carry out the corresponding procedure forF2.Thus, settingy2= 0 gives,

Setting y1= 0, yields

y2 = B (3.10) §ubstituting from (2.1) and (2.2) into equations (3.9) and (3.10) respectively yields equations (3.11) and(3.12) which give the time paths 9,O) and F2(t),in the special case when equation(2.2) is valid

(3.11)

and,

WITER

92

( 0=

bll~(t)b22~*(t)- b12~(t)b21~(t) bllK*(t)+b,2k?(t)

(3.121

where

K* (t) =

(i,t). ieA'

(3.13)

It follows that (3.14)

where Bl= (4lb22 -4 2 4 1 1 4 1

-622

'

and (3.15) where 4!=

(W22 -4 2 4 1 1 4 1 + 4 2

The set of commodity vectors attainable with one unitof the third commodity, is the convex hull of the three points,0, .fil(t)?and f2(t). Hence the growth of? (t) is determined by fl(t) and .fi2(t). As equations (3.14) and (3.15) make clear, their growth is exponential in t (when there is a positive eigenvalue) at a rate that depends on the number of people, their abilities and the resources allocatedto their respective areas, and to the effort expended on R&D in those areas,and on the pursuit of knowledge in the areas they rest on. Figure 7.2 shows the attainable production set at two points in time under the assumption thatB, is greater than2.78 In this example, under the assumptions made, the setof attainable productionsgrowsexponentially over Therefore so wouldproduction.

7, ~ O ~ E D GDISCOVERY, E , AND GROWTH Y3

FIG. 7.2.

If the model of production were, say a linear activity analysis model, the attainab~epro~uctionset could be characterized by the points at which the raysin the commodity space generated by the basic activities intersect the resource constraints.Of course, the analysis of the growth of the attainable set would be more complicated, because the direction of a ray determinedby a basic activities could change over time. If a recipe involving a new product were discovered, then it is either included in the setMC of some existing commodityc, or the classification is changed so that a new commodity is introduced. In that case, it is comof that commodity modity Cl"1, and the differential equation for the growth determines itstrajectory start in^ from the initial condition

If the setof potential developers is a subset A. of the set the sum in equation (3.15) is takenover the subset A. If the setof deve~opers is specialized by industry, so that thereis a set,i+,, whose knowledge is relan onlyif p = p', then the sumin eq~ation(2.1) evant to the coefficients ifapq is taken over i E A,,.8o

REITER *

I.

that in Example 1 ofPart 5.6, in whichbiologistsandpharmacistsdo of the research and R&D,the parametersd, depend on the skills and abilities members of the respective groups, and on the effort and resources they devote to discovery and to learning from others (i.e.,on the parameterse(i) and a(i)). Recalling equation (4), for all i and t, k*(i,t)= w(i)k(i,t), where,from (3), w(i) dependsontheeffortandresources e*(i) that i devotes to R&D, and on i's ability and skill parameters a(i). To comparethegrowth of attainableproduction sets of different economies,characterized by differentcomplexes of parametervalues, involves comparing the time pathsof the attainable production sets.In the present example, the attainable production set is determined by the point Fi(t) definedin 6.2.4. Using equations (3.11) and (3.12), we see that and since the~iolo~ists do not directly contribute any new recipes. ~ubstituting the solution forE2(t)from equation (29) Part 5, yields,

and,

aringthis sit~ationwiththeone

in which

12,

we seethat

7. ~ O ~ E D G DISCOVERY, E , AND GROWTH

This is the same expression as for the growth of 5(t)relative to EL(t) as in equation (30) of Part 5. Thesameresultapplies to theratio .Thus, &? 0) ; the additionto the relative rateof growth of knowledge dueto communication between the basic research community andR&D thecommunity carries over to the relative rate of growth of the attainable production set for given resources. Consider an example in which there are two countries each with a single sector that does both basic research and R&D. Suppose that Country1 is underdeveloped compared to Country 2. That is, K(1,O) is very small compared toK(2,0), where K(i,O) is the present stateof knowledge in Country i. Then k(1,O) is much smaller thank(2,O). The disparitybetweenthestates of knowledge in theresearchsector between the two countries reflects the backwardness of the current technology in Country 1 and the lack of knowledge of more basic subjects as well. Example l can be reinterpreted to represent the growthof two countries, one of which, Country2, is less developed than is Country1. Wemay assume that everything known in Country 2 is also known in Country 1. In that case, the matrix D' has the form

Let

Then, € o r

h,(6;'

0)

552221

so defined, the analysisof Example l, with d,, replaced by di2-di1,applies. are the equalof those Country 1 may contain people whose innate abilities in Country 2, butlackofeducatioandof train in^ an e~periencewithan advanced technology leads to the S 11 component of th than a(2). ~imi~arly, resources are scarcein Country 1, so that

REITER

r(1) is small and henceso are 2) hosts. Now, thedecisionisnotthesimplebinaryone of eitherworking together or working alone. The optimal working group size can range from one (work alone) to n (work together), or any value in between. We developed the concept of “scopeof control.”In the contextof load balif the former’s scheduler is able to send ancing, a host controls another host jobs to the latter’sserver (with possible queueing).A host’s scopeof control is the setof all hostsit controls.We use ra~domizationto determine the control sets: each Host i maintains a set of probabilities cii= Prob (Host i controls Host j), for all hostsj. If cli= 0 for all j, we have a partitioned system. Otherwise, we have various degrees of a pooled or shared resource system. See Fig. 11.4. Load balancing decisionsare made as follows. Each time a job arrives at a Host i,a control set is determined: For each Host], randomly decide if it is in the control set according to its current probability cII. The scheduler then selects the least-loaded hostin the control set,and the job is sent to that host’s server. To learn “optimal” controlsets, the control set probabilities are updated according to the learning automata scheme. If the host assignmentis a good decision, then the probability of including that hostin the control setin the in a lowering of the probability.A future is increased; a bad decision results

11. DECENTRALIZEDCONTROLUSINGRANDOMIZED COO~INATION

0

0 -

-\ \

\

FIG. 11.4. Scope of control.

good decision is defined as one that resultsin better-than-average performance (Le., the job completes execution faster than what is expected in a completely balanced system). The hosts communicate their state information with each other (i.e., they A host uses the received probabilexchange their control set probabilities). ities of other hosts to determine whether the decision it makes is expected to be good; this is what is actually used to reward or penalize decisions. Finally, as in our previous experiments, uncertainty is introducedby only periodically communicating state information, so that it becomes less timely until a new update is received. Themain result is that as the communication period increases (i.e., as of thecontrolsets moreuncertainty is introduced),theaveragesizes decrease. When there isvery low uncertainty, the hosts learn that it is best to work togetherin one large group (any host can off-load to jobs any other to work in smaller, but not nechost). As the uncertainty increases, it is best essarily singleton, groups (hosts off-load only to other hosts within their group). Finally, when there is high uncertainty, it is best to work alone (no off-loadingof jobs). Consequently, the scope-of-control randomization mechanism provides a wayfordecisionmakerstotakeuncertainty of stateinformationinto account. They learnby using randomizationto search the group (or control In fact, the mechset) space, allowing them to act a self-organizing as system. anism is very robust. Extending the simulated system by having heteroge-

PASQUALE

neous hosts (Le., some hosts have faster servers than others, the system self-organizes so that the hosts with fasterservers appear more frequently in control sets than those with slower servers). In another related study described in [ll], where the distance between hosts is taken into account (the greater the distance, the more communication delay, and hence the greater the unce~ainty),the system self-organizes so that hosts prefer to work with those hosts that arein closer proximity than other hosts.

We have shown that randomization can be used to attack the two fundamental problems of decentralized control, that of stat~informationuncertainty and thatof mutually conflicting decisions. These problems inhibit the coordination required to achieve the potentially high levels of performance offered by distributedsystems.Usingrandomizationis very appealing because it simplifies the interactions to achieve coordination between decision makers.~andomization can be usedas part of defensive mechanisms to by breaking symmetry. Furthermore, prevent mutually conflicting decisions by usingrandomization,onecaneasilyincorporateadaptivefeedback mechanisms to allowa decentralized control systemto self~rganizeinto a cooperatingwhole.Thisallowsrandomizationtobeusedtoproactively search for good collective decisions that are coordinated, in contrast to its complementary usein defensive mechanisms. We reviewedanumber of studieswhererandomizationwasused to difimprove the performance of various decentralized control systems given ferent constraints. With no communication, randomization provides a significantbutlimitedperformanceimprovement.Thelimitationdoesnot make the pooling of resources worthwhile; rather, it is better to statically partition the system. This changes when communication is possible and where the communicated informationofissufficient quality. Where communication of state information is periodic, there is generally a threshold for the communication period whereby, below the threshold, decision makers learn to eventually make optimal decisions (coordination is achieved), and above it coordination breaks down. Decision makers can learn whether it is worth working together (which requires coordination) or working alone (which does not), based on the quality of the communicatedstate information.By working together, higher levels of performance are achievable. By workingalone,performanceis more predictable because decision-makers do not interact, but the performance gain is significantly more modest. In fact, they can learn the optimal size for coordinated groups; this size is directly proportional to the quality of communicated state information.

11. DECENT~LI~ED CONTROLUSING ~ D O M I ~ ECOO~INATION D

Yet,despitethepower ofusingrandomization,we see very clearly throughoutourstudiesthe si~nificantdegreetowhichuncertainty will degrade performance. The two fundamental problems of decentralized control are the roots of this uncertainty; it is only by gaining a deeper understanding of them that wewill be able to develop additional methods to realize the potential powerof distributed systems,

I am grateful to my students, especiallyTed Biilard and Alex Glockner, with whom I explored manyof these issues. Most importantly,I thank the lateDr. Larry Rosenberg for his support and friendship. Larry's vision and enthusiasmfor a research agenda in coordination and collaboration remain an inspiration to me. This work was supportedby the National Science Foundationas partof a Presidential Young Investigator Award, Number IRI-8957603.

[ l ] T. W. Malone and K. Crowston, "What is coordination theory and how can it help design cooperative work Systems," Prw. ACM CSCWCont,Los Angeles, October1990. 121 J. HaIpern and V. Moses, owled edge and common knowledge in a distributed environment," ~ournal ofthe ACM, Vol. 37, pp. 549-587, July 1990. [3] Webster's Ninth New Collegiate Dictiona~.Springfield, MA Merriam Webster Inc.,1987. [4] J. Pasquale, "Randomized coordination in an autonomous decentralized system," 1stEEE Inti. Symp. onAutonom~ Decentral~~d Systems {ISXIS", Kawasaki, Japan, March93, pp. 77-82. [5] E.Billard and J. Pasquale, "Utilizing localand global queueing resources with uncertainty in state and service,"Prw. 2nd IEEE Intl. Symp. on Autonomous Decentralized Systems IS AD^], Phoenix, April 1995, pp. 258-265, 161 K, Narendra and M. Thathachar, Learning Automata: An I n ~ ~ u c t i oEnglewood n. Cliffs, NJ: Prentice-Hall, 1989, [7] E. Billard and J, Pasquale, "Adaptivecoordination in distributed systems with delayed comVol. 25, No. 4, April 1995, pp. munication,"IEEE Transactions on Systems, Man, and Cybernetics,

Pm.

546-554, 181 A. Glockner and J. Pasquale, "Coadaptive behavior in a simple distributed job scheduling

system," IEEE Transactions on Systems, Man, and ~bernetics,Vol. 23, No. 3, ~ a y f J u n93, e pp. 902-~. €91 E. Billard and J. Pasquale, "Effectsof delayed communicationin dynamic group formation," IEEE ~ansactionson Systems, Man, and Cybernetics,Vol.23, No. 5,SeptemberfOctober93, pp. 1265-1275. [101 E. Billard and J. Pasquale, "Dynamic scope of control in decentralized job scheduling," Prw. 1st IEEE Intl. Symp. on Autonomous Decentralized Systems (ISADS], Kawasaki, Japan, March 93, pp. 183-189. [l11 E. Billard and J. Pasquale, "Localized decision making and the value of information in decentralized control," Prw. 7th Intl. Cont on Parallel and Dis~ibutedomp put in^ Systems {PDCS], Las Vegas, October1994, pp. 417-425.

This Page Intentionally Left Blank

C H A P T E R

F. Donelson Smith

University of North Carolina, Chapel I

Hill

TI

Future hypermedia systems will integrate diverse information resources, systems, and technologies. They will be based on modular architectures p~ug~ompat(e.g., Thompson,1990) that separate orthogonal concerns into ible components such as change management, query and content search, notification,application-s~ecific concurrency control, computational semantics, and window conferencing. Someof these components, such as change mana~ement,maybe highly dependent on the semantics of a particular domain, whereas others will provide general support for all applications. The key point ...is that it is modular and open. This modularity is based on the observations that the functions the modules performare independent of each other, thatis o r ~ ~ o g o ~ u l i ~ i ~ ~ (Thompson, l i e s ~ ~1990, u l up.r234) i~.

Ortho~onalityimplies modularity; m ~ u Z a r implies i~ choice. The importance of this observation is thatevery service has a'cost associated with it. For example, transactions may be the appropriate concurrency mechanism for one application, while imposing prohibiti~elyhigh overhead on another. *Copyright 0 1993 by the Association for Computing Machinery, Inc. Thischapter waspublished previouslyin the ~ ~ e e do if ~ A C s ~ ~ y '93, ~ pages e ~ ~1-13, r and is reprinted here with permission from the Association for Computing Machinery.

S H A C ~ L F OSMITH, ~ , SMITH

Ideally, one should be able to use a service when it is needed without having to pay for it when it is not. In this chapter, we describe the architecture and implementation of our Distributed Graph Storage(DGS) system. We have designed itin a way that supports modular expansion to add services such as those enumerated earlier. A fundamental requirement has been that the basic hypermedia services for data storage and access should be inexpensive, efficient, and scalable. This is ~articularlyimportant because the performance of these basic services is an upper bound on the performance of the systemas awhole. The DGS has been developed as a part of a larger program of research that focuses on the processof collaboration and on technology to support that process. We are concerned with the ~ntellectual collaboration thatis required for designing software systems or other similar tasks in which groups of people work togetherto build large, complex structuresof ideas. Theworkofsuch groups-either directly or indirectly-is concerned with pro~ucingsome tangible artifact. For software systems, the artifact may includeconceptpapers,architecture,orspecificationdocuments,programs, diagrams, reference and user manuals, as well as administrative documents. A subtle but important point is that we view a group’s tangible creations as partsof a single artifact. Our research in the UNC Collaboratory project studies how groups merge their ideas and their efforts to build an artifact, and we are developing a computersystem(called B C forArtifact-BasedCollaboration;Smith & Smith, 1991) to support that process. B C has six key components (Jeffay, Lin, Menges, Smith,& Smith, 1992): the Distributed Graph Storage system,a set of graph browsers, a set of data application programs,a shared window conferencing facility, real-time video and audio, and a set of protocol tools for studying group behaviors and strategies.

In thissectionwegive a briefsummaryofkey shaped our storage service design.

requirem~ntsthat have

~ e r ~ a n e npersist t en^ storage: obvious but fundamental. ~ ~ a it^ r ~i r o~t e c ~ obecause n: the artifact effectively constitutes the

by all. There are, howevgroup’s collective memory, it must be sharable er, requirements for mechanisms to authorize or deny access to selected elements of the artifactby individuals or subgroups. C o ~ c ~ ~access: e n t because collaborators must work together, itis often necessary for more than one user to read or modify some of the partarti-

12. ~ C ~ I T E C TOFU~ ~

E

~ SYSTEM E D

~

fact at the same time. Data consistency semantics in these cases should be easily understood and provide minimal barriers to users' access to the artifact, ~esponsive perfor~ance: sufficient to support interactive browsing of the artifact, is required. ~cfflable: we are concerned about scale in two respects: the number of users in a group (and consequent size and complexity of the artifact), and the geographic dispersion of group members, To be scalable, it must be possible to distribute the system over available processing and network resources and to add resources incrementally as necessary. Performance (responsiveness) as perceivedby users must not degrade significantlyas the system growsin scale, Available: if data becomes unavailable because of system faults, users may be severely impacted. The system must, therefore, be designed to tolerate most common faults and continue to provide access to most or all elements of the artifact. ~eplicationof data and processing capacity is required to achieve high availability. User and artifact~ o b i l iusers ~ : will need to change locations and system administratorswill need to move dataor processing resources to balance loads and capacity. The system must support this mobility in a way that is transparent to users and application programs. There should be no location dependencies inherentin the storage system. Private data: these are created by individuals for their own use. include personal notes, annotations on documents, and correspondence, Users must be able create to and protect such data and still establish relationships among them and the public artifact. upp port for ffpplicff~ons: many applications used by a group are likely to be existing tools suchas editors,drawing packages, compilers, and utilities, which use a conventional file model for persistent storage. The system should make it possibleto use such tools on node dataeontent with no changes. L

The most basic element of the data model is the node, which usually contains the expression of a single thought or idea. Structural and semantic relas nodes.' tionships between nodes are represented explicitly as l i n ~ between 'Links to links are prohibit~d.

SHACKELFORD,SMITH, SMITH

The data model provides two mechanisms for storing information within a node: node attributes and node content.A ~ i 6 ~ t are e s typed, named variables for storing fine-grained information (approximately 1-100 bytes). Some attributes (such as creation time and size) are maintained automaticallyby the system. There may also be an unlimited numberof application~efined attributes. In comparison to attributes, node content is designed to reference larger amounts of information. This content can take one of two forms: 1. a stream of bytes (accessed usinga file metaphor) 2. a composite object (accessed usinga graph metaphor)

Applications control whether the content of a particular node isof Type 1 or of Type 2. Because Type I content obeys the standard file metaphor, it can be used to store the same typesof information as files (e.g., text, bitmaps, line drawings, digitized audio and video, spreadsheets, and other binary data). Applications that can read and write conventional files can read and I content is stored with the node write TypeI content with no changes. Type that contains it. se~aratelyas When a node has Type2 content, then the content is stored a composite object calleda subgraph? A s ~ 6 ~isrdefined a ~ ~as a subsetof in the artifact that is consistent with graph-theoret~c conthe nodes and links straints. Forexample,allsubgraphssatisfytheconditionthat if a link belongs to a subgraph, then so do the link's source node and target node. Nodes and links may belong to multiple subgraphs at the same time, but every node and link must belong to at least one subgraph. Our data model also providess ~ o ~typed i y subgraphs (e.g.,trees and lists) thatare guaranteed to be consistent with their type. Links can have both attributes and content associated with them. Moreover, the datamodel definestwo classes of links: structural and hyperstructural. Structural links (Slinks) are used tostore the essential structure of an artifact, By contrast,hyperstructurallinks(€"links) are lighterweight objects that represent relationships that cut across the basic structure (see Fig. 12.1). Subgraphs containing only structural links are called S-subgraphs; those containing hyperstructural links are called HS-subgraphs. The data model encourages users to compose a large artifact from small subgraphs using subgraph content. This organization can improve human co~prehensionof the artifact and increase the potential for concurrent 2Hereafter, Type1 content will be referred to as File content and Type 2 content will be called s u b g ~ content. a~~

12, ARCHITECTU~OF ~ P E ~ E D SYSTEM I A

FIG. 12.1.Examples of hyperstructural linking.

access to individual components. The best way to understand these mechanisms is by example. Figure 12.2 illustrates one way to organize the public and private materials associated with a large research project (node content is indicatedby dashed lines). One can observe thatFig. 12.2 subsumes the or~anizationof data in a conventional file system while providing additional mechanisms for storing metainformation about files(in attributes) and for representing (in links). semantic and structural relationships between files Subgraph SG 9 in Fig. 12.2 is the top-level subgraph of a document. A useful exercise isto compare this graph structure with the way that the conference paper would be storedin a conventional file system. The most striking difference is the number and size of the nodes that compose the document. Whereas a conventional document would normally be stored In a single file or a small numberof files, theDGS data model encouragesa userto divide documents into many smaller nodes and subgraphs. This maximizes the benefits of hyperstructural linking because each node expresses a single concept or idea. By dividing a document into different subgraphs, collaborators may be able to structure their materials for easier concurrent access. Although nodes are finer grained thai-t. traditional files,there are still times when one would like to reference information at an even finer level. For example, an application might wantto create alink that points toa specific

S ~ ~ ~ LSMITH, F O SMITH ~ ,

FIG. 12.2. Organizin~the public and private pieces of an artifact.

word within a node, rather than to the node itself. To achieve fine-grained linking like this, the data model provides the conceptof an anchor withina node. An anchor identifies partof a node’s content, such as a function declaration in a program module,a definition in a glossary, or an elementof a an onto a specific place line drawing.An anchor can be used to focusHS-link within the content of a node, When an HS-link is paired with one or more anchors in its source or target nodes,it is called an f f n c ~ o~ ~ -e 1~iThe n~. relationship between anchors and HS-linksmany-t~many. is Some attributes are called c o ~ ~ ao~ni b ~ tbecause es their values are independent of the context from which they are accessed. All objects-nodes, links, and subgraphs-can have common attributes, In addition, nodes and linkscanhavecontext-sensitiveattributeswhosevaluemaybedifferent

12. ARCHITECT^^ OF HYPER~EDIASYSTEM

depending on the context from which they are accessed.This second typeof attribute is calleda ~ a p ~ because a ~a subgraph i ~ ~ t provides e the context.

As shown in Fig. 12.3, the DGS has a layered architecture that can be configured in a number of different ways. The Application Layer contains the user interface and other code that is application-specific. The top layerof the DGS is theAppli~ation~ogrammingInterface (jlpr)which exports agraph~riented data modelto applications.An overview of this data model was presented in a previous section. Mostof the DGS is implemented in the bottom two layers: the Graph-Cache Manager (GCM) and the Storage Layer. The GC implements the data model and performs local caching; Storage the Layeris responsible for permanently storing results. Because theAPI isolates the application from the restof the DGS, appiication code is portable across different implementations of the bottom two layers. We currently support two different implementationsof the layer and two different methods for connecting the API with the GC

I

I

FIG. 12.3. Four implementations of the DGS layered architecture.

S

~

C

~ SMITH, ~ SMITH O ~

,

yields the four implementations that are shown in Fig. 12.3. In ~GS-M2,the application and the GCM run in different processes on the same machine; the Storage Layeris implemented as amultiuser, distributed storage server. I is the same except that the GCM is linked with the appIication to becomeasingleprocess.Theadvantageofthisdesign is betterlocal response time due to reduced Interprocess Communication(IPC). A disadvantage is that it increases the size of application executables, DGS-S I and DGSS2 follow a similar pattern except that the distributed storage server is replaced by a singleuser,nondistributed storage layer. e

The API for the DGS is a C* class library (for a complete description,see Shacke~ford,1993).Figure12.4shows the major classes in the inheritance hierarchy.The class Object definesoperationsthatarecommontoall objects such as the functions for manipulating the common attributes of an object. Subclasses inherit the API of their parentclass and extend the inherited API with more specialized functions. All node, link, and subgraph objects areidentified by an object identifier (OID) that is universal and unique, Once anobject is createdby the DGS, its OID is never changed and the value is never reused even if the object is deleted. To applications,an OID is an “opaque” (uninterpreted) key that can be used to retrieve the corresponding object. However, we discourage application programmers from making directreference toOlDs. Most operations can be performed without even knowing that OlDs exist.

Becausethe DGS datamodelis object-oriented,the objects of thedata model-nodes, links, andsubgraphs~xistas distinct entities within the storage system. Beforea user’s application can access the data withina particuobject lar object (see Fig. 12.5), theapplicationmustexplicitlyopenthe using its O ~ efunction. ~ ~ )

-ti

Ink FIG. 12.4. A P I class hierarchy.

12. A R C H I T E C T U ~OF H Y P E ~ E SYSTEM D ~

FIG.12.5. Information stored in nodes, links, and subgraphs.

Openo will fail if the user lacks the proper access authorizationsif the and in progress. Conflict can occur when request isin conflict with other requests different users try to access the same object concurrently.To specify allowMI defines three access modes for nodes, links, able concurrent accesses, the D G ~ ~ DGS-RE2B-N Tand E , subgraphs: and cations must sp these modes as a parameter to allows operations that do not change subgraph rnembersh tion, or attribute or content values. In the case of nodes, allowsanchorcreation and deletion,butonlywhenthe H§-link thatisbeinganchored. DCSread-writeauthorizationonthe N ~ is defined ~ ~only ~for nodes ~ and O allows ~ operation^ all of DCS except anchor creation and deletion. D C S ~ allows ~ ~ all E o~erations. The following rules govern concurrent access to an object:

9

For links and sub~raphs, multiple opens withDGS-REMI access and a single openwith D G S - ~ I T Eaccess are allowed concurrently (as is the weaker case of multiple D G S - m opens alone). For nodes, multiple opens withD ~ S - ~ - N O - ~ Caccess ~ O and R a single open with ~ ~ access are ~ allowed ~ concurrently T (as Eis the weaker ~ D G ~ ~ D - N opens ~ ~ ~alone). H O ~ case of multiple D G S - and/or

S ~ ~ ~ LSMITH, F OSMITH ~ ,

Thus, for nodes the design supports multiple nonannotatingreaders and a single writerOR multiple annotating readers. A consequence of this is thata writer is blocked from accessinga node that is being annotatedby a reader and vice versa. Changes to an object are not visibleto any applications with over~apping opens of the object until it is closed by the writer and then only to applications that open it after the close completes. Groupscancontrolaccess to parts of theartifact by specifyingaccess authorizationsfornode, link, andsubgraphobjects.Authorizations are An access expressed in an access control list that is stored with each object. con~ollist maps names of users or groups of users to categories of operations that theyare allowed to perform on the associated object.Two categories of authori~ations are defined: access and administer. Access a ~ t ~ o r i ~ a ~ ogive n s users permission to access the data associated with a particular object. A ~ ~ ~ n i sat ~e rt ~ o r ~ a give ~ o nusers s permission to perform operations such as changing the object’s access control list. Although the API doesnotdefineanexplicitannotatepermission, a similareffectcanbe accomplished by restricting the access authorizations associated withHSsubgraphs. Implementation

In thissectionwediscussthedistributedimplementations(DGS-M1and DGS-M2 in Fig, 12.3) with emphasis on key design decisions.

Givenanartifactcomposedfromsmallelementsanduseraccessvia interactive browsers, we believe many characteristics and access patterns of objects will strongly resemble those observed in distributed file systems supportingsoftwareteamsusingworkstations(Baker,Hartman,Kupfer, Shirriff, & Ousterhout, 1991; Kistler & Satyanarayanan, 1991). Our design is based on the notion that a scalable implementation can be achieved by applying design principles such as local caching, bulk-data transfer, minand imal client-server interactions pioneeredin high-performance, scalable file mslike AFS (Howard et al., 1 ,Sprite (Nelson, Welch, & ~usterhout, , andCoda(Kistler & Sa rayanan,1991). We also modelour a~proachesto data consistency, concurrency semantics, and replication after these distributed file systems. This provides a sufficient level of funcfull complexity of mechanisms (e.g., distion to users without requiring the tributed transactions) usedin database systems. in Fig. 12.6. A browser or appliThe basicstructure of the system is shown cation process acts onbehalfof a user to readandmodif objects. Each user’s wor~stationruns a single Graph-Cache

12. ~ C ~ I T E OF ~ T~ U ~E

~ SYSTEM E D

~

FIG. 12.6. DGS system structure.

services all applications running on that machine. Application requests are directed over local interprocess communication faciliti GCM maintains a local copy of node, link, and subgraph objects application processes and is responsible for implementi objects in the data model except for anchor table mergi responsibleformaintainingtheconsistencyoftyped S important to note that this design distributes the processing forcomple~ all object operations to the users’ workstations and thus minimizes the processing demands on shared (server) resources. en an application opens an object, the GCM,in turn, opens theobject atthestorageserver and retrieves itusing a whole-filetransfer,The received object is con~ertedfrom its representationin a file to anobject rep ed for fast access in memory.As the application makes in its localcach~. performs those operations on the copy e reflectedin the storage server only whenthe GCM closes the object and r~turnsthe modified file representation to the sto server. Each file retrieved from the storage server contains either a node (inc~udin~ data content, if present), a whole subgraph, or a group of links. An importantperformanceoptimization is thatconte )and link information for all nodes file, Thus, all of the data needed b

SHACmLFORD, SMITH, SMITH

display a subgraph is available froma single request (open) to the storage server.The structure of each typeof file is shownin Fig. 12.7. Nodes and subgraphs are stored individually, whereas linksare grouped according to the subgraph in which they were created. The file~rientedinterface to the storage server is designedto isolate itas much as possible from the representation and semantics of objects.The primary responsibilityof the storageserver, therefore, is to storeand control access to files indexedby an object’s OID. Storage servers are also responsible for maintaining access control lists, enforcing access authorizations, concurrency semantics, creating unique OIDs and anchor IDS, and merginganchortableinformationcreated by concurrent readers of the same node. The storage server must perform several checks before completing an open request. First, it must determine whether the user who is running the application has the correct authorizations to open the objectin therequestedaccessmode.Then,thestorage server mustdetermine whethertherequestedaccessmodeis in conflictwithanyoverlapping opens for the same object. An open requestwill fail if the user lacks proper ization orif the open conflicts with other opensin progress. mayneed tocommunicatewith m~ltiplestorage servers, at provide protection services and mappings from an m that is the custodian for that object. Object location the artifact store into nonoverlapping collections of graphs called partitions. Each partition is associated arti it ions form boundaries for administrative conam on^ servers, and re~licationof but this subvisible o~tsidethestorage serv object must same arti it ion for its entire lifetime because OlD its e partition numberof an object from its a levelof indirection (ap~rtitiondirecto-

FIG. 12.7. Struc~ureof object files.

12, A R ~ H I T E C T UOF ~ ~ E ~ E D SYSTEM I A

ry), it is possible to change the physical location ofobject an while preserving its OII) and, therefore, all its link andcompositionrelationshipswith other objects (seeFig. 12.8). Partition-location servers maintain a mapping of logical partitions to host(s) running server processes for that partition. The GCM extracts thepartition number fromthe OID of the object and uses find the host runninga storage server procthe partition location service to ess maintaining a directory for that partition (the GCM can also cache the partition location information for use in references to other objects). ~e espect that in most cases one storage server maintains both the partition directory and data storage for an object. Despite their importance, partitions are invisible to users. Only systemadministratorsandsystemprogrammers need to understand partitions. An WC interface to the storage servers is provided for administrative processes to use in creating new partitions, moving objects from one physical partitionto another, and performing backup and recovery operations. Storage servers are responsible €or managing partitions on disk, replicating partitions €or availability and fault tolerance in case of media or failures, and for recovering from most failures. The key to ourimp1 tion of fault tolerance is the ISIS system developed by Ken Birman colleagues at Cornell ~niversity(Joseph & Birman, 1986). In particular, we use ISIS process groups to maintain replicated copies of physical partitions and to provide the location in~ependenceof logical partitions. Each l o ~ i c ~ l onds to an ISIS process group. f erformance and scalability are two key requirements ent implementation withrespect to these

FIG.12.8. OID andobjectlocation.

S H A ~ ~ L FSMITH, O ~ , SMITH

have beguna seriesof benchmark experiments similarto those used to evaluate performance and scalability of distributed file systems such as AFS (~owardet al., 1988) and Sprite (Nelson et al., 19 .We have created several benchmark programs designed to stress different aspects of the system. The most interesting of these is a "synthetic browser" program that mimics the requests that result when users search for information in an artifact stored in the system. Load on the storage serviceis generated by running copies of the synthetic browseron several workstations. This program has of browsing behaviors. p~rametersthat can be used to produce a wide range In our first experiments we are using parameter values that represent the observed behavior ofhuman subjects in a series of experiments we conducted to understand how people would use a hypertext system for problem solving (Smith, 1992). With these values, each instanceof the program running on one workstation generates a load on theserver corresponding to approximately 10 users working with interactive browsing applications.We have also written an "artifact generator'' program that, based a number on of input parameters, creates a structure of subgraphs,nodes,andlinks to serve as data for the browsing benchmark. The results of our initial measurements have been very encouraging. The configuration for these measurements consisted of one storageserver running on a DECstation 5000/25c and up to seven workstations (DECstation 5000/120s) each runninga copy of the synthetic browser program. All workstations were connectedby a single ethernet segment. The most significant results are: 0 0

CPU utilization on the server is at most 0.5-1.0% per active user. Server response timesto requests from 50 users increasedby less than 20% over response times to requests from 10 users.

The results show that one server can support at least 50 users. More extensive benchmark experimentsare underway to validate this conclusion for a varietyof configurations. We are currently using the DCS for developin browsers and other collaboration support tools. We continue to make enhancements (mostly for operations and administration) and plan to havea version suitable for distribution to other groupsby Fall 1993.

In this section, we compare our design with severalhyperte~ systems that

have significant capability for supporting collaborating hn, Riley,Coombs, & e ~ o ~ 1992; i t ~Yan~elovich , etal.,

12. ARCHITECTU~OF H

~

E

~ SYSTEM E D ~

,HyperBase/CHS (Schutt & Haake, 1993; Schutt & Streitz, 1990), (Engelbart, 1984), Telesophy(Caplinger,1987;Schatz,1987), (Akscyn, McCracken, & Yoder, 1 ,and I"(Campbell & Goodman, Delisle & Schwartz, 198'7). These systems differ widely on factors such as the data model supported, scalability, concurrent reader-writer semantics, and protection. & Schwartz, 1990) supportrich DGS,HyperBase/CHS,Dexter(Halasz data models that include regates (named groups of objects), aggregates of aggregates, and aggregates as endpoints of links, Intermedia,HAM, and Augment do not use aggregatesin composition or linking. Telesophy's data modelhas ag~regatesbutdoesnotgivefirst-classstatustolinks. HB I (Schnase, Leggett,& Hicks, 1991) and Trellis (Stotts & Furuta, 1989) provide strong support for computation within hypertext but do not have aggregates. The DGS data model benefits from the graph-theoretical metaphor on which it is based and is the only system to provide strongly typed aggregate objects. are in the semanOther areas in which these systems differ substantially tics of concurrent reading and writing andin the access protection mechanisms (see Table 12.1). These systems also differin their capability to scale up to large numbersof users (and objects) while preserving the illusion of location transparency, Both Telesophy and the DGS have made scalabilitya central issuein their designs. However, theDGS provides more flexibilityin its data model and stronger consistency semantics.

Collaborative groups face many problems, but one of the hardest and most importantistomeldtheirthinkinginto a conceptual structure that has is coherent, consistent, and correct. Seeing that integrity as awhole and that construct as a single, inte rated artifact can help. But groups must also be able to view specific partsf the artifactin order to understand and manage it.Ourdesignwasguided by these requirements, along with others discussed earlier.The graph-based data model permits us to both partition the artifact and to compose those pieces to build larger components and the whole. The distributed architecture, in turn, permits us to build a system number of users, and that can scaleup in terms of the sizeof the artifact, the their g e o ~ ~ ~ pdistances hic from one another. ~e observe that mostof the academic research in hypermedia is based on the sort of modular architecture that was described at the be ning of this chapter. Althou~hmany communities view h~ermediaas an plication,wetaketheperspective(also ex~res§edin 91) that hypermedia hasa broader role to play. In our o

TABLE 12.1 C o n c u ~ n t ~ ~ e r - w rSemantics iter and Object Protect Co~urrentReaderWriter Semantics Augment HA

Objects in the Journalare readonly. Access to Journal entries can be restricted atsub~ssiontime

Can have multiple readers of documen&that have been s u b ~ ~toe the d Journal system not Could

bedetermined

Access Control Listsp optional^:

access, ~ n o ~ tupdate, e, and destroy permissions

are provided to Access control willbe based on warn a p p l i c ~ of o ~c o n c u ~ n t user roles such as manage^' and activity, but these markersare “secretary” (not yet implemen~). advisory I natw. AI1 applicatio~ are notified when data is changed, so that they can update their view (if desired).

Hy~r~as~CHS Activity markers

Inte~edia

S

T~lesop~y

Supports multipleusers d n g and Provides read, wri&, and annotate a n n o ~ n gand , a single writer. ~rmissionsthat can be First user to write an object locks users and groupsof users. out other potential writers.

Uses an opti~stic concu~ncy method. When a writer ~ e m top save a node, Wshe may be denied because someone elsehas concurrently written to the same node. In this case, the human user must manuallymeqe the two conflicting versions.

Owner can protect a h e from ~m ~ f i ~ oro read n access. In items, but notto modify e items.

~

Supports multiplec o n c u ~ n t could not be determined readers and writers. When writes overlap, the last writer completely overwrites the workof others.

S

Access Control Lists: access (read

and a n n o ~ e .

e

s

~

12. ~ C H I T E C T U OF~ H Y P E ~ E D SYSTEM ~

ion, hypermedia is not just an application, abut newisparadigm forthe way we work andcollaboratewith each other. As such,willit be an essential component of the next generationof operating system support. Our experiences with DGS strongly indicate that it is possible to achieve the richer functions needed for hypermedia storage with cost, performance, and scalability comparable to the best conventional distributed file systems AFS), (e.g., As we look to the future, additional issueswill weexplore pertainto widearea network access, dynamic change notification, graph traversal, and sup port of a richer setof graph and set operationsand queries. Many of these e~ensions lend themselvesto the sortof modular approach that is suggested in the §trawman Reference Model~hompson, 1990).

A number

of individuals and org~izationshave contributedto this project. Gordon Ferguson and Barry Ellege contributed to a Smalltalk prototype that preceded theDGS. Rajaraman ~ishnan, Shankar Krishnan, XiaofanLu, Mike ~agner, and Zhenxin ~ a n have g contributed to the implementationof the DGS. This work was supportedby the National Science Foundation (Grant# 3) and by the IBM Corporation.

Akscyn, R M,, D. L.McCracken, andE. A, Yoder (1988, July).KMS: A distributed hypermedia sysComm~nications of the ACM,31(7), 820-835. tem for managing knowledge org~izations, in Baker, M.G., J. H. Hartman, M. D. Kupfer, K. W. Shirriff, and K. J. Ousterhout (1991, October). Measurements of a distributedfile system. Operating Systems Review, Special Issue: Proceedi~s ofthe 13thACMSymp~iumon Operati~SystemsPrinciples (Pacific Grove, CA), 25(5), 198-212, Campbell, B. and J. M. Goodman (1988). HAM: A general purpose hypertext abstract machine. Communicati~s ofthe ACM, 31(7), 856-861. In O O S U '87 Caplinger, M. (1987, October). An information system based on distributed objects, Proceedi~s,pp. 126-137. Deiisle, N. M.and M. D. Schwartz (1987, April). Contexts:A partitioning concept for hypertext, ACM Transactions on Ofice informationSptems, 5(2), 168-186. Engelbart, D. C. (1984, February). Authorship provisions in A U G ~ EIn~ P. r ~ e e d i of ~ the s 1984 Conference,San Francisco,CA, pp. 465-472. Haan, B,J., P. Kahn, V. A. Riley, J. H. Coombs, andN. K. Meyrowitz (1992, January). IRIS hypermedia services.Communications of the ACM,35(1), 36-51. Halasz, E and M, Schwartz (1990). The Dexter hypertext reference model.In Proceedings ofthe ~ I ~ ~ yStandard~ation p e ~ ~ ~t o r ~ h (Gaithersburg, op MD), pp. 1-39. Howard, J.H., M. L.Kazar, S. G.Menees, D. A. Nichols, M. Satyanarayanan,R. N. Sidebotham, and ,February). Scale and performance in a distributed file system. ACM Transactions on Computer Systems, S( l), 51-81. ~~~~

S ~ C ~ L F SMITH, O ~ ,SMITH Jeffay, K., J. K Lin, J. Menges, F.D. Smith, and J. B. Smith (1992). Architecture of the artifact-based ~ sCSCW '92 Conference on C o m p u t e r ~ u ~ collaboration systemmatrix. In ~ ~ e e dofiACM ported C ~ p e ~ Work, ~ v eCSCW A ~ h i t e c ~ r epp. s , 195-202. Joseph, T. A. and K.P. Birman (1986, February), Low CO ment of replicated data in faultSystems, 4(1), 54-70, tolerantdistributedsystems. ACM ~ansactionson Kistler, J. J. and M. Satyan~ayanan (1991, October). Disconnected operation the in Coda filesys~ 13th s ACM Symp~iumon tem. Operating Systems Review, Special Issue: ~ ~ e e dofi the opera ti^ Systems ~ i n c ~ l(Pacific e s Grove, 213-225. Nelson, M. N.,B. B. Welch, and J. K.Ousterhout in the Sprite network file system.ACM T~nsactionson Computer Systems,6(1), 134-154. Schatz, B. R. (1987). Telesophy:A system for manipulatingthe knowledge of a community.In P m c e e d i ~ osf Globecom'87 (New York), pp.1181-1186. ACM. Schnase, J,L., J. J. Leggett,and D. L. Hicks (1991, October). HBk Initial des& and implementati~ of a hyperbase man~ementsystem (Tech. Rep. T ~ U - 91-003). H ~ Hypertext Research Lab, Texas A&M University. Schutt, H. and J.M. Haake (1993, March). Server support for cooperative hypermediasystems, In ~ypermedia'93,Zurich. Schutt, H. A. and N. A. Streitz (1 .Hyperbase: A hypermedia engine based on a relational i ~Es ~ European 9 Conference ~ on Hyperdatabase management system.In ~ ~ e e dofthe tat, ~atabases, Indices and Norma~ve~ n o ~ l epp. ~ 95-108. e, Shackelford,D. E.(1993, January). The ~ i s ~ i b u t Graph ed Storage System: A uses ma~ual forappliUniversi~ cation p r ~ m ~ e(Tech. r s Rep.TR93403). Department of Computer Science, The of North Carolina at Chapel Hili. Smith, D. K.(1992). Hypermedia vs. paper: Userstrategies in brows NA materials (Tech.Rep. TR92-036). Department of Computer Science, The Universityof h Carolina at Chapel Hill. Smith, J.B. and F. D. Smith (1991). ABC: A hypermedia system for artifact-based collaboration. In ~ e e d i of~ACM s Hy~rt~t'91, Cons~u and c ~Authori~, on pp. 179-192. Stotts, P. D. and R. Furuta (1989). Petri-net-based hypertext: Documentstructure with browsin semantics. ACM Transactions on Info~ation Systems, 7(1), 3-29. Thompson, C. W.(1990, January). Strawman reference model for hypermedia systems.In Pro c e e d i ~ s the o f NIST H y ~ r tStan~ard~ation ~t Workshop(Gaithersburg, MD), pp. 223-246. Yankelovich, N. et al. (1988, January). Intermedia: The concept and the construction of a seamIEEE compute^ 21(1), 81-96. less .information environment.

P A R T

This Page Intentionally Left Blank

C H A P T E R

Ronald C. Arkin Tucker Batch

Georgia Institute of Technology

e

~oordinationtheory has most often been considered to be the domain of man-machine interactions. Indeed, throughout most of this volume youwill see that the majority of research reported involves computer co~laboration with humans to some degree. The research presentedin this chapter, however, takes a different tack: It is concerned with the content of inforrnation of mobile that is necessaryforsuccessfulcooperationbetweenteams robots. Mostof the communication studies we referare to at subhuman levels (i.e., derivedfrom animal studies). At the ~eorgia Tech Mobile Robotics Laboratory, a robot system desi me tho do lo^ has been developed and refined for both single and multiagent robotic systems. These systems are implemented in both simulation and on mobile robots [3, 131. The approach relies on two key points: an objective metric of systemperformance,andaniterativecycle of simulationan in~tantiationon real systems. Through simulation, the designer can ~uickly discover which sensors, actuators, and control parameters are most critical. ~arametersare varied as performance is measured and cornpared to that of other configurations.The goal is to find a system that m ~ i ~ i z(or e sminimizes) theperfor~ancemetric. Finally, the configuration is ported to a real 'A closely related versionof this paper appeared in Issue 1.1 ofAUTONO~OUSROBOTS and is reproduced with permission.

ARIUN AND BALCH

robotic system for testing.In this chapter, the approach is applied to communication in reactive multiagent robotic systems. To discover how communication impacts multiagent robotic system performance, three societal robot tasks were devised. The performance in simulation of a team of robots is measured for eachof these tasks for threedifferenttypes of communication.Theexperiments are designed so that performance for each type of communication can be compared acrossdifferent tasks.In all, asix~imensionalspace of task, environment, and control parameters was explored including: task, communication type, number of robots, number of attractors, mass of attractors,and percentageof obstacle coverage. The simulation results were supportedby porting the control system to ateam of Denning mobile robots. e

K

Multiagent robotic systems constitutes a very active area of research. A large body of literature exists regarding systems rangingin size from two to thousands of robots. Dudek et al.[191 provide a taxonomy of these systems classified along the dimensions of group size, reconfigurability, processing ability, and communicationrange,topology, and bandwidth.Theresearch in this chapter concentrates on relatively small group sizes, typically on the of order two to ten agents. Large-scale swarm robotic systems [27] are not considered. in the According to the taxonomy cited earlier, our work fitscategoriesof LIMGROUP (small number of robots), C O ~ - ~ O Nand E ,COM-INF (robots either have no communication or can be heard within the entire range of the simulaW, tion), TOP-BRO~(broadcastcommunicationmethod), B ~ ~ L O and B ~ ~ Z E (have R O limited bandwidth or no communication), ARR-DYN (robots canreconfigurethemselvesindependently), PROC-FSA (usesfinitestate automaton processing), and homogeneous (all agents are of the same type). Fukuda was among the first to study multiagent robotic systems in the context of what he refers to as cellular robotics[24].This pioneering work is mainly concerned with heterogeneous agents, The research reported in this study is for homogeneous societies, where all the agents are functionally MIT’s AI Laboratory [16,391 have studequivalent. Recently, researchers at ied aspects of subsumption-based reactive control using robot societies consisting of up to 20 agents. In particular, learning methods have been evaluApplications of multiagent systemsare also being investigated in ated [a]. militaryenvironments in boththeUnitedStatesandEurope [42, 341. ~traterrestrialplanetary exploration has also been proposed as a useful target domain for these societies [40].

13. C O ~ ~ ~ I C A T IAND O NCOO~INATIONIN ROBOTIC T W S

Foraging has been one ofthe most widely studied tasks to date for multirobot teams. Floreano [Zl] described nest-based foraging strategies using a neuralnetwork architecture. DrogoulandFerber1201reported results of simulations of foraging robots demonstrating the spontaneous evolution of structure such as chains from extremely simple agents. A pressing question, and one that the research described here addresses, is the roleof communication in multiagent robotic systems. Arkin [5] previthe ously reported that successful task-achieving behavior can occurineven absence of communication between agents. It is the goal of the study reported in this chapterto understand what improvementsin performance canbe gained by adding communication above noncommunicative methods. Along these same lines, ~ltenburgand Pavicic created a multirobot society consisting of a group of smallrobots conductinga searchand retrieve task (one robot only per object retrieved) using either an infrared or incandescent recruitment signal. The authors reported an approximately 50% improvement in performance for target acquisition using this type of signal. The work as reportedin [1 J is very preliminary. in synthetWerner and Dyer[51] studied the evolution of communication ic agents and have demonstrated that directional mating signals can evolve in these systems given the presence of societal necessity. MacLennan 2371 hasalsostudiedthisproblemandconcludedthatcommunicationcan evolve in a societyof simple robotic agents.In his studies, the societies that evolved communication were84% fitter than thosein which communication was suppressed. An order of magnitude better performance was observed in simulation research when learning was introduced. Franklin and Harmon, conducted atERIM [22 J,used a rule-based cooperative multiagent system to study the role of communication, cooperation, and inference and these how relationships leadto specialized categoriesof cooperative systems. Regarding communication, they recognized that information need not be explicitly requested by a receiverfor it to be potentially useful to the multiagent system as awhole. Yanco studied communication specifically in the context of robotic systems. In her research [5Z], a task was defined requiring communication to coordinate tworobots, Ernie andBert. The robots have a limited vocabulary that self-organizes over time to improve the performance the of task,which involves mimicking the behavior of a leader robot. Noreils [43] described coordinatedprotocolsasabasisforencodingcommunicationsignals between robots for navigational tasks. Formal theoretical methods are also ~ a n gE501 beingapplied in a limitedway tothisproblem.Forexample, looked at distributed mutual exclusion techniques for coordinating multirobot systems. The research we report herein is motivated by the desire to create a design methodology for multiagentreactive robotic systems. To effective-

ly design these systems it is important to choose correctly the number of agents and the communication mechanisms of a robot society for a particular task. This goal is decidedly different than the studies reported earlier,

Nature offersa wealth of existing successful behaviors that robot designers can often directly apply to their work. Because communication is important in many natural societies it is appropriate to look to them for inspiration, influOur strategy for creating multiagent systems has been significantly enced by biological and ethological studies.In [101, we reported the dimensions by which communication can be described in these systems. Some in animal societiesare specificexamplesoftheroleofcommunication reported next. One of the most commonly studied social biological systems is that of ants. Excellent referenceson their social organizationand communication methods are available[29,25]. Ants typically use chemical communication toconveyinformationbetweenthem. Goss etal. E261 studiedforaging behavior in ants, creating computer models thatare capableof replicating various species’ performance for this task. Franks [23] has looked in particular at the behavior of armyants in the contextof groupretrieval of prey regarding the relationships ofmass to objects retrieved and velocity of return. Tinbergen’s influential work on social behavior in animals [49] described a range of behaviors including: simple social cooperation involving sympathetic induction (doingthe same things as others), reciprocal behavior (e.g., feeding activity), and antagonistic behavior; mating behaviors involving persuasion,appeasement,andorientation;familyandgrouplifebehaviors involving flocking, communal attack (mobs), herding behaviors, and infectious behaviors (alarm,sleep, eating); and fight-related behaviors involving re~roductivefighting (spacing rivals), mutual hostility (spacing group indipeck-orde and viduals), ing). An interesting study nmental impact foraging on behavior in fish is presentedin [18]. The factors considered include food supply, in the hunger, danger, and competition. Mob behavior and communication whiptailwallaby [31] also providesanunderstandingfor the emergent of multiple agents andthe natureof com~unicationthat suproupbehavior.Studies in primateshavebeenconducted organization of colonies [Z] relative to their environment. Finally, research in display behavior in animals C411 provides insights in relation to the state-basedcommunication mechanismsdescribed laterin this chapter.

13. COMMUNICATION AND ~ O O ~ I N A T IN I OROBOTIC ~ TEAMS

The task a robotic systemis to perform dictates to some extent the sensors and actuators required.It is not as apparent how the task impacts control system and communicationparameters.Our researchfocuses on three Consume, and Graze. Foraging consists of search~ng generic tasks: ForQ~e, the en~ronmentfor objects (referred to as attractors) and carrying them back to a central location. Consuming requires the robot to perform work on the attractorsin place, rather than carrying them back. Crazing is similar to lawn mowing; the robot or robot team must adequately cover the environment. Even more complex tasks can be constructed using these basic tasks 3.5. as bui~ding blocks. This is discussed further in Section "he Forage task for a robotis to wander about theen~ronmentlooking for items of interest (attractors).Upon encountering one ofthese attractors, the robot moves toward it, finally attaching itself. After attachment, the robot returns the object to a specified home base. Many ant species perform Forage task as they gather food. Robots performing this task wouldPO tially be suitable for garbage collection or specimen collection in a hazure 13.la shows a simulation of two robots foraging for seven attractors and returning them to a home base (the simulation en~ronmentis described in Section 6). In the simulation, obstacles are shown as large black circles, attractors are represented as small circles, and the paths of the robots are shown as solid or dashed lines. They leave dashed linesas they Forage

Consume

Graze

FIG. 13.1. Multiagent simulations of three tasks: (left to right)Forage,Consume, and Graze. The paths of the robots are marked by solid and dashed lines while obstaclesare shown as black circles. Each Simulation includes two robots and seven attractors.

ARKIN AND BALCH

wander, and solid lines when they acquire, attach, and return attractors the to home base. The mass of the attractoritem dictates how quickly a robotcan carry it. The heavier the attractor, the slower the speed. Several robots cooperating up to the m ~ i m u m speed of an indican move the attractor faster, but only vidual robot. Like Forage, the Consume task involves wandering about the environment to find attractors. Upon encountering an attractor, the robot moves toward it Forage task, however, the robot and attaches itself to the object. Unlike the in place after attachment. The time requiredto performs work on the object do the in-place work is proportional to the mass of the object.It is not necessary for the robot to carry the object back to home base. Applications might include toxic waste cleanup, assembly,or cleaning tasks. Figure 13.lb shows a simulation of two robots consuming seven attractors. Note that this task is performed in exactly the same environment as the forage task shownin Fig. 13.la. The robots leave dashed lines as they wander, and solid lines when they acquire and move to the attractors. The mass of the attractoritem dictates how quickly a robot can consume it. The heavier the attractor, the more time it takes. Several robots cooperating can consume anattractor faster.For this task the rateof consumption is linear with the numberof robots and has no ceiling. The Graze task differs fromForage and Consume in that discrete attractors are not involved. Instead, the objecttoiscompletely cover, or visit the enviarea that has not been ronment. TheGraze task fora robot is to search for an grazed, move toward it, then graze over it until the entire environment (or some percentageof it) has been covered.It is assumed that the robot possesses some means to “graze” and that it grazesover a fixed “swath,” The size of the task is dictated by the proportion of environment that must be covered before completion. Figure 1 3 . 1 ~shows a simulation of two robots grazing over 95%of the environment. The robots leave dashed lines as they wander, and solid lines when they graze. Crazing robots might be used to mow, plow, or seed fields, vacuum houses [36], or remove scrub in a lumber producing forest. The size of the swath that a robot can graze, and the percentage of the area that the robot must grazeover both affect how long it takes to complete the task. Multiple robots can complete the task fasterif they avoid traversing already grazed areas and if they can find ungrazed areas quickly.

13. COM~UNICATIONAND COO~INATIONIN ROBOTIC TEXMS

A number

of items contributeto the complete specification of a task, includ-

ing task factors, environmental factors, and the sensor and motor capabili-

ties of the robots. Table 13.1 enumerates and summarizes the parameters available in our simulation. We consider these parameters to be the most important: 0

0

0

~ u m ~ofeat~uctors. r Clearly the number of

attractors the robotsmust collect or consumewill affect how long it takes to accomplish the task. Mass of attractors. In general terms, anattractor’s mass can be thought of as a utransportability~ factor for the Forqe task, or a ‘~orkability” factor for theConsume task. Graze coverage.For the Graze task, the total size of the area and the percentage required to be grazed directly impacts the time to cover it.

Sections 7 and 8 report e~perimental results on how each ofthese factors affect performance. For this work, only the three basic tasks and the behaviors necessary for robots to performthem are considered. The results for these tasks are important because more complex tasks are easily described as combinations of simpler ones. Considera robot removing scrub froma forest; after working for a period of time, it must returnto a refueling station. The scrub removal portionof the task is analogous toGraze, whereas refueling is similar to Consume. Another complex task, ~ o u n ~ i ~ o v e ris~aamovement tc~, tactic utilizedby Army Scouts. Usually employed by two groupsof two ground vehicles, it allows safe penetration into hostile areas. Each group moves forward a short disit moves forward. A behavior tance, then waits and ‘kovers” the otherasgroup v eber built ~ a as famore c ~ specialized and coordito perform ~ ~ n ~ ~ ~ o can are selected, nated Cons~metask. Once appropriate waypoints for each group virtual attractors can be placed there. The behavior would erne twoelement group successively moves from attractor to attractor, Other research in our laboratory is underway that investigates how complex behaviors may be specified as combinations of basic behaviors[%l, The research includes a language that allows individual robots, and societies of robots to be described formally. Formal operators allow basic, or primitive, behaviors to be grouped into more complex assemblages. These assemblages are further combined to form the overall behavior of the robot.The language includes operators that coordinate individual robots into cooperating groups.

TABLE 13.1 ~ x ~ ~ m e n Values ~ P ~ t e r

men^ Range

Factor

Number of attractors Mass of to^ GrazeCoverage

5 avg 95

Obstacle Coverage Obstacle radius

15%

1 .Oto 4.0 ." "

1 to5

mum Velocity Attractor Sensor Range Obstacle Sensor Range C o ~ ~ Range ~ o C o ~ u n i c ~ Type on GrazeSwath Consume Rate

fixed n

2 Wstep

20 ft 20 ft 100 ft

No

2ft

0.01 mass fixed units/step

fixed fixed fixed No,State, Goal fixed

Obstacle Sphen: of Influence 5At Obstacle Repulsion Gain l .o Robot Repulsion Sphere 20 ft Robot Repuision Gain( w ~ r ~ 0.5 Robot Repulsion Gain (ucqu~re) 0.l Robot Repulsion Gain (&fiver, graze) 0.1 (ac uire) 1.o (~e~ve~) l .o Rate ( c ~ ~ ~ u ~ e ) 1*o

~ u t e Unless . nded otherwise, the values are t h e same for all the tasks.

fixed fixed fixed fixed fixed fiied fixed fixed fixed fixed

13. C O ~ ~ U N I C A ~AND O N COORDINATIONIN ROBOTIC TEAMS

For clarity, we describe the robot behaviors somewhat less formally thanin this related work, but the same recursive philosophy applies.

A schema-based reactive control system is used in this research.To provide

the reader appropriate background,a brief summary of reactive controlis first provided, followed by some of the special characteristics of schemabased systems. Reactive control is a paradigm that emerged in the mid-1~80sas a new approach to controlling robots. It arose in response to the perceived problems in hierarchical robotic control systems that required a heavy reliance on internal world models. Reactive behavior-based control avoidsas much as possible the use of symbolic representations of the world, preferring manytightsensorimotorcouplingsthatgroundtherobot’sperceptions directly. Reactive control is characterizedby several distinct features: The basic component isa behavior consistingof a coordinated perceptual and motor process. Perception and actionare tightly coupled. Reliance on explicit world models and representational knowledge is avoided during execution. * They are particularly well-suited for dynamic and unstructured do~ains because they rely entirely on immediately perceivedsensory data.

8

8

8

Brooks’ subsumption architecture is one well-known example of this control paradi~m[151. Other representative examples include[3,30,45,38,48]. These reactive strategies differ in several significant ways inclu~in~ the organization and nature of the underlyin~behaviors and whether arbitration, action-selection, or concurrent processing is used to select of thewhich behaviors are active at any given moment. Space prevents a complete tutorial on reactive systems,so the interested reader is referred to[S]for a more complete review. Schema-based reactive control has been widely used with success in our laboratory for both simulation studies and real robot implementations [3,5, 12, 13, 141. Some features distinguishing schema-based robotic control from other reactive approaches include: 8

A dynamic networkof processes (schemas)Is used rather than a strictly layered system such as foundin subsumption.

ARKIN AND BALCM

‘Instead of choosing only one behavior (arbitration), several behaviors

are allowedto concurrently contribute to the overall action of the robot. Potential field techniques[32,33]are used to encode the robot’s behavby gravitational or ioral response. Forces analagous to those generated electrical fields repulse or attract the robotas a resultof various environmental stimuli. * Adaptation and learningare facilitated through this flexibility [17,46,47] by permitting access to theunderlying numeric parameters of the control system. * ~euroscientific, psychological, and ethological studies provide motivation for schema use.[7]

0

In schema-based control, each of the active behaviors (motor schemas) computes its reaction to its perceptual stimuli using a method analogous to potential fields [3]. The traditional potential field method computes the attracenerated by anattractorand combines them with repulsive forces generated from obstacles, computing a global force field based on potential energy computations drawn from gravitational or electrostatic analogues.It must be noted that unlike traditional potential fields, in our work only the robot’s immediate reaction at its current location and its current perceptions of the world is computed. All of the independent behavioral vector force computations are summed and normalized and then sent to the robot for execution. This perceive-react cycle is repeated as rapidly as possible. Problems with local minima, maxima, and cyclic behavior which are endemic to many potential fields strategies are handled by several methods including: the injection of noise into the system [3];resorting to high-level planning [6];repulsion from previously visited locales [14]; continuous adaptation [17]; and other learning strategies[46,47].The Appendix contains information on the specific computations of the individual schemas used in this research. Individual schemas are primitive behaviors that are combined to generate more complex emergent behaviors. A group of schemas that together result in a task-achievingbehavioriscalledan f f s s e ~ ~ Behavioral l~~e. assemblages are often arranged in a sequence, so that the accomplished in a step-by-step manner with each assembla robot accomplish one step of the task. Assemblages for accomplishing the ~ o ~C o~ ~ es ~,and ~ eGraze , taskes are describedin the next section.

~perimental results were generated for the tasks described in Section 3by comparing performance of proposed robotic systems to baseline, or control, performance results, The baseline data was computed by first selecting

13. COMMUNICATIONAND COO~lNATlONIN ROBOTIC TEAMS

a reasonable set of control parameters, then running a statistically significant number of simulations. Values for these parametersare based on previous research 151. In this section, the behaviors for executing the three tasks (Forage, Consu~e,and Graze) andtheirbaselineparameters are described. At the highest level, the tasks themselves are assemblages that are represented as finite state acceptors(FSAs) consisting of several states. FSAs provide an easy means for both expressing and reasoning about behavioral sets by providing formal semantics [ll]. Each state corresponds to a separate assemblage in which a constituent setof motor schemas is instantiated if that particular state is active. ~erce~tuffL Tr@ers cause transitions between states. Each active motor schema hasa perceptual schema associated with it to provide the information necessary for the robot to interact with its environment.

For the Forage task, the robots can be in one of three states:~ ~ n d e r , f f c ~ ~ i r e , and ~eLiuer.All robots begin in the ~ f f ~ dstate. e r If there are no attractors within the robot's field of view, the robot remains in ~ f f n ~until e r one is encountered. When an attractor is encountered, a transition to theffc~uire state is triggered. While in the ffc~uire state, the robot moves towards the attractor and when it is sufficiently close, attaches to it. The laststate, d e b er, is triggered when the robot attaches to the attractor. Whilein the deLiuer state the robot carries the attractor back to home base. Upon reaching home base, the robot deposits the attractor there and reverts back to the ~ f f ~ d e r state. Figure 13.2 shows theFSA for Forage? For each state, the active schemas and their parameters are: Wff~der State noise: high gain, moderate persistence to cover wide a area of the environment. avoid-static-obstac~e for objects: sufficiently high to avoid collisions. avoid-static-obstacle for robot^:^ moderately high repulsion to force individual robots apartand more efficientlycover the environment. state when the detect-attractor: perceptual schema that triggers acquire the robot sensesan attractor. 2This taskwas described earlier in [g]. The "forage" state mentioned there corresponds to the Ywander" state here. 3Avoid-sta~ic~bstacle Is also used for nonthreatening moving objects. Otherschemas such as escape and dodge can be used for noncooperative moving objects when appropriate.

ARKIN AND BALCH

FIG. 13.2. The Forage FSA.

~c9uireState noise: low gain,to deal with local minima. high to avoid collisions. avoid-static-obstacle for objects: sufficiently avoid-static-obstacle for robots: very low gain, to allow robots to converge onthesame attractor andthus cooperate, butavoidcollidingwithone another. to the detected attractor. move-to-goal: high gain to move the robot detect-attachment: a perceptual schema that triggers a state transition to deliuer when the robotis close enoughto attach to the attractor. * ~ e l i v e State r noise: asin acquire, low gain to deal with local minima. avoid-stat~c-obstacle for objects:inasacquire, sufficientlyhigh to avoid collisions. avoid-static~bstacle for robots: same asin acquire. move-to-goal: high gain, with home baseas the target. detect-deposit: a perceptual schema that triggersa state change when the robot reaches home base. Specific values used for schema gains and parameters in this study are listed in Table 13.1in Section 3.4 (see Appendix for additional information on gains and parameters).

The FSA and behaviors for theConsume task (Fig. 13.3) are similar to those used in Forage. In fact, the schemasand their gains are identicalin the wander and ac9uire states. The consume state, however, is unique to to this behavior. In the consume state, only one motor schema, consum~attractor is activated. It reduces the mass of the attractor at a fixed rate over time. When the attractor is fully consumed (mass zero) it is deactivated and the robot transitions back to the wander state. Theonly parameter applicablein

13. C O ~ ~ ~ N I C A AND T I OCOO~INATION ~~ rrv ROBOTICTEAMS

FIG. 13.3. The ~ o n s u FSA, ~e

the Consume state is the rate at which an attractor is consumed. This value is fixed at 0.01 units/time step for all experiments (Table 13.1 in Section 3.4). For the Graze task, thewander and acquire states areagain similarto thoseof Forage and Consume. The primary difference is that detect-attractor in the

ande er state is replaced with a similar detect-ungrazed-area schema. Detectungr~ed-areahasthesamefixedsensorrangeas detect-attra~tor, butit detects ~ n g r ~areas e d insteadof attra~tors.Each robot startsin the er state and searches for ungrazed areas.Upon encountering one, ittr~sitions to theacquire state and moves toward it. When the robot arrives at the graze site, it tr~sitions to thegraze state.The graze state is quite different from the corresponding states in the other FSAs. While in the graze state, the robot tends to move along its current heading as it “grazes” over a fixed swath of the environm~nt. As long as there continues to be ungrazed areas directly ahead, the robot remainsin the graze state. The active schemas for this state are:

noise: low gain, to deal with local minima. avoid-static-obsta~le for objects:high enough to avoid collisions. avoid-static~bstacle for robots:very low, to allow robotsto graze close by, but avoid collisions. probe: moderate gain,to encourage the robot to keep moving along its current heading towards ungrazed areas. graze: performs the actual graze operation over afixed swath. detect-grazed-area:perceptualschemathattriggers astate chan~e once the robot has completely grazed the local area. For simulation purposes,Graze is implemented by maintaining and marking a high resolution grid corresponding to the environment. Initially, the

entire grid is markedas ungrazed. As robots graze, they mark visitedareas on the grid accordingly.

ARKIN AND BALCH

FIG. 13.4. The Graze FSA.

Gains and parameters for each of the schemas activein the graze state are listed in Table 13.1,

Three different types of communication are evaluated in this research.Using a minimalist philosophy, the first type actually involves no direct communication between the agents. The second type allows for the transmission of state information between agents in a manner similar to that found in display behaviorin animals 1411.The third type (goal communication) requires the transmitting agentto recognize and broadcast the location of anattractor when one is located within detectable range. Each of these formsof eommunication is described next. For this typeof multiagent society no direct communication is allowed. The robots are able to discriminate internally three perceptual classes: other robots, attractors,and obstacles.None of this information, however, is communicated to other agents.Each robot must rely entirely on its own perception of the world. Arkin has shown in previous work[S] that this basic information is enough to support cooperationin robot retrieval tasks( ~ o r ~ e ) . Cooperation in this context refers to the observed phenomena of recruitment, where multiple agents converge togetherto work on the same task. The baseline results (Section7) show that cooperation also emerges in the C o ~ s u and ~ e Graze tasks as well. When state communication is permitted,robots are able to detect the intere r , or deliver) of other robots.For the results reportnal state ( ~ f f ~ dacquire, ed here, thecommunication is even simpler thanthat, where only one of bit

13. C O ~ M U N I C ~ T I OAND N COO~INATIONIN ROBOTIC TEAMS

data is transmitted: with zero indicative of an agent beingin the wander state and one indicating that itin is any state other than wander (i.e.,acquire, deliver, c o n s ~ ~ore gaze). , In [S], this typeof communicatio~ was shownto provide a distinct advantage over no communication for performance of the Forage task.Communicationisoftenconsidered a deliberate act, but state communication is not necessarily “intentional” because information can be relayed by passive observation.The sender does not necessarily explicitly broadcast its state, but allows others to observe it. In nature this type of communicationisdemonstratedwhenananimalchangesitspostureor external appearance, such as a dog raising its hackles or exhibiting flight behavior in response to fear. To take advantageof state informationin reactive controI, the behavioral assemblages for each task are modified slightly. Froma robot’s point of view, the most important states to look for in another robot are those where the other robot has found an attractor or an area to graze; that means that the other robot has found useful work. If the robot goes to the same location, it be able to assist cooperativeis likely tofind useful work as well, or at least ly. The appropriate statesare acquire, deliuer,consu~e,or graze; in the wander state the robot has not yet found any workto do. For all three tasks, the behaviors are modified so that a robotwill transition to acquire if it discovers another robotin acquire, ~eliver, consu~e, or graze. Inasmuch as the robot may not yet know the locationof the attractor, it follows the other robot instead. Once the attractoris detectable it heads directly for it.

Goal communication involves the transmission and reception of specific goaloriented information. Implementation on mobile robots requires data to be encoded, transmitted, received, and decoded. Goal communication differs from the other two levels in that the sender must deliberately send or broadcast the information. A natural example of this type ofcommunication is found in the behaviorof honeybees. When a bee discovers a rich sourceof nectar, it returns to the hive and communicates the location with a “dance“ which encodes the direction and distance from the hive to the source. For reactive control, goal communication is implemented by modifying the behavioral assemblages in the same manneras described for state communication. However, instead of following the transmitting robot that discovered the attractor, a receiving robot moves directly toward the communicated locationof the attractor.The intent is that the agent may now follow a more direct path (beeline) to the attractor. This very rudimentary form of communication only broadcasts the goal that the transmitting agent is involved with. Another modeof communica-

A M N AND BALCH

tion, not yet explored, involves the transmissionof all detected attractors independent of whether the transmitting agent is already acquiringor delivering one. This would present more options for the receiving agent, perhaps choosing to move to the closest attractor independent of whether or not the transmitting agent would benefit from its help. This additional form of communication is left for future work. The implementation of goal and state communication requires explicit signaling and reception of the communicated information. State communication can be implemented simplyby mounting a binary signal atop the robot which is either on or off depending on the robot’s internal state. This communication, although trivial, is explicit as it requires the deliberate act of invoking the signal. Information pertinent to cooperation might be gatheredby other means as well.The internalstate of a robot could be inferred by observing its mowrnent (e.g., recognizinga robotin the wander state due to apparent random movements), thereby placing a larger perceptual burden on the receiving agent.Robotscanalsocommunicatethroughtheirenvironment. In the graze task, robots leave evidence of their passage since the places they visit are modified. This fact is observable by the other robots. These types of communicationare referred to as j ~ ~ Z ~asc jthey t do not require a deliberate act of transmission. Implicit ~ommunication was found to be an important mode of cooperation in simulations of the graze task. Because this communication emerges from the interactionof the agent and the environment, it cannot be “turned off.” Thus comparative analyses of performance with and without implicit communicationare not meaningful. E

The simulation environment should provide an accurate estimate of robot performance in the real world. Simulation is important because it offers a means to test many robot system configurations quickly. To be useful, the simulation must report performance in terms of the prescribed performance metric and realistically emulate the environment and the robot’s interaction with it. Furthermore, the simulation must allow hardware, control, and environmental variables to be readily manipulated. Each robot (i.e., agent), is an identical vehicle controlled by one of the task assemblages described before. Each agent’s currentstate, however,is dependent solelyon its own perception. Therobots execute their tasks in a

13. COM~UNICATIONAND COO~INATIONIN ROBOTIC TEAMS

64 x 64 unit environment. The units are dimensionless, but for convenience of comparison to real robot implementations they represent1 foot. Time is measured in steps. Each step is one iteration of the program that calculates the robots’next positions. The robots are able to sense their locationin the en~ronment,and detect obstacles, attractors and other robots within a fixed radius field of view. They are able to grasp and carry attractors, consume attractors, or graze as the task dictates. The simulation automatica~ly enforces the limits and rules set forth in the task specifications, aswell as sensor-actuator limits. The robots are allowed to move without~estriction within the 64 x 64 environment, but they may not move outside of it.

What is ‘~erformance~? Because one goal of this research is to report the impact of communication on robotic societies,performance must be objectively measurable. §election of a performance metric is important because these metrics are oftenin competition (i.e., cost vs. reliability). Some potential metrics for multiagent robotic systems are: Cost: Build a system to accomplish the task for theminimum cost. This may be appropriate for many industrial tasks. Use of this metric will tendtoreducethe cost of thesystemandminimizethenumberof robots used. * Time: Build a system to accomplish the taskin minimum time. This metric will lead to a solution calling for m the~ i m u m number ofrobots that can operate without interference. * Energy: Complete the task using the smallest amount of ener appropriatein situations where energy stores arelimited (e.g., space or undersea applications). * ~eliabili~-§urvivability: Build a systemthat will havethegreatest probabilit~to complete the task even at the expense of time or cost. This may be useful for certaintactical military appIications.

*

The task metric can also be a numeric combination of several measurements, Whatever the metricis, it must be measurable, especiallyin simulation. For thisresearch,time to complete the task was chosen as the primary performance metric. It is easily and accurately measurable and conforms to what is frequently thought of as performance. No claim is made ho~ever best metric; robot path lengthor energy consum~tionmay be that this is the equally useful. In the simulation studies described herein, performance is measured by countinghowmanyiterations the simulation program executes before the task is completed.

A W N AND BALCH

There are a few initial conditions for some tasks that prevent the robots from completing it. For example,if an attractor was somehow placed within a circleof obstacles, the robots would never be able to reach it. Sucha scenario is not solvableby any robot system without the capacity to move the obstacles. Other scenarios, however, may ultimately be solvable, but may potentially defeat the purely reactive strategies presented here. To provide for these situations, the simulation is allowed to continue for 8,000 steps before failure is declared.As most runs completein less than 2,000 steps, it isikelythatthesystem will never completethetask if itdoesnotdo so failureisdeclared.Theobjectiveis to evaluatethe impactcommunication makes on performance,so it is not important to know why the In cases system failed, just to measure how it improves with communication. of failure, the run is recorded as having taken 8,000 steps. This approach reports optimistic performance because the run might never have completed (infinite steps). But, to show improvement over a failure case, the system must actually complete the taskand in less than8,000steps. ~ v i r o ~ m e ~Factors tai As much as can

be known about the target system’s operating environment should be incorporated into the design process for the control system. If these factors are known a priori, they can be included in the simulation. Important environmental factors include: Mobility factors:Is the terrain mountainous or flat? What percent of the environment isserved by roadways? Obstacle coverage: What percent of the environment is cluttered with obstacles? Metric a priori knowledge: Does the robot havea good map of the area or is it completely unknown? Static or dynamic:Is the environment filled with moving objects, thus reducing the utilityof maps, or is the environmenta static one? For this study, a static flat environment with randomly scattered obstacles is assumed.No a priori knowledgeof the obstacles’ location is available. Obstacle coverage is varied from5%to 20%of the total area, with15% as a baseline. otor and Sensor

As a step in the robot system design methodology, realistic bounds on the

expected motor and sensor capabilities of robots are set. These bounds help reduce the search space for an optimum solution. The affect of communica-

13. COMMUNICATIONAND C O O ~ I N A ~ OINNROBOTIC TEAMS

tion on performance is the main thrust of this research, so fixed values repIf the goal were resenting the expected capabilities of the robots were used. todetermineoptimalsensorormotorrequirements,thoseparameters could be varied as well. Table 13.1 (Section 3.4) shows the experimental motor and sensor values usedin the simulations. 7.

To build a baseline database of performance measurements,a configuration of environment,control,andtaskparameterswasselectedempirically in the (Table 13.1). The baseline databaseserves as a control for comparison evaluation of the communication experiments described next. The database is generated by runnin~thesimulationusingthebaselineconfiguration Forage, C o ~ ~and ~ Graze. ~ e ,For each parameters for each of the three tasks: task, the numberof robots and the number ofattractor objects (or percenta om bin at ion of robots and attracage of graze coverage) is varied. For each tors, a measure of performance is taken by timing runs on 30 different randomly generated scenarios. Overall performanceis the averageof those 30 runs. For each run, the simulation records the number of steps taken, and whether or not the run timed-out (failed). The baseline performance measurements were made with no communication allowed between the robots. This control is then compared with the when state orgoal communication is performance in each of the three tasks allowed (Section 8). From these comparisons, one can see quantitatively how these modes of communication impact performance. Performance data is visualized as a ~imensionalsurface with the X axis reflecting the number of robots and the Y axis indicating the number of attractors orpercent coverage4 (seeFig. 13.5). The Z,or height, axis shows the average time to complete the task for that combination of robots and attractors (smaller numbers are indicative of betterPerformance).Each point on the surface represents the the results of 30 simulation runs. The plots for all three tasks share a similar shape. Notice that the back three surfaces. Thisisexpected leftcorner is thehighestpointonthe because that location represents the case where one robot by itself must complete the most work (seven attractors for forage and consume, 95% cov‘For Graze,the percentof area to be grazed is varied in increments of 13.57%. This allowsthe difficulty to be varied in seven discrete steps from 13.57% to 95%. Results can be directly compared to Forage and Consume tasks with oneto seven attractors.

FIG. 13.5. Time to complete a task in movement steps, for one to five robots and one toseven attractors with no communication. From leftto right: plots for the Forage, Consume, and Gruze tasks. Lower numbers in the height axis (movement steps) indicate better performance. Noticethat the shapes of the graphs are similar, indicating that the tasks share a common relationship between performance, numbers of robots and numbers of attractors. The highest point (worst performance), at the back, is the case where only one robot must complete the task alone. The lowest point, at the front, is the case where five robots share the least amountof work.

erage for graze). Similarly, the right front is the lowest point, as the largest number of robots (five) complete the least amount of work (one attractor), It is also apparent for all three tasks that performance initially improves In some cases, persharply as more robots are added, but then tapers off. formance does not improve much at all with more than fourrobots. This is important if robots are expensive. To illustrate, suppose a robotic system forForage the task should be both fast and inexpensive. Performance is thena combinat~on of the timeto complete the task and thecost of the system. Ultimately, the designer must balance thei~portanceof cost versus speedof completion, but one approach is to amortize thecost of the robotic system over its expected lifetime. Thus the cost of one run is the overall cost divided by the expected number of runs. For this example, suppose the amortized cost of each robot per run is valued the same as 300 time steps. Then if N is the number ofrobots, and 7' is the time to complete the task, the overall performance is: P=N*300+7' Using timing measurements taken forForage and adding in amortized cost, a ~imensionalsurface is generatedforthenewperformancemetric(Fig. 13.6). A system with tworobots is generally best for three or more attractors. If the environment is expected to contain only one or two attractors, one robot is the best choice.Even though morerobots may be faster, the overall goals ofthe designer may call for fewer.

13. COMMU~CATIONAND COO~INATIONIN ROBOTIC T W S FIG. 13.6. Optimizing in the Forage task for t h e and cost. This graph illustrates a more complex performance metric than time steps alone. Performance here is defined as time to complete the task plus the number of robots times robotcost (300 units). Lawer points on the graph indicatebetterperformance. For one or two attractors, overall performance is optimizedwith one robot, while two robots are better with three or more attractors.

*

i

0

iI

Another effective tool is spee~upmeasurement. A plot of speedup reveals how much more efficient several robots are than just one in ~ompletinga task. If fli,]] is the performance fori robots and] attractors, the speedup at that point is:

So, if two robots complete the task exactly twice as fast as one robot, speedup is 1.0 (higher numbers are better). Mataric introduced a similar e is equalto 1.0, the metric of robot perform~cein [39].A n ~ h e r speedup performance is saidto be &near.~u~eriinear performance is greater than 1.0, and su~iinearis less than 1.0. Realize, however, that in some cases more robots will be faster for actual task completion time, but still offer sublinear speedup. Figure 13.7 shows speedup plots forForage, Consume,and Graze without communication. Note that speedup for all tasksis generally higher for larger numbers of attractors. Researchers in other branches ofcom ence have found that randomized search tasks are often completedin superlinear time on parallel systems [28].Because the~ a n ~behavior er usedin all three tasksesse~tiallysolves a randomized search task, itis not s~rprising that performanceis superlinear when this behavioris heavily utilize^, as is the case when thereare large numbersof attractors. Surprisin~ly, spe~dupin the Consume task is sublinear at all but one point (Fig. 13.7b).Thebehavior in the cons~mestate canatmostofferlinear speedup (the limit is set by the specification of the task).So an en~ronment with rnassive attractors will force the speedup to be limited near 1.0. This

ARKIN AND BALCW

FIG. 13.7. Speedup in a task as the number of robots and attractors are var-

ied with no commun~cation. Left to right: plots forthe &rage, Consume, and Graze tasks. Speedup values greater than1.0 indicate Nrobots perform more than N times better than one robot alonein that situation.Of the three tasks, Graze shows the bestspeedup.

hypothesis was testedby reducing the average massof the attractors, then rerunning the simulations.In the baseline runs, attractor mass varies from 2.0 to 8.0 units, but for these experimental runs, mass was reduced to 1.0 to 4.0 units. Reducing attractor mass allows the robots to spend more time wander in^ (a superlinear task) insteadof consuming (at most linear). The speedup for C o ~ s with ~ ~ lower e mass attractors is shown in Fig. 13.8. At every point on the surface, speedup is better for low mass attractors than for high mass. In fact, in many cases speedup is superlinear. on the surSpeedup in the Graze task is superlinearat all but three points face (Fig. 13.7). In the very worst case, speedup dips to 0.97. Situations requira high percentage of graze coverage resultin the best speedup; the peak is 1.21 for five robots and 95%coverage. In cases wherehigh graze coverage

FIG. 13.8. Comparison of speedup in the Consume task without communication for attractors ofdifferentmass.Speedupwhen attractors average 5.0 mass units, left, and 2.5 on the right. Overall speedup is better in the case of the lower mass attractors.

13. C O ~ ~ U N I C A T I OAND N C O O ~ I N A T ~ OIN N ROBOTIC TEAMS

is required, robots spend more time in wander as theylook for the last bit of area to graze. Again, because wander is a superlinear time task, the best speedups should be expected for those regimes. Speedup resultsare summarized in Table 13.2. 7.3. Timeo~ts

A ~ ~ eoccurs o ~ when t a simulation

run exceedsa time limit (for these experiments, the limit is8,000 steps). A timeout mechanism isnecessary to avoid lockups in infinite loops in the event the society is unable to complete the task for that particular random world. Frequencyof timeouts for each combination of robots and attractors is measured and plottedin Fig. 13.9. The frequency of timeouts serves primarily asa measure of data quality.In situations where timeout frequency is higher, the experimenter cannot know for sure how long the runs would have takenif they were allowedto complete. Some runs may have completed whereas others may have run inde~initely. When there are relatively fewtimeouts,theperformance is knownwith greater certainty. As would be expected, most timeouts occur when fewer robots must solve a task with more attractors or a higher graze coverage requirement. Baseline resultsserve as control a for experimental comparison in assessing the impact of other communication modes on performance. It is important to derive and understandfully these basic results before testing more complex robot configurations.Impo~antresults for the baseline configuration are: For a given number of attractors, more robots complete a task faster than fewer robots. * For a given number of robots, it takes longer to complete a task with more attractors.

*

S

Co~me o

TABU 13.2

speedup ~

Average ~ p e e ~ ~ p

Task

C

U of

~ (low ~ mzw) e

Graz~

0.93 0.82 0.89 1.07

Best

Worst

1.15 1.01 1.26 1.21

0.640.65

0.66 0.97

FIG. 13.9. Percentage of runs that end in a timeout for one tofive robots and one to sevenattractors withno communicat~on.A timeout occurs when the task is notcompletedwithin 8O , OO timesteps. Lower numbersarebetter. Timeouts occur more frequently for small numbers of robots with a large numbers of attractors. From left to right: plots for Forage, Consume,and Gruze tasks.

Some performance metrics may result in a system that is optimized with lower numbers of robots than for other metrics. Speedup is greaterin scenarios where larger numbers of attractors are present. Speedup in the Consume task is mostly sublinear, but can be superlinear for lower mass attractors. Speedup in the Graze task is mostly superlinear. Timeouts occur more often for low numbers of robots and high numbers of attractors.

. Fi~ure13.10 showsa typical simulation runof two robotsfora gin^ for seven attractors with no, state, and goal communication. Insp~ctin~ the ima~es ht reveals an apparent improvement in the ‘~rderliness~ of the The quantitative experimental results sum~arizedin Table 13.3 confirm these qualitativeim~ressions. Figure 13.5a in Section 7.2 shows a ical performance plot for Forage, in thiscasefornocommunication(beperformanceislower).Eachdata nts the resultsof 30 different simulation runs. The plots for no, a1communication are quite similar in contourbut there is improvement in performance evidencedby lower surfacesas the communication becomes more complex. The statistical analysis in Table 13.3 summarizes these observations.

FIG. 13.10. Typical simu~ation runs of the Forage task. From leftto right: runs with No,State, and Goal~ommunication. The overall distance the robots travel, qualitatively apparentas the lengthof their trails,is reduced as more compIex communication is permitted. The simulations required5,145, 4,470, and 3,495 steps, respectlvely, to complete.

S

TABLE 13.3 u of Performance ~ ~Ratios forNo, State, and Goal Communication

Task

Average ~mpr~vement Worst

Best

Forage State vs. No Co~unication66%

Goal vs. No C o ~ u n ~ c a t i o59% n 34% Goal vs. State Co~unication

46% State vs. No Co~unication

Goal vs.-16% No co~unication44% Goal vs. State Communication Goal vs. State (low mass attractors)

16% 19% 3%

10% 6% -4% -19% -1%

-5%

-7%

-19%

-9% 5%

-30%

19% 19%

0% 0% 0%

23%

raze State vs. No Communication

Goal vs. No Co~unication 0% Goal vs. State communication

1% 1% 0%

ARKIN AND BALCH

To quantify the difference between performance with and without com(Fig. 13.11). At each point, munication, a performance ratio plot is computed by the performanc~ without the performance with communication is divided communication. Results greater than 1.0 imply improved performance. For instance, a valueof 1.1 indicates 10%improvement. For all the cases tested, State communication improved performancein the Forage task an average of 16%. On the average,goal communication is 3% better than statecommunication in the Forage task.

The impact of communication on performance of the Consu~etask is similar to that in Forage. Figure 13.12 shows a typical simulation of two robots consuming seven attractors with no, state, and goal communication. A surprising result is that the simulation with goal communication actually takes longer than the one with state communication. This slight increasein run time with goalversus statecommunication is typical for this task. A representative exampleof the basic performance data for simulations in Fig. 13.5b, Section 7.2. Again, the conof the Consu~etask is plotted earlier tours for all three forms of communicationare quite similar.A comparative analysis reveals thaton the average, state communication offers10% a performan~eadvantageovernocommunication.Goalcommunicationis 4% worse on the average than state communication. Goal communication, however, is still 6% better than no communication at all. Table 13.3 summarizes these results. Recall that speedup in the Consu~etask is linkedto attractormass (Section 4). Attractor mass may also impact the benefit of communication. Analysis of the data from runs with low mass attractors reveals that goal communication pe~ormanceis almostindistin~ishablefrom that ofstate communication (1% worse). Future research may determineif this result is justan anomaly or if en~ronmentaland task parameters might shift this trend.

FIG. 13.11. Performanceratioplot for the Forage task for Goal versus State communication.A ratio of 1.0 indicates performance is the same. Largernumbersindicate better performancewithGoalcommunication.

13.COMMUNICATION

AND COO~INATIONIN ROBOTICTEAMS

FIG.13.12.The Consume taskwith No, State, andGoalCommunication.The case with State Communication is visibly better than the simulation with No Communication. Performance with State and Goal Communication is about the same. The simulations required4,200,3,340and 3,355 steps, respectively, to complete.

8.3. ~ o m ~ u n i c a t i oinnthe Graze Task

The surprising result from Graze task simulations is that communication hardly helps at all. Plots of basic performance data for each of the different levels of communicationare not shown because theyare visually identical (see Fig. 13.5~for the case with no communication).On average, statecommunication is only1% better than no communication. Performance with goal communication is virtually indistinguishable from that with state communication (0%difference). Table 13.3 summarizes these results. As robots graze they inevitably leave a record of their passage: the graze swath. This physical changein the environment is actuallya form of implicit communication. The robots leave marks that advise others where work has or has not been completed. This result is important because it implies that for tasks where such implicit communication is available, explicit corn~unicationis unnecessary. The p~rformanceimprovement each typeof communication offers for each task are summarized in Table 13.3. Several important conclusions may be drawn: ~ommunication improves performance significantlyin tasks with little implicit communication(Forage and C o ~ s ~ ~ e ) . * ~ommunication appears unnecessary in tasks for which implicit communication exists(Graze).

AIUUN AND BALCH

ore complex communication strategies (Goal) offer little benefit over basic (State) communication for these tasks (Le., display behaviora is rich communication method).

The ultimate goal of this research is a working multiagent robotic system; simulation serves only as a development tool. To demonstrate the sirnulation results, andto move towarda completely functiona~ society, the behave , Graze must be instantiated on mobile robots. iors forForage, C o n s ~ ~and Our laboratoryisequippedwiththreemobilerobotsbuilt by Denning Mobile Robotics: George, Ren, and Stimpy. Thesethre~wheeledmotorized vehicles are appro~mately32inches in diameterwithRenandStirnpy measuring3’feet talland George standing4 feet. Theyare equipped with sensorsthattracktheirposition(shaftencoders),andobstacledetection devices (ultrasonic range sensors). The Forage task described in Section 4 was ported and tested on Ren and st of the required schemas had already been coded, but the lack ng omnidirec~ional sensor system for attractor and robot detecby simulating the tion complicated matters. The problem was circumvented sensor within an embedded perceptual schema utilizing shaft encoder data. Spatial locationsof attractors and moving robots are maintained in continuously updated shared files. Fidelityis maintained by coding the perceptual schema so that it does not ‘keveal” the location of attractors or other robots until theyare within sensor range. A two robot run of the Forage task is shown in Fig. 13.13, Mostof the parameters are those from the baseline simulation runs, but because the test area is rather small, attractor sensor range was reduced from 20 to 10 feet. The range at whicha robot begins to be repulsed from an obstacle (the sphere of influence) was set 2atfeet. Thereare three attractors (boxes) and one obstacle (chair) in the environment. Both robots were initialized at home base. This run was made without communication. At the beginning of ~ r and are repulsed by the run (Fig. 13.13), the robots enter thew f f n ~state, each other. They immediately detect separate attractors. After tagging their r e s ~ e ~ tattractors, i~e the robots deliver them to homebase.Againthe robots cycle towander. Only one attractor remains(in the foreground). The so Ren ttractor iswithinRen’s sensorrange,butoutsideStimpy’s, ap~roachesit alone.As Ren returns the attractor to home base within Stimpy’s sensor range. Stimpy responds by approachi

13. ~OMMUNI~ATIONAND ~OO~INATION IN ROBOTIC T W S

FIG, 13.13. Two Denning robots, RenandStimpy,demonstrate the Forage task (upper left). Ren tags an attractor (upper right). Stimpy “tags” an attractor (lower left). Ren and Stimpy deliver the attractors to home base (lower right).

helping to deliver the attractor. A (hand-drawn) reconstructionof this run is shown in Fig. 13.1~.

All three levels of communication for the Consume task have been implemented and tested on Ren and Stimpy. A scenario for the two robots with one attractorwas usedin testing theConsume behavior (Fig. 13.15). ~ ~ t h o u ~ h

of and the qualithe scenario is simple it serves to illustrate the advantages tative differences between the three levels communication described in Section 5. Runs on mobilerobots ared i r e ~ t ~ y ~ o mwith ~ a rsim~lations ed of the same scenarioin Fig. 13.15. In the test scenario, two ro~otsand one attractor are arrange one robot is immediate~y within sensor rangeof the attractor,wh other is justoutsi~esensor ran~e.In the s~mulations, the attractor is 20 feet ot. If no communication is allowed, one robotshoul~inithe attractor.The other robot should move a ~ a ydue , to If commu~icationis allowed, both robots shoul interrobot rep~is~on, tially move towardthe attractor~ecauseat least oneof them sense^ it.

).tome

0

AllfBmn

FIG. 13.14. A reconstruction (from above) of the Forage demonstration with two mobile robots. Initially (left) the robots arerepulsed from one another and detect separate attractors, which they deliver to homebase (center). Later (right)they cooperatein returning the last attractor to homebase.

Robot 2

.. * : ,.".*.: !*"."Q

I.

"";Robot

1

Robot 2

Robot 2

f

Attractor

Rcn

?LQ

Robot 1

Attractor

Robot l Attractor

Rcn

"----do

Stimpy

Attractor

Stimpy Attractor

FIG. 13.15. Comparison of simulated Consume task runs (top row) with runs on mobile robots (bottom row). The traces made by the simulated and real robots for each level of communication are similar. The differences are primarily attributable to thefact that parameter values usedin the robot experiments were adjustedto account for the limited space in our lab.

l

13. C O ~ ~ U N I C A T I OAND N C O O ~ I N A T I O NIN ROBOTIC T W S

These predictions are borne outin the simulations shownin the top row of Fig. 13.15. The simulations were runin the environment describedin Section 6 using the baseline control parameters (Table13.1). In the case ofNo Communication,Robot 1 immediately moves to the attractor and begins con sum in^ it (top left). Robot 2 moves away, and continue§ to search for attractors in the ~ f f ~state. ~ eEventually r it too falls within sensor range of the attractor, moves toward it, and helps consume it. In the case of State Communication(topcenter),Robot 1 againinitiallymovestowardthe attractor. Robot 2 begins to follow it (dotted line), then transitions to the f f c ~ ~state i r e (solid line) when it comes within sensor range of the attractor. Finally, in the caseof Goal Communication (top right), both robots immediately move to the attractor and consume it. A. ~ualitative differen~e betwe~n State and Goal Communication is visible in the paths Robot 2 takes to the attractor in Fig. 13.15 (top row). With State Communication, Robot2, initially outside sensor rangeof the attractor,makes a curved path to the attractor because it can only follow Robot1 initially (top center).a hen Goal Communication is allowed, however, Robot2 can proceed directly to the attractor (top right). Now compare the simulations (top row) with runs on the robots Ren and Stimpy (bottom row). Because the sensor range of the robotsis set at 10 feet, the scenario was altered for runs onmobile robots so that the attractor is only 10 feet away from the lower robot. The telemetry is shown at half the scale of the simulated runsto account for the smaller scale of the scenario. ~ualitatively,performance for mobile robots with No Communication is quite similarto simulated performance(Fig. 13.15 bottom left), Initially, Ren does not sense the attractor and explores the left side of the laboratory instead. But eventually, it comes within sensor range and moves to the attractor.When State Communicationis allowed Ren follows Stimpyto the attractor, making a curved path (bottom center). Finally, when Goal Communication is allowe ,Ren travels directly to the attractor (bottom right). The path of the lower robot for the casesof State and Goal Communication is somewhat different in simulation than on mobile robots. On mobile robots, the lower robot curves away from the upper robot much more than in simulation. This isa resultof two factors. First, the scale of the telemetry r~reationsare half that of the simulations. Thus, the effects of interrobot repulsion are visually e ~ a gerated.Second,theperce obstacle detection (a ring of ultrasonic sensors) is not sop to ignore robots: Robots are detected as robots and as sion between them isf~rther exa gerated. This proble better o~nidirectionalsensors and ~erceptualprocesses are incorporat~ into our research.

A W N AND BALCH

The impact of communication on performancein reactive multiagent robotic systems has been investigated throughe~ensive simulation studies. Performance results for three generic tasks illustrate how task and environme canaffectcommunicationpayoffs.Initial results fromtestingonmobile robots are shown to support the simulation studies. Three levels ofcommunicationwereinvestigated:nocommunication, state communicationi and goal communication.hen state communication is allowed, robots are able to determine the interal state of other robots. oal communication is allowed, robots transmit goal~rientedinformation to one another. To evaluate the impactof communicationia baseline of performance was evel loped for robots p e r f o r ~ i n ~~ o r a ~Ceo, ~ s ~and ~ e~, r tasks a without ~ ~ communication, The baseline results were com~ared then with performance in thesetaskswhenstatecommunication,thengoalcommunication is allowed. Im~ortantresults establishedin the baselineexperime~tsinclude: en number of attractors, more robots complete or a given number of robots, it takes lon er to complete a tas in scenarios where largernum ers of attractorsare

13. CO~MUNICATIONAND COO~INATIONIN ROBOTIC TEAMS

will facilitate their use and to establish formally provable properties (Le.,

necessary and sufficient conditions) regard in^ their specifications.

This chapter is dedicated to the memory of Larry Rosenberg who provided support for this research through the National Science ~oundationunder grant ~1~-9100149.

This appendix contains the methods by which each of the individ~al tive schemas used in this research compute their component vector results of all active schemas are summed and norma~ized prior to transmission to the r o ~ ofor t execution. ov~t~~oal: for a goal. v

~

~tot goal t r a cwith t variable gain. Sethi

a~justable ~ d ~

~= i

irection towar~s ~erceived goal ith varia~legain an

A W N AND BALCH

d = Distance of robotto center of obstacle (ldiEctim

*

= along a line from robotto centerof obstacle moving away from

obstacle

Noise:Randomwanderwithvariablegainand persistence. Used to overcome local maxima, minima,cycles, and for exploration. = ~djustablegain value

~ m ~ n i ~ d e

Random direction that persists for ~ adjustable)

~ d i ~ c= t i ~

*

~

~ steps ~ (~ ~ ~ e ~ i ~~ is ~ ce n ec e

Probe: Usedin Graze for favoring continuedmotio~in the current directional heading. V m ~ = i adjustable ~ d ~ gain value or0 if no ungrazed areas detected = Straight ahead along an extrapolated path from the current

location only if grazed area ahead. Direction not important if no ungrazed area ahead as gain is0.

[ l ] Altenburg, K.and Pavicic, M., 1993. Initial Results of the Use of Inter-Robot Communication for a Multiple, Mobile Robotic System,~ o r kNotes i ~ of the ~orkshopon dynamical^ Interacting Robots at I J ~ - 9 3pp. , 95-100. [2] Altmann,S. 1974. Baboons, Space, Time, and Energy.American Zoology, 14221-248. [3] Arkin, R. C., 1989. Motor Schema Based Mobile Robot Navigation, International Journal of Robotics Research,v01 8(4), pp. 92-112. [S] Arkin, R C., 1992. Cooperation without ~ommunication:~ulti-agentSchema Based Robot Vol. 9(3), pp. 351-364, Navigation,Journal ofRobotic Syste~s, [6] Arkin,R. C., 1992. Integrating Behavioral, Perceptual, and World Knowledge In Reactive i~ Agents, ed. P. Maes, Bradford-MITPress, pp. 105-122. Navigation. D e s ~ nAutono~ous 171 Arkin, R C., 1992. Modeling Neural Functionat the Schema Level: Implications and Results al Net~orksin Invertebrate N e u r ~ t h for Robotic Control.~ i o l ~ i cNeural ed. R. Beer, R Ritzmann, andT. McK~nna, Academic Press, pp. 383-410. Robotics [8] Arkin, R. C., 1993. Survivable Robotic Systems: Reactive and Homeostatic Control. and Remote Systems for H ~ a r d Environ~ents, o~ ed. M, Jamshidi and P, Eicker, PrenticeHall, pp. 135-154. [g] Arkin, R C., Balch, TV,Nitz, E., 1993. Communication of BehavioralState in Multi-agent RetrievalTasks, Roc. 1993 IEEE Internation~lConferenceonRoboticsandAutomation, Atlanta, GA, vol. 1, p. 678. [101 Arkin, R. C. and Hobbs, J. D., 1992, Dimensions of Commun~cation and Social Organization in Multi-Agent Robotic Systems,From animals to animals 2: Roc. 2nd Internationul Conference on Simulation of Adapti~e ~ehavior, Honolulu, HI, Dec. 1992, MIT Press, pp. €111 Arkin, R C. and MacKenzie, D.,1994. “Temporal Coordination of Perceptual Algorithms for Mobile Robot Navigation”,to appear in IEEE rans sac ti on^ on Robotics and Automation.

13.COMMUNICATION

AND COO~INATIONIN ROBOTICTEAMS

[121 Arkin, R. C., Murphy, R. R., Pearson, M., and Vaughn, D,, 1989. Mobile Robot Docking Operations in a Manufacturing Environment: Progressin Visual Perceptual Strategies,Proc.IEEE International ~orkshopon Intell~entRobots and Systems '89,Tsukuba, Japan, pp. 147-154. [131 Arkin, R C., et al., 1993. Buzz:An Instantiation ofa Schema-Based Reactive Robotic System, €'roc. Internationul Confe~nce on ~ntell~ent Autonomous Systems: I A S , P~ttsburgh, PA., pp. 418-427, [141 Balch, T. and Arkin, R. C., 1993. Avoiding the Past:A Simple but Effective S tiveNavigation, h x . 1993 IEEE Inte~ationalConference onRobotics Atlanta, G A , vol. 1, pp. 678-685. [151 Brooks, R. A., 1986. A Robust Layered Control System for a Mobile Robot, IEEE Journal of Robotics and Automa~on,RA-2, No. 1, p. 14, 1986. [161 Brooks, R., Maes, P., Mataric, M,, and More, G., 1990. Lunar Base Construction Robots,IEEE International ~orkshopon Intell~entRobots and Systems (IROS 'go), Tsuchiura, Japan, pp. 389-392. [l71 Clark, R. J., Arkin, R C., andRam, A., 1992.LearningMomentum:On-linePerformance Enhancement for Reactive Systems. Proc. I992 IEEE International Conference onRoboti~and Automation, Nice, France,May 1992, pp. 111-116, [l81 Croy, M. and Hughes, R., 1991. Effects of Food Supply, Hunger, Danger, andCompet~tionon choice of Foraging Location by the fifteen-spined stickleback. Animul Behauior,1991, Vol. 42, pp.131-139. [191 Dudek, G., Jenkin, M., Milios, E., and Wilkes, D., 1993. A Taxonomy for Swarm Robots.&oc. I993 I E E E / International ~ Conferenceon Intell~entRobots and Systems (IROS).Yokohama, Japan, pp. 441-447. 1201 Drogoul, A. and Ferber,J., 5992. From Tom Thumb to theDockers: Some Experiments with Foraging Robots. F m Animals to Animals:Proc. 2nd International Conference on the Simulation of A~aptiue Behauior, MIT Press/Bradford Books, Honolulu, HI, pp. 451-459. [21] Foreano, D., 1993. Emergence of Nest-based Foraging Strategies in Ecosystems of Neural on the Simulation of Networks. From Animals to Animals:h.2nd Inte~ational Confe~nce Adaptiue Behauior,MIT Press/Bradford Books, Honolulu, HI, pp. 410-416. [22] Franklin, R. F. and Harmon,L. A., 1987. Elements of Cooperative Behavior. Internal Research and Development Final Report 655404-1-F, En~ronmentalResearch Institute of Michigan (EWM), Ann Arbor, MI. 1231 Franks, N., 1986. Teams in Social Insects: Group Retrieval of pre by army ants. Behau. Ecd. Sociobid., 18:425-429. [24] Fukuda, T-,Nakagawa, S., Kawauchi, Y., and Buss, M., 1989, Structure Decisionin Self Organising Robots Based on Cell Structures~EBOT.IEEE International Conference on Robotics and Automation,Scottsdale, Arizona,pp. 695-700. [25] Goetsch, W., 1957.77ze Ants. University of Michigan Press, 1261 Goss, S., Beckers, R., Deneubourg, J., Aron, S., and Pasteels, J., 1990.How Trail Laying and Trail ~ o l l o ~ can n g Solve Foraging Problems for Ant Colonies. Behauio~l~ e c ~ n i s m of s Selec~on,ed. R. N. Hughes, Nato AS1 Series, Vol. G. 20, Springer-Verlag, Berlin, pp. 661-678. [27] Hackwood, S. and Beni, S., 1992. Self~rganizationof Sensors for Swarm Intelligence, I992 IEEE Inte~ationalConferenceon Robotics and Automation,Nice, pp. 81!%-829. [28] Helmbold, D. and McDowell, C., 1990. Modelling SpeedupCn) Greater than n, IEEE ~ r a n s ~ c v01 1(2), pp. 250-256. tions on Parallel and~ i s ~ i b u t Systems, ed [29] Holldobler, B. and Wilson, E., 1990. The Ants, Belknap Press, Cambridge Mass. [30] Kaelbling, L. and Rosechein,S,, 1990., "Action and Planning in Embedded Agents",in Designi~ Autonomous Agents,Maes, P. (ed), MIT Press, pp. 35-48. [31] Kaufmann, J., 1974. Social Ethology of the~ i p t a iWallaby, l Macropus Parryi, in N o ~ h ~ a s t ern New South Wales.Animal Behavior,22:281-369. €'roc. 1321 Khatib, O., 1985. "Real-time Obstacle Avoidance for Manipulators and Mobile Robots", IEEE Int. Conf Robotics and Automation, St. Louis, p. 500.

ARKIN AND BALCH 1331IQ-ogh,B., 1984. A Generalized Potential Field Approach to Obstacle Avoidance Control,

SME-RI Technical Paper MS84-484.

[34] Lee, J.,Huber, M,, Durfee, E., and Kenny, P,, 1994. UM-PRS: An Implementation of the Proce-

~ / ~ A Conference S A on dural Reasoning System for Multirobot Applications. To appear Intell~entRobots in Field, Factow Service, and Space i ~ R W). F ~

(351 MaKenzie, D. and Arkin, R.C., 1993. Formal Specification for Behavior-Based Mobile Robots, Roc. SPIE Conferenceon Mobile Robots W., Boston, M A , pp. 94-104. [X] MacKenzie,D. C. and Balch, T. R., 1993. Making a Clean Sweep: Behavior Based Vacuuming. W o r k i Notes ~ of 1993AAAI Fall Symp~ium: Instantiati~ Real-World Agents. Ralei [37j MacLennan, B., 1991. Synthetic Ethology: An Approach to the Study of Communication.In Artj~cialLifeII, SFI Studies in the Sciencesof Complexity, vol.XI, ed. Farmer et al.,AddisonWesley. [B] Maes, P,, 1991. Situated Agents can have Goals, in D e s ~ nAutonomous i~ Agents, Maes, P. (ed), MIT Press, pp. 49-70. [39] Mataric, M., 1992. Minimizing Complexity in Controlling a Mobile Robot Population. Roc, IEEE Inte~ationalConference on Robotics and Automation, Nice, FR, May 1992. [40] Miller, D., 1990. Multiple Behavior~ontrolled Micro-Robots for Planetary Surface Missions. h . 1990 IEEEInterna~onalConferenceon Systems, Man, and~bernetics,Los Angeles, CA, November 1990, pp. 289-292. [41] Moynihan, M., 1970. Control, Suppression, Decay, Disappearance and Replacement of Disal 2985-112. plays. Journal of ~ e o ~ t i cBiology, [42] Noreils, F. R., 1992. "Battlefield Strategies and Coordination between Mobile Robots",Roc. 1992 IEEE/W InternationalConferenceon Intell~entRobotsandSystemsiIROS), pp. 1777-1784. [43] Noreils, F., 1993. Coordinated Protocols: An Approach to Formalize Coordination between Mobile Robots. hoc. 1992 I E E E / ~Interna~onalConferenceon Intell~entRobots and Systems iIROS),pp. 717-725. ] Parker, L., 1993. Adaptive Action Selection for Cooperative Agent Teams.From Animals to . 2nd Inte~ationafConference on the Simulation of Adaptive Behavior, MIT Animals: h PressBradford Books,Honolulu, HI, pp. 442-450. in Des~ning [45] Payton, D. W., 1991. Internalized Plans:A Representation for Action Resources, Autonomous Agents,Maes, P. (ed), MIT Press, pp. 89-103. [46] Pearce, M,, Arkin,R. C., and Ram, A., 1992, The Learning of Reactive Control Parameters

through Genetic ~gorithsm,Pnx. 1992 Internationai Conference on Intell~entR ~ o t ~ and cs Systems iIROS~,Raleigh, N.C., pp. 130-137.

[47] Ram, A., Arkin, R. C., Moorman, K., and Clark, R., 1992. Case-based Reactive Navigation:A

case-based method for on-line selection and adaptation of reactive control parameters in autonomous robotic systems, Technical ReportGITXC-92/57, College of Computing, Georgia Tech. [48] Slack, M. G., 1990. Situationally Driven Local Navigation for Mobile Robots, JPL ~ b l i c a ~ o n 90-17, Jet Propulsion Laboratory, Pasadena,CA. [49] Tinbergen, 1966.Social Behavior inAnimals, Methuen & Co., London. [ 5 0 ] Wang, J., 1992. "Distributed Mutual Exclusion based on Dynamic Costs",Roc. IEEE International ~ y ~ p ~ on i uIntell~ent m Control, Glasgow, UK, pp. 109-15. [51] Werner, G. and Dyer, M. 1990. Evolution of Communication in Artificial Organisms. Technical Report~ C ~ - ~ - 9 AI 0 -Laboratory, 0~, Universityof California, Los Angeles. [52] Yanco, H. and Stein, L., 1993. An Adaptive Communication Protocol for Cooperatin Robots. From Animals to Animals:Roc. 2nd International Conference on the S i m ~ l a ~ oofn Aduptive ~ehavior,MIT Press~radfordBooks, Honolulu, HI, pp. 478-485.

C H A P T E R

e e

erhard Fischer

University of Colorado, Boulder Microsoft Research

ay~ond ~c~all Jonathan Ostwald

University of Colorado, Boulder University of California, lwine TwinBear Research, Boulder,CO

rank Shipman

Texas A&M University

For a number of years we created software-based design environments solely to support individual designers. Recently, however, we turned our attention to the problemof supporting long-term collaboration. This takes place when an artifact functions and is repeatedly redesigned over a re~ati~ely long period of time (e.g., many years). Such artifacts are increasingly com-

FISCHER ET AL.

mon in a wide range of domains, including the design of buildings, spacebased habitats, software, and computer networks-to name a few that we have lookedat, Our initial plan was to modify our previous architecture of design environments to add more support for the evolutionary development of the design and the knowledge about the design. Because we anticipated this would create "messy" information we proposed a processof seeding, evolutionary growth, and resee of the informatiQnin the design environment, In thecourse ofworkingnetworkdesigners, building andevaluating prototype systems, and revising our initial theoriesof collaborative design practice we developed some new outlooks that we think are quite useful. This chapter discusses these results. This chapter beginsby describing our viewof design and long-termindirect collaboration.We then discuss the role that knowledge plays in collaborative design. Next the three phases of seeding, evolutionary growth, and reseeding are each describedin detail. A discussion section describes experiences from observing and working with network designers experiences affected our system prototypes. After reviewin will be of value to othwe conclude with the lessons we learned and believe ers interestedin supporting long-term asynchrQnous design.

Teamwork is playing a larger role in design projects ~ ~ e M ~Lister r c o , 87; Hackman, 74; Johansen 881. Such projects are increasingly large, complex, g in duration. Thedesignprocesstakesplace over many by extendedperiods of maintenanceand years,onlytobefollowed ts from many different domains must coordinate their e separationsof distance and time.In such projects, contion is crucial for success yet difficult to achieve. difThis ficults is due in large part to ignoranceby individual designers of how the ~ecisionsthey make interact with decisions made by other designers. A large part of this, in turn, consists of simply not knowing what has been ngsand other types of direct communication are the commonly used means for coordination and collaboration in design projects, but in many situations~speciallyones involving lon~-termcollaboration-these are not feasible. Design projects that extendover many years can involve a high turnover in personnel. Much of the design work on systems is done as maintenance and redesign, and the people doing this work are often not members of the original design team. But to be able to do this work well,or sometimes atall, requiresc o l l ~ ~ o rwith ~ ~ ithe o ~ original designers of the sys-

FISCHER ET AL.

But after awhile, people beginto leave messages that contain more than just, “call me back.” For example: Hi, this is Denny, I’ve found a problem with theAR 30 report. The due datesare

incorrectforcustomernumber day’s close.

.Weneed togettheserightbeforeFri-

A~though it takes time, people learn that it is more efficient to place an item on another person’s “electronic stack” or “inbox” than to interrupt what they were doing with a phone call. And although people usually prefer to speak to someonein person, they nevertheless learn how to make good use of the technology. Leaving a detailed message makes moresense for many of the messages and requests. It can be better to give the person time to research the questionand call back, rather than surprise him with a problem and expect instant diagnosis, As~chronous communication beginsto be used wheres ~ c ~ r o npreo~s viously dominated. Although phone mail is at first thought of as an inconvenient backup to the more preferable voice-to-voice, it becomes a useful service in its own right, and more than just a back-up. In arguing against the assumption that the goal of computational media shouldbetoemulateface-to-facecommunication,HollanandStornetta [l9921 cited email as the ”paramount success of computationally-mediated informal communication.’’ The interesting aspect related to the point argued here is the statement:

It meets our critical litmus test of being used by groups even when in close physical proximity.In fact, in our own experience, it is not uncommonto send email an office. to someonein the next adjacent office, or even someone sharing Like voice mail, in certain situations, email has changed from a tolerable substitute to a preferred medium. Electronic mail plays an important role as a successmodel of computer ~ ~ focuses on prosupported communication. Although much C § research media that emulate face-to-face meetings, the success of email is a reminder that there is more to good communication support thanemu la tin^ face-to-face communication.A benefit froma synchronous communication is that it is archived somewhere. Whether this is digital voice recording, or electronic mail, it is available for later retrieval. Clearly this does not solve the problem of information retrieval later, but it is a first step. The two examples, voice mail and electronic mail, illustrate that asynchronous communication comes to be preferred tos~chronouscommunicationforcertaintypesofinformation.But in this chapter wemake the stronger case that not only is as~chronouscommunication an improvement in some cases, itisabsolutelynecessaryforlong-termdesign.To

14.SEEDING,

EVOLUTION~YGROWTH, AND RESEEDING

tern, People who are not in the project groupat the same time need to collaborate in long-term design, Much research in supporting collaborative work has gone toward supporting synchronous communication (e.g. COLAB [Stefik et al. 871 and GROW

[Ellis, Gibbs, Rein 911). Earlier, we argued that lon~termcollaborative design demands support beyond synchronous communication. Even when it is posto have direct communication, there is still sible for collaborating designers much potential in providing tools primarily intended foras~chronous use [Hollan, Stornetta921. The p~imarydistinctions between synchronous and asynchronous communication are taken from the Computer Supported Cooperative Work field ( ~ S C and ~ , are represented by the matrix in Fig. 14.1. The following two examples illustrate the benefits of technologically supported asynchronous communication: voice mail and electronic mail. Consider the installation of voicemailsystems in large corporations. Experience suggests that once people become used to the of idea asynchronous communicationby phone, they make better use their phone communication. At first people leave messages like: Hi, this is Denny, I guess you’re out,so call me back.

Time me

~ifferent

me

ifferent FIG.14.1. Asynchronouscommunication. A matrix of CSCW perspectives developed by Johansen [Johansen 881. Our work focuses on technolo~ical support for Asynchronous (i.e., different time, different place) communication. Email is the prototypical exampleof asynchronous communication.

14.SEEDING, EVOLUTION~YGROWTH,ANDRESEEDING

model design communication onlyafter email or voicemailis to underutilize an important resource for design teams. Careful analysis led us to introduce a key distinctionin asynchronous communication, that of predictabi~ity. e

In long-term collaborative design tasks, communication between designers is not only as~chronouswith respect to time and place, but it is alsoindirect in the sense that the senders and receivers of information are not known a priori. Figure 14.2 introduces predictability into the w e ~ ~ - ~2oxw2 n matrix of C ~ (compare C ~ Fig. 14.1). Also note that predictabilityp the participants as well. Long-term projects are unpredictable with As we argue later, the team members and users who need to communicate. this attribute caused us to pursue what we describe as embedded communication, where the communication isin a sense “embedded”in the design artifact rather than being stored separately. ~ o n ~ t ei ~r ~~i,r e c t c o ~ ~ u nis j cof ~ tparticular jon importancein situations where:

direct communication is impossible, impracticalor undesirable communication is shared around artifacts * designed artifacts continue to evolve over long periods of time (e. over months or years) * designers need to be informed within the context of their work

* 0

Time Different Different Predictable Unpredictable Same

ace Same ifferent Predictable Different ~npredictable

FIG. 14.2. A classification of Different CSCW perspectives. This classification scheme extends thematrix shown in Fig. 14.1. Our focus is on the unpredictable communication that occurs throughout the design lifecycle of complex systems. Not only is the time and place unpredictable: the participants themselves are not alwaysknown over thelong lifecycles of complex systems[Grudin Ma].

FISCHER ET AL.

Support for indirect coordination and collaboration must go beyond what electronic mailandmostproposed CSCW software could provide. This support should allow team members to work separately-across substantial distances in space and time-but alert them to the existenceof potential interactions between their work and the workof others. Where such interactions exist, support should be provided for collaboration and conflict resolution. Designers must be able to interact with design artifacts created by previous designers. Technology enabling this could effectively create virtual cooperation between all designers who ever workedon the project. These challenges motivated not only a new conceptual model of design en~ronments, but also a new model of the design process.

Our conceptual framework grew from systems augmenting an individual expert solving a problem to support for cooperating experts collaborating over many years and different places. In the process of our initial research, we formulated a design process model, which we now believe is an important aspect of designing systems for collaboration. The process model is motivated by how large software systems, such as GNU Emacs, Symbolics’ Genera, Unix, and the X Window System, have evolved over time,In such systems, users develop new techniques and extend the functionality of the system to solve problems that were not anticipated by the system’s authors.New releases of the system often incorporate ideas and code produced by users. In the same way that these software systems are extensible by programmers who use them, design environments need to be extended by ~ o ~ ~ e s ~ ~(our e r term s for the users of design environments) whoare neither interested nor trained in the (low-level) details of computational environments [Nardi 931. We illustrate our conceptual framework in the domain of computer network design, which involves complex artifacts thatare continuously modified and redesigned, The domain itself is also constantly changing as new technologies are developed. Knowledge acquisition isa crucial issuein the creationof effective information systemsof all types (including expert systems, hypermedia systems, and design environments). There have been two extreme approaches: one is to inputinformation in advance of use,typified by expertsystems [Buchanan, Shortliffe841, and the other is start to with an empty system and allow its information base to grow and become structured a consequence as of use, characterizedby initial proposals for argumentative hypertext [Conk-

14. SEEDING, E V O L ~ I O ~ ~GROWTH, RY AND RESEEDING

lin, Elegeman 88; McCall, Schaab, Schuler831. Neither approach is adequate

for the information needsof designers. approach fails for numerThe ‘~ut-al~-the-knowIedg~in-at-th~beginning’’ ous reasons. It is inadequate for domains in whtch the domain knowle undergoes rapid changes (the computer network domain being a prime example). Tr~ditiona~ knowledgeacquisitionapproachesthatrequire domain designers to articulate their knowledge outside the context of problem solving or during an initial knowledge acquisitiQn phase fail to capture tacit knowledge [Polanyi661, because designers know more than they can tell environment developers. Tacit knowledge is a part of human expertise that surfaces onlyin the contextof solving specific problems. approachrequirestoomuch The “just-provid~an~mpty-framewor~’ work of designers in the contextof a specific project. The difficultiesof ca turing design knowledge from design projects are well known [Fischer et 911. Documentinginterfereswiththethinkingprocessitself, ~isrupting design and requiring substantial time and effort that designers would rather invest in design.Designerstypically find itdifficulttostructuretheir thoughts in a given format, regardless of the format used [McCall91]. In addition, domain designers often lack thenow ledge and the interest to formal931. ize knowledgeso it can be computationally interpreted [Shipman model is between the two extremes of “put-all-th~knowledge-in-atinning” and “just-pro~id~an~mpty-framewor~.” Designers are more ed in their design task at hand than in maintaining the knQwledge base.Atthesametime,importantknowledgeisproducedduring design activities that should be captured. Rather than expectdesi~nersto spend extra time and effort to maintain the knowledge base as they design, we provide tools to help designers record information quickly and without regard for how the information should be integrated. In our model, knowledge base maintenanceis periodically performedby environment develo ers and domain designersin a collaborative activity. Ourdomain-independentdesignenvironmentarchitectureplaysan important rolein the continual developmentof design environments. It provides a structure for domainknowledgeandmechanismsfor nowl ledge as itisneeded to support the design task at hand developedourdomain-independentarchitecturethroughnumerous attempts to createdomain~rienteddesign environments [Fischer 921. The architecture consistsof the following five components: (a) a cQnstruction component, (b) an argumentation component, (c) a catalog of interestin design examples,(d) a specification component, and (e)a simulation component. The individual com~onentsare linked by knowledge-based mechanisms: a construction analyzer (built as a critiquing system [Fischer et al. 93]), an argumentation illustrator, and a catalog explorer ~Nakakoji 931. Design environments contain information encoded usinga varietyof repre-

FISCHER ET AL.

sentational formalities. Construction kits and criticsare considered formal representations of design knowledge because they are interpreted by the computer. Argumentation isa semiformal representationin which informal ual and graphic records are linked by formal associations. Our process model for continual development of design environments from an initial seed through iterations of growth and reseeding is illustrated in Fig. 14.3.

en~ronment Theseedingprocess, in whichdomaindesignersand developers work together to instantiatea domain~rienteddesign environment seeded with domain knowledge. The evolutionary growth process,in which domain designers add information to the seedas they use it to create design artifacts. The reseeding process,in which en~ronmentdevelopers help domain designers to reorganize and reformulate information so it can be reused to support future design tasks. To illustrate the evolution of design environments, we discuss seeding, eve lution through use, and reseeding in detail in the following three sections. € n ~ r o n ~ nfbrnain t

En~ronment Domain

c

FIG. 14.3. Seeding, evolutionary growth, and reseeding: process a model for domain~rienteddesign environments. During seeding, environment develop ers and domain designers collaborate to create a design environmentseed. During evolutionary growth, domain designers create artifacts that add new domain knowledge to the seed. In the reseeding phase, environment developers again collaborate with domain designers to organize, formalize, and generalize new knowledge.

14. SEEDING, EVOLUTION~YGROWTH, AND RESEEDING

~ E T W [Fischer O ~ et al.92; Shipman 931, a design environment supporting computer network design, is used for illustration.

A seed

is builtby customizing the domain-independent design environment architecture to a particular domain through a process of know~edge construction, ~ l t h o u ~ theh goal is to constructas much knowledge as possible during seedbuilding, for complex and changing domains complete coverage is not possible. Therefore, the seed is explicitly designedto capture design knowledge during use[Girgensohn9 4 . Domain designers must participate in the seeding process because they have the expertise to determine when a seed can support their work practice. Rather than expecting designers to articulate precise and complete system requirements prior to seed building,we view seedbuilding as knowledge construction (in which ~ o w l e dstructures ~e and access methods are collaboratively designed and built) rather than as knowledge acquisition (in which knowledge is transferred from an expert to a knowledge engineer and finally are elicitexpressed in formal rules and procedures).New seed requi~ements ed by constructing and evaluatingdomain~riented knowledge structures. The seeding process for the~ E design ~environment O (see Fig. ~ 14.4) was driven by observations of network design sessions, prototypes of proposed system functionality, and discussions centered on the prototypes. In design sessions,a logical map of the network being designed served toground design meetings, discussions, what-if scenarios, and disagreements. The logical map was chosen as the central representation of the artifact in network design, anda prototype construction kit was implemented based on the logical map [Fischer et al. 921. Evaluation of the ~ E seed indicated ~ that O ~ in theform of critiques, designersneedsupportforcommunication reminders, and general comments [Reeves, Shipman 92al. Pointer, annotation, and sketching tools were integrated into the construction so kittalking about the artifact takes place within the artifact representation space. An important lesson we learned during the seedingof ~ E T ~ O base our design discussions and prototyping efforts on existing artifacts. Discussing the existing computer science network at the~niversityof Colorado, Boulder, was an effective way to elicit domain knowledge because it provided a concrete context that triggered domain designers' knowledge (often in the form of "war stories").We found high-level discussions of general domain concepts to be much less effective than discussions focused on existing domain artifacts. Information to seed ~ E T W was O ~acquiredfromexistingdatabases containing information about network devices, users, and the architectura~

FISCHERET AL.

FIG. 14.4.An

environment supporting computer network design. A screen image of the N E T W O ~seed. Shown is a palette of network objects (upper ht)and the constructionareawherelogicalnetworksareconfigured (upper left).

layout of our building. The N E T W O ~ seed contains formal representations of approximately 300 network devices and60 users. Autocad databases created by facilitiesmaintenancepersonnelprovidearchitecturaldetailsof about 100 rooms. This information is represented in NETWO tionkitand in the underl~ngknowledgerepresentationformalisms.The informal partof the N E T W O seed ~ includes notes from the systems administration class,knowledge aboutthe various research groups, and electronic mail of the network designers.

During the use phase, each design task has the potential to the addknowlare edgecontained in thesystem. New constructionkitpartsandrules re~uiredto support design in rapidlychangingdomains [Fischer, Cirgen-

14. SEEDING,EVOLUTIONARY G R O ~ HAND , RESEEDING

sohn 901. Issue-based information in the seed can also be augmented by each design task as alternative approaches to problems are discovered and recorded. The information accumulatedin the information space during this phase is mostly informal because designers either cannot formalize new knowledge or they do not want to be distracted from their design task. Our approachto this challenge is to view the design environment seed as a mediumforcommunication aswellasdesign.Ourcritiqueofcurrent n systems is that they function as “keepersof the artifact,”in which one depositsrepresentations of theartifact b designed.Butourexperience has shown that designers integrate desig and discussing in such a way as to make separate interpretation difficult [Reeves 931. Talking about an artifact requires talking with the artifact. Therefore, later interpretation of the discussion requires that the discussion be embeddedin the context in which it was originally elicited. The most important implication of this view is that design artifacts must not be artificially separated from the communication about them. This integration was seen in two ways. First, in design sessions videotapedforanalysis,deicticreferences(referringtoitems by theuse of “these,” “those,” “here,’, and so forth) were frequent. A long-term study of network designers showed that users took advantage of embedded annotations and made frequent useof deictic references [Reeves931, Second, discussion about the artifact guided the incremental design process. Designers took every opportunity to illustrate critiques of each other’s work. Only rarely was a detailed comment made and not accompanied by changing the artifact. The logical map mentioned earlier servednot only to represent the real network, but alsoas amedium through which changes were considered and argued (Fig. 14.5). It focused as well as facilitated discussion. Frequently,in arguing over desi n artifacts, specific issues led to discussions of larger issues, Collaborating designers preferred to ground discussions in design representations. Thelogicalmaps servedto pointoutinconsistencies between an appealing idea and its difficulty of implementation; remind participants of important constraints; and describe networkstates beforeand after changes.

n knowledge is of little benefit unless it can be delivered to designers when it is relevant. Periodically, the growing information space mustbe structured, generalized, and formalized in a reseeding process, which increases the computational support the system is able to provide to designers [Shipman, McCall94].

FISCHERET AL.

FIG,14.5. Logical map with embedded discussion. The logical map serves to abstract away low-level details while allowing discussions about the artifact to beembedded inthe design.

The task of reseeding involves environment developers working with domain designers. Aftera period of use, the information space can abe jumble of annotations, partial designs, and discussions mised in with the originalseedandanymodificationsperformed by the domain designers. To make this information useful, the environment developers work with the domain designers in a process of organizing, generalizing, and formalizing the new information and updating the initial seed. ~or~aRiza~ioR* The org~izationalaspect of reseeding

is the information space, through modificationby designers using the system, eventually becomes inconsistent. For esample, as new techno lo^ becomesavailable,informalnotesmightcontradictformal e about network devices. Whenaninformationspaceisdisorganized,itbecomesdifficultfor designers to locate information that they need. This also makes it more difficult to find the 'tight" place to add new information, thereby compounding the problem. Disorganization can occur when information about the same topic has become located in separate parts of the information space, or find it. when information has been putin a location where designers cannot R. Reuseofinformationbetweenprojects re~uiresthe generalizationof task-specific information entered during use. The goal is to create more generally applicable informationby integrating

14. SEEDING, E V O L U ~ O N ~GROWTH, Y AND RESEEDING

information about specific situations. This is related to the need for reorganization when variations of the same ideas have been addedin project~riented parts of the information space. An example of generalization in the network domain is that while documenting changes to a design, information concerning the conversion of sections of a network toa new networking standardwill likely appear with each conversion of a subnet. In order to bring this information together into a coherent whole, the subnet-specific details need to be abstracted so that the information isof use in new situations. As described earlier, systems supare forced to evolve or become obsolete. Although some of this evolution can occur as evolutionary growth, still there istheneedforconcertedefforts by environmentdesigners to update domain information. As an example of the difficulty and potential for support in updating the information consider the problem of updating the initial formal structures created during our seeding of ~ E T W QThere ~ . is frequent change in the number and type of devices, people, and places that the en~ronmentneeds to represent. One potential for supporting this process comes from the existence of online sources for someof this information. The sourcesof online informationindividuallycontainonlypart of theinformationneeded.Also,the sources use a varietyof representations and identifiers for the information. As a result, the process ofupdatingturnsout to be heuristic. Still, the automation of changes aids the knowledge engineers in their updatingof the informationspace,requiringthemtodetermineandhandleexceptional cases and to disambiguate between multiple potential cross-references. Q ~shown the needto devise semiautoThis experience with~ E T W has matic methods for updating information from other online sources during reseeding. Although much informationin computer network desi online, such as user and device profiles, other domains are also likely to havepotentialonlinesourccis of information.Supportbasedonthese sources can aid both domain designers and environment designersin keeping the design environmentup to date. ation.

r partn.is that Atask final of reseeding is the formalizationof domain information entered during evolutionary growth but not in a format useful for providing knowledg~based support. Ekperience has shown thatin many cases users cannotor will not 94). formalize information on their own [Shipman, Marshall An example of this is that designers may make comments in notes to themselves or one another about the characteristics of a network device

FTSCHER ET AL.

and yet notenter thatinformation into the environment's description of the device, Designers are notbeing reticent, but cautious in takingtimeperforming tasks whichare not necessary tocompleting their primary task, the design and instantiation of the network. During reseeding, both because the primary task is the improvement of the information in the design environment andthe involvement of the environment designers, theformalization of information entered informally durevolutionary growth becomes possible.

Experience with the incremental deve~o~ment of the col~aborative desi en~ronment~ E T has ~ shown O ~the utility (and perhaps necessity) using a seeding, evolutionary growth, reseeding process. At the same time it inted out the need for further types of support to be provided by environments for sucha processto be successful. First, the design environment needs to become more than just the storage mechanism for a design, and must start becoming a medium for communication between designers. This leads to the need for embedding communication with the design artifact, Once an artifact and discussion of the artifact develop over some time it becomes necessary to provide prior context to enable the comprehension of comments and design decisions. Information must be allowedto enter the environment in an informal representation but formal representations are needed to provide knowledg~ based support. Thus, computer support for formalization of knowledge is important for successful design environment development. Because of the size of the information space and the potential for missing relevant information entered by other designers, mechanisms are requiredto help the location and communication of information. This has led to an initial investigation of computational agentswhich convey designdecisionsand opinionsto other designers,

Analysis of design sessions in several domains (kitchen, architecture, networks) showed that design artifacts ground discussions so strongly that the e itselfis difficult to understand withoutaccess to video. Deictic ref(this,overthere,here,that)arefrequent,yetcomputer-based design systems suchas CAD, do not support this referencing, This led to us embedded communication: including the discussion about a design artifact in the artifact itself.Though intended to address deixis, this also addressed we see in design rationale systems, namely that the artifact

14. SEEDING,EVOLUTIONARY GROWTH, AND ~ E E D I N G

and its rationaleare stored separately. This quickly leads to inconsistencies between what was done, and what was supposed to have been or what done, user think should have gotten done. Competent practitioners usually know more than they can say [Polanyi 661, and conversation leaves many things tacit. One could attemptto overcome these tacit aspects by forcingdesignerstomakemoreknowle explicit, Against this approach, which might be labeled the~ t ~ a nofn y explicit” [Hill 891, issues of designer attention and desireare seen as motivation for providing computational media in which the designer’s natural level of tacit knowledge is respected. The design process suggests the need for a medium in which the design artifact emerges, and which allows the designer to undergo frequent “shifts in stance’’ [Schoen8 Observations of collaborating designers usingNE facts serve as medium for communication. Further theartifactguidetheincrementaldesignprocess.Whencommunicatin asynchronously via textual annotations, network designers integrated the notes and the artifactin ways that made separate interpretation difficult.

.

ist

In the same way that evidence of physical history guides cognitive tasks, computational media should provide cues of use which [Hill et al. 921. For example, as auto parts manuals beco vide visual and tactile cues to guide further use. In the S ical wear and tear can be a resource, computational media should embed the historyof an artifact in the artifactso that it canserve as guide to further use, Computationa~ media represent the potential to provide history access mechanisms that go beyond what is possible with physical artifacts, as such providing accessby various perspectives such as date, user, design chan and relation to other design units. The approach here was to recognize thefluid nature of the des ess and create a computer-based en~ronmentin which the artifact an end-product and moreof a process. If capturing the design process can be done in a way that does not interfere, then otherswill be bette learn from design ob sequence later. Understanding a designisbest dying theprocessaswellastheproduct Ullman 911. Users draw heavily from past experience in solving current ~roblems [Lee 921. Computational tools should therefore support this human tendency to reuse previous experience. However, a complicating factor is the tendency for people to isre remember" an event according to plausible inferenceratherthanexactrecall[Reder 821. Thereismuchpotentialfor in restoring the context computer systems toserve as external memory aids

FISCHER ET AL.

past design decisions [Anderson 85; Reder 82; Suchman 871. The context becomes all the more importantas collaboration increases,In the contextof collaborative design, it is not enough to provide user history, there should be artifact history. Wolf and Rhyne [19921 argued that the process by which information is created and used can be important for understanding of the end productof a work group. In a study done to gain insight into how to facilitate information retrieval in computer-mediateddesignsessions,theyanalyzedhow groupparticipantsusedvideotapetoaccessmeetinginformation.They found that people searched for information using four main access methods by participant: they remembered person X doing some action * by communication medium: people recalled what medium was used *

(e.g., whiteboard, overhead transparency) by time: people used relative time (”midway through the meeting”), duration (“25 minutes into the discussion”), and clock time (“we only got throu~hItem 1.2 by 5 o’clock”) by relation to other events: people used eventsas markers before or after other events.

Thesefindingsofhowpeopleusevideotapeforinformation retrieval serve as challenges for computational history mechanisms. utchins’ [l9901 study of team navigation of large ships also motivates history for collaborative artifacts: The work a chart does is performed on its surface-all at the device interface, as it were-but watching someone work with a chart is much more revealing of what is done to perform the task than watching someone work awith calcula[p. 2171 tor or a computer.

As~chronouslycommunicating designers do not have the possibility of ‘~atchingsomeone work witha chart.’’ However,by keeping the artifact history, the interaction is available for watching ata later time. To relate it to Hutchin’s study, imaginea chart which could replay the interaction that took place and show the instruments as they were used. Design history also provides an approach to design rationale. Though design rationale appears to have great promise [Kunz, Rittel701, there have been few recorded successes [Yakemovic, Conklin901. The designers must perceive a benefit for the extra cost of documenting their reasoning [Reeves, hipm man 92bl. History is therefore a potential candidate for an interaction CO nitive cost associated with having history tool, because there is no extra support. Yet itprovides the benefit of restoring the context of previous work, others’ as well as one’s own. *

14. SEEDING,EVOLUTIONARYGROWTH, AND ~ E E D I N G

The benefit of history relatedto design rationale is thatin domains such as network design, which involve tw~imensionalsketches and graphical representations, designers can often deduce rationale by seeing the process of how something came to be [Chen, Dietterich, Ullman 91; ~uffner ~llman 911. A logical map of the current network hides many tradeoffs and compromises that were made in the past, yet which still affect current decisions. Having the historyof the evolution clarifies some of the tacitknowle~gethat is represented in the static logical map. One side benefit of some groupware aids is that they also help theindividual. For example, one designer said, “What did I do last?” Though the history was primarily viewedas a tool to help one understand other of one’s own work, called “reflexwork, itis also useful for reminding oneself ive CSCW” [T~imbleby, Anderson, Witten 901. Usually adding multiuser features complicates the system for single users, but history is an example of both kinds of use. Research in human memor~has sh “what y inferring is p~ausib~e given what they can remember” [Anderson 851. Memory performance improves the more closely the current context matches the past physical, emotional, and internal context.Much of recall invokes plausible inference rather than exact recall[Reder 821. History toolsare needed to support collaborative design. The motivation in the work donein situated cognition relatingto confor this argument lies text [Carraher, Carraher, Sch~iemann sf;, Lave 883. Design environments can ‘ capture only a portion of the whole context, namely the dates when a user made certain changes. Yet this small portion can be important i laborative design. Each designer on a project team understands onlya portion of the overall design artifact.As large projects evolveover time and turnove tion take their toll, will it become increasingly important for comp design environmentsto help capture the evolution of an artifact a its current state.The history serves to remind designersof how came to be and what the context was when certain decisions were made.

Study of colla~orati esignprojectsshowedthat desi~nersr from informalto for f ple, only allow the ‘ theimportance of “in part of the §ystem, because they exertenormo~§influe eve~opment and implementation proces

FISCHER ET AL.

distinction between formal and informal data, we support a smooth migration from informal representations such as textual notes and sketches, to more formal representation such as 00 class diagrams, or standardized graphical items such as palettes. To address the difficulties involved in formalizing information by both en~ronmentdevelopers and domain designers during all phasesof design en~ronmentdevelopment, we have developed tools that support the process of formalization [Shipman931. One class of tools suggests possible formalizations that could betoadded the information space based on the varietyof information that the system already has available from both the seed and information added during use. The formally represented information, along with the placement, textual by these tools. content, and textual attribute values, can be used in ~ E T looks ~ for O vocabulary ~ in textual values For example, one tool of attributes thatmight relate to other objects and suggests replacement (or augmentation) by a relationship between the two objects. h example of this would be a workstation (c3d2)in the design that has an attribute “disk server’’ with the value “c3dl” as a text string, This tool would suggest the recasting of this attributeto be a relation, instead of a string, pointin in the design that represents the device c3dl. Tools can also make use of possible references found in the textual display of objects to suggest new attributes and relations. As an example we discuss the text annotationin Fig. 14.5, which was taken verbatim from an electronic mail message between network designers. ~ecognizingsome of the references to concepts already formally represented in the system provides domain- and design-specific cues as to the content of the object. Based on the occurrence of these text references, the system suggests new attributes for the mail message,as shown in Fig. 14.6. In this example, the system has recognized references to devices,and places already known to the system. Further, these new attributes can be used later to locate related information.

When many designers participatein a project,it is difficult for each to know him, the whole project.How can each be informedof the changes that affect and how can a designer find old knowledge that affects a future modification? The issue of how to make information in group memor~avai~ableto designers presents formidablechallen~es. Our prior work on critics [Fischer et al. 931 and information delivery er, ~akakoji911 for individual work has provi~edthe potential solucreating mechanisms that automatica~ly vo~unteerinformation.

14. SEEDING, E V O L ~ I O N ~ GROWTH, Y AND WEEDING

FIG. 14.6. Property sheet with suggested attributes. Suggested attributes for an impo~edelectronic mail message appear in bold in the property sheet. The text of the mail message is shown in Fig. 14.5.Whenan attribute (e.g., "Devices Involved") is selectedby the user, it appears in the top portion of the property sheet, where it may be edited.

the design en~ironmentvolunteer information leads to issues such to determine what information to volunteer, how to volunteer it, how to let designers interact with the environment to determine what information is volunteer~dto them andto other designers.In the contextof WORK, we have begun investigating the use of agentsto sup~ort communication. ~ o m ~ a rto e dother work on agents providing information we have on how designers, rather than text analysis algorith with compl~xinterest profiles, can directly determine what mation is ~ o l ~ ~ t e eTo r e ddo. this we have focused on how e rest of the information in the colla~orativedesign enviare created and evolve throu~hall phases of environment initial setof agents are created during seedi

FISCHERET AL.

FIG. 14.7. Agent editorshowingcurrentstatus of anagent.Designerscan define and edit agents in ~ E ~ agent O editor. ~ By s selecting a trigger, a query, and an action, designers can decide the information to be displayed, the situation in whichto display it, and the manner in whichto display it.

.

14. SEEDING, EVOLUTION~YGROWTH, AND RESEEDING

[Fischer et al.911. Critics use knowledgeof design principles for the detection of suboptimal solutions constructedby the designer. One of the challenges for critiquing systems is avoiding work disruption. Our systems accomplish thisby making the critics sensitive to the specific design situation [Fischer, ~akakoji911, incorporating a modelof the user [Fischer et al. 911, and giving users control over when and which critics shouldfire[Fischer, ~irgensohn901. Partof the JANUS project [Fischer, orch 89b] focusedon building critics which are useful in spite of highly intelligent (i.e., more intelligent than a designer). Simple critics can be informative because they are based on domain~ o w ~ ethat d~e designers might not have (e.g., network designers are not necessarily familiar with relevant knowledge about fire codes for bui~dings). Problems often arise in the use of collaboration technology when some people are required to do additional work to support a system that primari1y ~enefitsother users [~rudin 94bI. This is particularly important for §ystems that would incorporate design rationale information, a task [Fischer et al.911. The system and process we have descr

environment W

FISCHER ET AL.

entation sets our approach apart. In particular, domain orientation is an interesting perspective from which to view two major challen S for shared and evolving information spaces: the development of classification conventions that support information location, and the ability to actively deliver information to users when it is relevant to their task at hand [Fischer931. et al. Systems designed for general information storage and retrieval face the difficult task of developing information categories that make sense to the individuals who share the information [Berlin et 931. al. General categorization schemes are dependent on the group members that develop and use them, and therefore will change as group members come and go. Design domains, on the other hand, are characterized by domain-specific conventions that have relatively precise and stable meaning to domain practitioners. Domain conventions have developed over time to enable designers to conceptualize design problems and to communicate important ideas. The relative stability ofdomainconventionsmakesdomain-orientedsystems less sensitive to turnoverin group personnel. ~eneral-purposeinformation spaces can have only a limited notion of the user’s task at hand. Domain-oriented design environments exploit domain semantics and the design context [Fischer et93)al.to actively notify designers when there is information about which they should know. Active inforthe design mation delivery helps designers to detect inconsistencies inearly process, and to learn about design concepts of which they were unaware.

This chapter has describeda process model for the evolution of domain-oriented design environments through use.We consider design en~ronments as seeds thatgrow by accumulating design knowledge as they are used to supportdesigntasks. ~eriodically,a reseedingprocessisnecessaryto ensure that new knowledge is accessible to the design environment’s computationa~ mechanisms and therefore is accessible to environment. Weclaim thatsuchanapproach is nee design in complex and open-ended domains, in which new designowle edge surfaces in the contextof design tasks. A seed is a collection of knowledge and procedures capableof growingof sustaining gro~h-through interactionwithdomaindesignersdurin use. It stimulates, focuses, and mediates discussion-and thus captur~uring the incremental growth phase. The seed must be capturing the information elicited from the use of the system. Thereisnoabsolute re~uirementforthe compl~teness,correctness, or ecificity of the informationin the seed.In fact, it is often its shortcomin in these respects that provoke input from designers.

14, SEEDING,EVOLUTIONARYGROWTH, AND RESEEDING

Evolutionary growth during system use is a process of adding information related directly or indirectly to the artifact being designed. Thus, the artifact (in our case, the network logical map) is the foundation for evolutionary growth. Duringthe growth phase the designerswho use the system are primarily focused on their task at hand. Information input ishighly situation specific-tied to a specific artifact and statedin particular rather than in general. For a while, information growsin an orderly manner, but eventually order breaks down and the system beginsto degrade in usefulness. Reseedingisnecessarywhenevolutionarygrowthstopsproceeding smoothly. During reseeding, the system’s information is ~estructured,generalized, and formalized to serve future design tasks. The reseeding process creates aforum to discuss what design information capturedin the context of specific design projects should be incorporated into the extended seed to support the next cycle of evolutionary growth and reseeding. Tools conn environmentssupportreseedingbymakingsuggestions about how the information can be formalized.

The authors would like to thank the members ofthe Human-Computer Communication group at the University of Colorado, who contributed substanin this paper. tially to the conceptual framework and the systems discussed In particular, Kumiyo Nakakoji provided invaluable assistance. The research wassupported by the National Science Foundationundergrants No. IRI9015441andMDR-9253245,and NYNEX Science andTechnologyCenter (White Plains,N.Y.).

[Anderson 8 5 1 J. R Anderson, Cognitive Psychology and Its Implications (2nd Edition), W. H. Freeman andCo., New York, 1985. [Berlin et al. 931 L. Berlin, R. Jeffries, V. L.ODay, A. Paepcke, C. Warton, Where Did You Put It? Issues in the Design and Useof a Group Memory, HumanFactorsin C o m p ~ tSystems, i~ ACM, 1993,pp. 23-30. l ~ ~ RConference ~ ~ 9Pr~eedings, 3 [Buchanan, Shortliffe841B. G. Buchanan, E. H.Shortliffe,Rule-Based ExpertSystems: The ~ Y C l ~ ~perimentsof the Stanford HeuristicPr~ammingProject, Addison-Wesley Publishin pany, Reading, MA, 1984. [Carraher, Carraher, Schlie~ann851 T. N. Carraher, D. W. Carraher, A. D. Schliemann, Mathe~ , 3, matics in the streets and in the schools, British Journal of eve lop mental ~ s y c h o l Vol. 1985,pp. 21-29. [Chen, Dietterich,Ullman 911 A. Chen, T,G.Dietterick, D. G.Ullman, A Computer-Based Desi History Tool,~ ~ e eof the ~ 1991 j ~ Ss ~ ~and e ~sa n~u fna c ~ rSystems i n ~ Confe~ence, SME (Society of Manufacturin~Engineers), Austin, Texas, January1991,pp. 985-994.

FISCHER ET AL. [Conklin, Begeman881J. Conklin, M. Begeman, gIBIS: A Hypertext T d for ~ p l o r a t Policy o ~ Discussion, Transactions of Office Information Systems, Vol.6, No. 4, October 1988, pp. 303-331. [CSTB 901 Computer Science and T e c h n o l ~Board, Scaling Up: A Research Agenda for Sobare Engineeri~,Communicationsof the ACM, Vol. 33, No. 3, March 1990,pp. 281-293. [DeMarco, Lister 871 T. DeMarco, T. Lister, People~are:P r ~ u c ~ Projects ue and Teams, Dorset, New York, 1987. [Ellis, Gibbs, Rein911 C. A. Ellis, S. J. Gibbs, G. L. Rein, Groupware: SomeIssues and Experiences, Communications of the ACM, Vol. 3 4 , No. 1,1991, pp. 38-58. [Fischer 921 G. Fischer, Domain-oriented Design Environments, Proceedings of the 7th Annual Knowle~~Based Sobare Engineering (KBSE-92) Conference(McLean, VA), IEEE Computer Society Press,Los Alamitos, CA, September 1992, pp. 204-213. [Fischer et al. 911 G. Fischer, A. C. Lemke, R. McCall, A. Morch, M a ~ n gArgumentation Serve Design, ~ u m a nComputer Interac~o~, Vol. 6, No. 3-4,1991, pp. 393-419. [Fischer et al. 921 G. Fischer, J.Grudin, A. C. Lemke, R. McCall, J,Ostwald, B. N. Reeves, E Ship man, Suppo~ingIndirect, Collaborative Design with Integrated ~ o w l e d g ~ B ~Design ed Environments,Human Computer Interaction,Special Issue on Computer Supported Cooperative Work, Vol. 7, No. 3,1992,pp. 281-314. [Fischer et al. 931 G. Fischer, K.Nakakoji, J. Ostwald, G. Stahl, T. Sumner, Embedding ComputerBased Critics in the Contexts of Design, Humun Factors in Computing Systems, I ~ E R ~ ~ 9 3 Conference Proceedings,ACM, 1993, pp. 157-164. [Fischer, Girgensohn 9 0 3 G. Fischer, A. Girgensohn, End-User Modifiabilits in Design Environments, Human Factors in Computin~Systems, W 9 0 Conference Proceedings(Seattle, WA), ACM, New York, April 1990, pp. 183-191. [Fischer, Lemke 881 G. Fischer, A. C. Lemke, Construction Kitsand Design Environments: Steps Human~omputerInteraction,Vol. 3, No. 3, Toward HumanProb1em”Dorn~n Communication, 1988, pp. 179-222. [Ffscher, McCall, Morch 89a] G. Fischer, R. McCall, A. Morch, Design Environments for Constructive and Argumentative Design, ffumanFactors in Computi~Systems, cH1’89 Conference Pr~eedings(Austin, TQ,ACM, New York, May 1989, pp. 269-275. [Fischer, McCalI, Morch 89b] G. Fischer, R. McCalI, A. Morch, J ~ U SIntegratin~ : Hypertext with a ~ o w l e d g ~ B Design ~ e d Environment,Proceedings o f H y p e ~ e x(Pittsbu t~9 New York, November 1989, pp. 105-117, [Fischer, Nakakoji 911 G. Fischer, K. Nakakoji, Making Design Objects Relevant to the Task at Conferenceon Arti~cialIntelligence, AAA1 Hand, Pr~eedings ofAAAI-94 ~inth ~ational Pressnhe MIT Press, Cambridge,MA, 1991, pp. 67-73. i~ Design Enuironments, [Girgensohn 921 A. Girgensohn, End-User M ~ i ~ a b iinl Knowle~e-~ased Ph.D. Dissertation, Department of Computer Science, Universityof Colorado, Boulder, CO, 1992, Alsoavailable as TechReport CUGS-59592. N a p ~ n ~ o m p u tsupport er for working with dia[Gross 961 M. D. Gross, The Electronic Cocktail grams, Design S ~ d i e sVol. , 17, No. l, 1996, pp. 53-70. [Grudin Ma] J.Grudin, Evaluating Oppo~unitiesfor Design Capture, in T. Moran and J.Carroll (eds.), Des@ Rationale: Concepts,Techni~ues,and Use, Lawrence Erlbaum Associates, Inc., Hillsdale, NJ, 1994, (in press). [Grudin 94b] J. Grudin, Computer-Supported Cooperative Work History and focus,IEEE Computer, Vol. 27, No. 5, 1994, pp. 19-26. An approach [Hackman, Kaplan741 J,R Hackman, R E. Kaplan, Interventions into group process: to improvin~ the effectivenessof groups, ~ e cSciences, ~ i Vol. ~ 5,1974, pp. 459-480. [Hil~ 891 W. C. Hill, The Mind at AI: Horseless Carriageto Clock, AI ~ a ~ ~ jVol. n e10,, No. 2, Summer 1989, pp. 29-41. [Hill et al. 921 W. C. Hill, J. D. Hollan, D. Wroblewski, T. McCandless, Edit Wearand Read Wear, ~ u m a nFactors in Computi~Systems, cH1’92 Conference~ ~ e e d i n (Monterrey, gs CA), ACM, May 1992, pp. 3-9.

14,SEEDING,EVOLUTIONARYGROWTH,

AND ~ E E D I N ~

[Hollan, Stornetta 921 J. Hollan, S. Stornetta, Beyond BeingThere, Proceedj~sofACM CHl’92 Conference on Human Factors in ComputingSystems, ACM, New York, 1992, pp. 119-125. [Hutchins 9 0 1 E. Hutchins, The Technologyof Team Navigation, in P. Galegher, R. Kraut, and C. Egido (eds.),Intellec~alTeamwork, Lawrence Erlbaum Associates, Hillsdale, NJ, 1990,ch. 8. [Johansen 881 R. Johansen, Groupware: Computer Supportfor Business Teams, The Free Press, New York, 1988. [Kuffner, Ullman911T. A. Kuffner, D. G. Ullman, The information requestsof mechanical design engineers, Design Studies,Vol. 12, No. 1,January 1991, pp. 42-50, [Kunz, Rittel701W. Kunz, H. W. J, Rittel, Issues asel em en^ ~ I n f o ~ a t iSystems, on Working Paper 131, Center for Planning and Development Research, University of California, Berkeley,CA, 1970. [Lave 881J. Lave, C~nitionin Practice, Cambridge University Press, Cambridge,UK, 1988. Into HistoryTmls for User Support,Unpublished Ph.D. Dissertation, [Lee 921 A. Lee, Inve~~ations University of Toronto, 1992. [Lemke, Fischer 901 A. C. Lemke, G. Fischer, A Cooperative Problem Solving System for User Interface Design,Proceedi~sof AAAI-90, Eighth ~ationalConferenceon Arti~cial Intell~ence, AAAI Pressflhe MIT Press, Cambridge,MA, August 1990, pp. 479-484. [McCall91]R. McCall, PHI:A Conceptual Foundationfor Design Hypermedia,Design Studies,Vol. 12, No. 1, 1991, pp. 30-41, [McCall et al. 901 R. McCall, P Bennett,P. d’Oronzio,J. Ostwald, F. Shipman, N. Wallace, PHIDIAS: EuropeanConferenceon Hypert~t Integrating CAD GraphicsintoDynamicHypertext, (ECHT’SO), 1990, pp. 152-165. [McCall, Ostwald, Shipman911 R. McCall, J. Ostwald, E Shipman, Supporting Designers’ Access Proceedings ofthe 1991 Conference to Information Through Virtually Structured Hypermedia, on Intelligent Computer Aided Des@, Elsevier, New York, NY, 1991, pp. 116-127. [McCall, Schaab, Schuier831 R. McCall, B. Schaab, W. Schuier, An Information Station €orthe Problem Solver: System Concepts, C. in Keren, andL. Perlmutter (eds.),Applications ofMiniand Microcomputers in Informa~on, ~ ~ u m e n t aand t i ~Libraries,Elsevier, New York, 1983. [Nakakoji 931 K. Nakakoji, Increasi~Shared Understandi~of a Design Task Between Designers and Design Environments: The Role ofa Speci~cationComponent, Unpublished Ph.D. Dissertation, Department of Computer Science, University of Colorado, 1993,Also available as TechReport CU-CS-651-93. [Nardi 931 B.A. Nardi, A SmaH Matter o f P r ~ r a m m The i ~ , MIT Press, Cambridge,MA, 1993. [Polanyi 661 M. Polanyi, The Tacit Dimension,Doubleday, Garden City,NY, 1966. [Reder 821 L. M. Reder, Plausability judgment versus fact retrieval: Alternative strategies for sentence verification,Psychol~icalReview, Vol. 89, 1982, pp. 250-280. [Reeves 931 B. N. Reeves, Suppo~ingCollaborative Designby Embedding Communication and History in Design Artifacts,Ph.D. Dissertation CU-CS-694-93,Department oi Computer Science, University of Colorado, Boulder,CO, 1993. [Reeves, Shipman 92a] B. N. Reeves, F. Shipman, Supporting Communication between Designers with Artifact€entered Evolving Information Spaces, Proceedings of the Conference on Computer~upported~ ~ p @ r a t Work i v e (CSC~92),ACM, New York, November 1992, pp. 394-401. [Reeves, Shipman 92b]B. N. Reeves, F. Shipman, M a k i ~it Easy for Designers to Provide Design Rationale, Working Notes of the A A A I 1992 Workshop on Design Rationale Capture and Use, A A A I , San Jose, CA, July 1992, pp. 227-233. [Repenning, Sumner921 A. Repenning, T. Sumner, Using Agentsheetsto Create a Voice Dialog Design Environment,Proceedings of the 1992ACM/SI~PPSymp~ium on Applied C o m p u ~ ~ , ACM Press, 1992, pp. 1199-1207. [Schoen 831 D. A. Schoen, The Re~ectivePractitioner: How Professionals Think in Action, Basic Books, New York, 1983. [Shipman 931 F. Shipman, Supporti~ ~nowledg~Base Evolution with Incremental Formal~ation, Ph.D. Dissertation, Depa~mentof Computer Science, Universityof Colorado, Boulder,CO, 1993, Alsoavailable as TechReport C U € ~ ~ 9 3 .

FISCHER ET

AL.

[Shipman, Marshall9 4 1 F. M.S h i p m ~and C. C. Marshall, ~ o r m a Considered l~~ Harm~l:&per& ences, ~ m e ~ i nmemes, g and Directions, Technical Report I S ~ S A - ~Xerox ~ 2 PARC, , 3333 Coyote Hill Road, Palo Alto,CA 94304,1994, [Shipman, McCall 941E. Shipman, R McCall, Supporting KnowledgeBase Evolution withIncre mental Formalization, Human Factors in Computing Systems, CHI94 Conference Proceedings, ACM, 1994, pp. 285-291. [Stefik et al. 871M,Stefik, G. Foster, D. G. Bobrow, K Kahn, S. Lanning, L. Suchman, Beyondthe Chalkboard: Computer Support for Collaboration and Problem Solving in Meetings, Communications oftheAGM, Vol. 30, No. 1,January 1987, pp. 32-47. [Suchman 871 L.A. Suchman, Plans and Situated Actions, Cambridge University Press, Cambridge, UK,1987. [Terveen, Selfridge, Long 931 L. G. Terveen, P. G. Selfridge, M. D, Long, From Folklore to Living Design Memory, Human Factom in Computing Systems, I ~ ~ R Conference ~ ~ 9 Proceedings, 3 ACM, April 1993, pp. 15-22. [Thimbleby, Anderson, Witten 901H. Thimbleby, S. Anderson, I. H. Witten, Reflexive CSCMS u p porting Long-Term Personal Work, Interacti~~ i t Computem, h Vol. 2, No. 3, 1990, pp. 330-336. [Walker et a1 871J. H.Walker, D. A. Moon, D. L. Weinreb, M. Mc~ahon,The S y ~ 6 ~ i Genera cs Proam mi^ ~nuironment,IEEE Software, Vol. 4, No. 6, November 1987, pp. 36-45. [Wolf,Rhyne 921C. Wolf, J. Rhyne, Communicationand Information Retrieval with a Pen-based Meeting Support Tool, ~ o c e e d i dthe ~ s Conferenceon Computer~uppo~e~ Cooperatiue Work (CSC~92),ACM, New York, November1992, pp. 322-329. [Yakemovic, Conklin901K. C. Yakemovic, E. J. Conklin, Reportof a Development ProjectUse of an IssueBased Information System, Proceedings of the Conference on C o ~ p u t e r ~ u p p o ~ ~ d C~peratiueWork (CSC~90),1990, pp. 105-118.

C H A P T E R

Q Q

Starr Roxanne Hiltz Jerry Fjerrnestad Youngjin Kim Ajaz Rana Murray Turoff

New jersey Institute of Technology One type of computer-based systemto support collaborative work groupware”; Johnson-Lenz& Johnson-Lenz, 1982;Ellis, Gibbs,& Rein, 1991) is most often called a Group Support System, or GSS. Other terms that have been rt (GDSS; DeSanctis &C Gallupe, used include ~ r o ~ue ~c i s i oS ~u ~ ~ oSyste~s 1987) and ~ Z e c ~ o~~ ei ce tSyste~s i ~ ~ (~unamaker,Dennis, Valacich, Vogel, George, 1991). DeSanctis and Gallupe’s(1987) seminal paper,‘a~oundationfor the Stu of Group Decision Support Systems” has been extremely influential in providing a cQmmon framework forresearch. They definedGDSS as comb in in^ ‘~ommunication, computer, and decision technologies to support formulation and solution in group meetings’’(p. 589). Various types of tools or structuresfor interaction can help a group to avoid process lossesand to achieve better decisions or outcomes. For example, the use ofcom~uter the ommunication can allow everyoneto i n p ~ t s i m ~ l t ~ n e Q ~ s ing greaterequalityofparticipation.Failureto ~uanti~ res canbeovercomebyproviding appro~riatevotin scales and tools, whereas failure to efficiently organize and communi information about ideas and preferences can be overcome by the statis

HILTZ ET AL.

analysis and displayof the resultsof rating or voting. They also presented a ‘~ontingency” theory to help explain why GDSS is not always beneficial; it depends on whether the natureof the technology and structuring provided is appropriate for the group size (smaller vs. larger), the type of task, and the CQmmunication mode, of which they identified two: same place (FtF, or “decision room”) and different place, or dispersed. The term roup Support System (GSS) has cometo be used as more general and inclusive than GDSS, which is often used to imply decision room, same time settings.GSS can apply to many stages and types of group work, not just decision making, and to computer support for groups that are working asynchronously through wide area networks as well as at the same time. ~ i s ~ i ~ roup u t e S~pport ~ Systems embedGDSS type tools and procedures within a Computer-Mediated Communication (C C) system to support collaborative work among dispersed groups of people. “Distributed” has sew eral dimensions:t~mporal,spatial,and technological, The majority ofGDSS research has been conducted in “Decision Rooms,” where the participants are meeting at the same time and same place. CMC-based systems can be used synchronously (same or different places, but at the same time), or ~ 5 y ~ ~ ~ r o ~ oThe u 5central ly. focusof the programof research reported here ous ~rQups, in which interactionis distributed in time as well as group members use the system to work to~etherto reach a mplete theircooperati~e work over a period of time, with each

15. DIST~BUTEDGROUPSUPPORTSYSTEMS

Many GDSS systems that are decision room based simply include a setof tools and proceduresin the “package” that is always provided as partof the system; thus, the effects of medium of communication are confounded with the effects of specifi~tools and procedures. Our objective has been to isolate specific tools and procedures and explore their effectiveness in the a$ynchronous environment. Very few otherGDSS experiments have looked Of the 120 GSS experimentspublished attheasyncnronouscondition. through the end of 1995 that we have been able to identify (~jermestad& Hiltz, 1996), besides the threeNJIT experiments published to date, only five other experiments have used an as~nchronouscondition, plus two that compared s~chronous and asynchronous conditions, and only one other inves, has tigator(Chidambarami 1989 ChidambaramiBostrom, & ~ y n n e1990) conducted more than one experimentin the asynchronous mode. New Jersey Institute of Technology’s project is an integrated program of theory building, software tool development and assessment, and empirical studies (both controlled experiments, and as opportunitiesarise, field studies). The project investigates the effectiveness of different types of tools and within the ~istributed proceduresforvarioustypesoftasksandgroups,

n. ~efirst$ummarize

.

HILTZ ET AL.

nous vs. as~chronous); group size and task type. For task type, we currently use ~ c ~ r a(1984) t ~ “task s circumplex.” The graphical representation of this

typology (not included here) differentiates tasks on two dimensions. The first dimension classifies tasks on the basis of outcome: intellectua~ (e.g., a decision) or behavioral (e.g.,a “product” or action). The second dimension uses the typeof behavior of group members (convergent or cooperative, vs. conflicting).We are presently focusing on the four types of cooperative tasks: Generatns(Type 1: ~lanningandType 2: Creativetasks)andChoosing solu a specificproblem(Type 3: lntellectivetasks,whichhave a 4: “correct” or optimum decision whose quality can be measured; and Type Preference, for which the objective is to reach agreement), We extended this initial framework to produce a more comprehensive of different theoretical foundation thatwill enable us to compare the results studies and to compare our results to those of other researchers.One com& Turoff, 1993) reviewed the major research pleted paper (Fjermestad, Hiltz, models that have been used for studying G§§’s and derives and presents an integrated, compr~hensivemodel of the factors that are utilized for their investigation. It shows that in the short time since the publication of the ~e§anctisandGallupe(1987)foundationpaper,thenumber of research dimensions includedin various models has more than doubled (fr There hasalso beena shift in research emphasis from the techno interaction among the technology, the task, and the group topr comes. The model organizes all the variables that have been used in G§§ research into four dimensions: contextual, intervening, adaptation, and outcomes. A concise overview of the theoretical framewor~,showing the version we began with, is shown in Table 15.1. There, the contextual factors are shown at the top; the intervening variables, and the outcomes or dependent variables are at the bottom. The model served as the theoretical framework for all individual studies carried out within the program of research. The Conte~tualfactors are all external or driving variables that comprise the environment or conditions for the decision-makin task. For any one experiment, theyare (relatively) fixed or contr ed. These include characteristics of the group, task, environmental and o the particular technolo (G§§) beingused. Intervening factorsare related to the emergent structuring of the group interaction, both derived from and adding to setthe of condi~ions created by group decision sessions. For example, the methods used by the group may vary as to session length, numberof sessions, and presto sesence and roleof a facilitator. These factors can change from session thus are dynamic rather than static. ined by the effectsof single elements

15. ~ I S T ~ B U T EGROUP D SUPPORT SYSTEMS

TABU 15.1 Theoretical F ~ e w o r kDistributed Group Support Systems

GDSS Medium Tools Procedures Sync~onou~Async~nous ~ u i p m e n tAccess interface Training

Type ~ u i v ~ i ~ ~ a l y ~ i ~ t y Difficulty Time Importance ~Joy~ility

Computer ~ t u d ~ s ~ l l s Language skills ~mogr~hics Self~onsciousness/aw~ness Values I n t e ~ ~ o nOrientation al Initial Quality

up Size

B~dwidth Social pmsence I n f o ~ ~ richness on Constraints Ease of Use Ease of Learning

Leadership Homogeneity structure Identity~sto~ Initial consensus II

i

~ r o u p / ~ r o ~~e s~s p t ~ ~ o n 1. Level of effort 2. ~ r o p o u ~ ~ d ~ i n g 3. Emergent s t ~ c t u ~ n e ~ o ~ ( ~ u a l i t y / d o ~ n ~ c e ) 4. Structuration: Faithful vs. ironic use; Attitudes; Comfort, Respect; Con~o~reinventions h v e l of consensus

Quality Completion (inst~ctiondependent) Comtness (task type dependent) Absolute % Improve~ent Relative Collective I n t e ~ i g e n ~ Creativi~(task type dependent)

Consensus Absolute Improve~nt

SubjectiveS~sfactionwith Group, Group Solution Own Performance, Task ~ ~ i l i ~ t o r / leader p e r f o ~ ~ c e GDSS: Tools; Functions;I n t e ~ a ~ ; Procedw Co~uni~on M~ium Discussio~GroupProcess

(such as techno lo^ and taskcharacteristics),but by a complex and continuous process in which those elements are appropriated by the group. The four dimensions of the construct (level of use, attitudes toward the GS level of consensus,and levelof control) are measured in all studies via questionnaire items designed and validated by Scott Poole. For each of these aspects of group ~ppropriation,there can be effective or ineffective modes.

HILT2 ET AL.

For example, the group may use the GSS facilities little or not at all, even though theyare instructed to, or theymay use it in a very different manner than was intended. Finally, the outcomes,or dependent variables,are the resultof the interplay of the intervening, adaptation, and contextual factors. They include efficiency measures (e.g., calendar time to decision), effectiveness measures (e.g., number of different ideas generated or decision quality), and subjective satisfaction measures. Itshouldbenotedthatworkiscompletedonfurtherdevelopingand applying the integrated framework to a comprehensive comparison of over 100 published experiments on GSS’sto date (Fjermestad & Hiltz, 1996). Different outcomes have been observed depending upon the initial set of indeby intervening varipendent variables and the group processes (influenced ables) that resultin a specific adaptationor “adaptive structuration” (Poole & DeSanctis, 1990) of the technology provided.By focusing on differences in the variables controlled and studied, a foundation is provided for understanding differencesin findings. ‘

NJIT’s EIES 2 is a CMC enhanced with GDSStools, that provides the foundation that allows continued evolution and the incorporation of additional functionality (Turoff, 1991). As a result of this capability, we were able to enhance the system and to create a “developer’s kit” to Ph.D. allowstudents, or others, use to a version of Smalltalk to develop their own features or interface characteristics.EIES 2 is based on an object-oriented database and a compiler for theX.409 communication database specification language. This base allows the evolution of new object types as theyare needed. To support grou~rientedobjectives aCMC system must allow other computer resources to be integrated within CMC the environment. The approach we have chosen uses the metaphor of an ‘hctivity’’ that can be attached to any communication item. “Doing” an activity executes a program or procedure on the host computeror the networkof users’ computers. rk has been completed on many kinds ofactivities.One replicate the functionality provided by Minnesota’s SA Aided Meeting Management) for a groupto createand revise acommon list of alternatives,and then applyseveral typesof voting proceduresto this list: vote for one, vote yes or no on each alternative, and rate or weight each alternative. Another activity, “Poll,”allows the construction, response to, ts from a pollor survey,within theCMC environment. A sponse” Activity, supports processes such as Nominal roup Technique, dialectical inquiry andbrainstorm in^, as well as applica”

15. D ~ S T ~ CROUP ~ ~ T SUPPORT ~ D SYSTEMS

tions to collaborative learning. Each participant must independently (and possibly anonymously) respond to a problem or question, before seeing the responses of others. Many other activities have been developed to support other kindsof group tasks, includinga class “gradebook activity forVirthe tual Classroom~M. It should be noted that the series of studies reported here were conducted between 1990 and 1995 with a VT100 based full screen menu-type interface, rather thana GUI (point and click Graphical User Interface,) The VT100 type interface has the advantage of being usable on anyPC or even a‘ ~ u m b terminal”; whatever the subjects may have had available to use at home or at work would suffice. Later, we completed work on a Web-based GUI that can be used with browsers such as Netscape, for those who have the necessary equipment and prefer thisstyle of interaction.

. Eachof these experiments represents an attempt to find appropriate tools and processes to support four different types of task in the McGrath “task circumplex”; they examined: Voting tools and sequential procedures fora preference task; * Conflict vs. Consensus structures plus experience for a planning task; * The effects ofFtF vs. distributed asynchronousCMC as it interacts with a structured design procedure, for a creative task, software design * Question-Response tool and the Polling tool for an intellective task (peer review) * Designatedleadershipandsequential vs. parallelproceduresfora mixed task, choosinga stock portfolio. concise summary of the methods and findings of the experiments is presented in Tables 15.2and 15.3. Each group had1 or 2 weeks to complete each decision (depending on the experiment). This isa relatively long time period, compared to the 10 or15 minutes that some experimental tasks usedin decision room GDSS experiments have taken. Unless otherwise noted, all subjects were undergraduate and graduate students from the Computer Science and Management degree programs atNJIT and Rutgers. Students participated as a course assignment, and were graded; alternate assignments were offered for those who chose not to participate. It should be noted that in asynchronous groups interacting over a week or more, group size cannot be truly controlled. Despite the grade incentive, some students dropped out of the group interaction, perhaps because of illness or computer problems,

A

cl

p

U

a

I

.*

15. ~ I S T ~ ~ U TGROUP E D SUPPORTSYSTEMS

and thus decreased group size below the starting number. When group size is reported, it refers to the ending group size, not the number who were trained and began a task. In all experiments, if this effective group size fell below two, the group was dropped from the analysis. Mostof the studies also used expert judges to rate some aspect of the quality of the group outcomes.In all cases, at least three judges were used. They consisted of faculty members or advanced(‘YSD”) graduate students with expertise in the area. Coding and rating procedures were developed and refined during pilot studies, and judges were trained with pilot study transcripts beforebeing given the experimental datato rate or code. Thebrief overviews that follow are essentially extended abstracts of parts of dissertations that total 300 to 550 pages; obviously, in just a few pages, many details such as acomplete listof hypotheses with justifications, and specificsof measurement of variables, must be omitted.

This experiment, carried out by Dufner (Dufner, 1995; Dufner, Hiltz, Johnson, & Czech, 1995; Dufner, Hiltz, & Turoff1994),is a replication (modified for of thedoctoral implementation in as~chronousmode)andextension research conducted by Watson (1987) at the University of Minnesota. The study, which was preceded by a full year of pilot studies, extends the ats son research to include an investigation of adaptive structuration, media richness, system and task expectations, and training (Dufner 1989, 1995; Hiltz, Dufner, Holmes, & Poole, 1991). The Foundation Task, developed at Minnesota and used for this experiment, can be classified as a preference allocation task based on the ~ c ~ r a t h Circumplex model (McGrath, 1984). The subjects play the role of a foundation board, and are to reach consensus on how to allocate funds among “applicants” representingvery different kinds of objectives, such as cutting local taxes, helping the homeless, or improving the town library. .I. I. assigned were . Individuals randomly to groups, as much as was possible given time constraints and schedules. These groups were then trained for approximately 3 hours in the use of the medium (EIES 2) and in working together to performa group decision-making task. All groups were given a suggested agenda (“define the problem,” etc.), as used by Watson, and their conferences were seeded with root comments correspondingto each stage in this suggested set of activities. TOOLS groups were also given trainingin the use of the “List Activity” of lists, and (an electronicflip table) for group generation and management in the useof the “Vote Activity,” which provided three forms of voting on the

HILTZ ET AL.

items on the list. Groups assigneda to SEQUENCED condition were instructed as follows: “You must all work on the same agenda item together. The group decidis when to move to a new agenda item. You do not have to follow the agendaorder. However, you must all work on the same agenda item together. You are asked not to work ahead of or following the group.’’ The “not sequenced” groups were not clearly instructed that they were free to work in parallel; they simply were not given these instructions. After training, each group was given5 business days to perform the experimental task. Groups were instructed to communicate only through the medium.No formal facilitation was provided to the groups, although technical assistance was given when anyone asked for help. There were a totalof 31 groups with 119 subjects; group size varied from 3 to 8 subjects. l t ~ . Frompilotstudies(Dufner,

e knew that groupsin asynchronous mode encounter coordination problems (Dufner, Hiltz, that cause frustration with the medium. Therefore, we hypoth TOOLS and SEQUENT~LP R O C E ~ U would ~ S makesignificant contributions to subjectively reported perceptions of medium richness (Zmud, Lind, & Young, 1990) and satisfaction with the process in the asynchronous environment. No significant difference in the SEQUENCED versus NOT SEQUENCE^ groups was found, Re-examining the manipulation, we decided that we could not determine whether this was because imposing a sequenced process truly makesnodifference, or whetherthemanipulationwasnotstrong enough. Therefore, we noted that examining the sequenced vs. parallel process should be tried again in a subsequent experiment. The TOOLS Groups perceived more media richness, reporting that the medium was significantly more dependable, convenient, flexible, and wideing than did the groups not supported with tools. The TOOLS groups also perceived the system as more personal, more rich, and as providing more feedback and more immediate feedback thandid the groups not supported with tools. TOOLS groups were also found to be significantly more likely to recommend the system for future meetings; to have enjoyed their participation in the discussion; and to have a higher opinion of the overall quality of the discussion thandid the groups not supported with tools. These experimental results seem to indicate that user perceptions of media richness and of the quality of group processes can be improved by providing voting tools that support group discussions, at least for preference type tasks, where the primary goal is to reach consensus. This contrasts with the findings by Watson (1987) for the same task and the same type of tools in a synchronous (Recision Room) environment, where tools created few si~nificantdifferences.

15. D I S T ~ ~ U GROUP ~ D S U ~ P SYSTEMS O ~

Despite the significant and consistent positive effects of providing the list-

ing and voting tools on subjective perceptions, there were no significant results on other dependent variables measuredin this experiment, including changes in level of consensusandthegroup’sequality of influence.

Because this experimentdid not have decisions that could be rated on quality, the effectivenessof tools such as “List” and “Vote” should be examined forintellectivetasksorothertasksforwhichqualitymeasurescanbe obtained, in the future.

The basic objective of this longitudinal experiment (Fjermestad, 1994; Fjermestad et al., 1995) was to examine the performance and attitude changesof groups involved with strategic decision making in a computer-mediated~ommunications (CMC) environment. The two independent variablesof interest weredecisionapproachandexperience.Decisionapproachconsistedof dialectical inquiry (DI; Schwenk, 1990), whichis a structured approach to induce conflict, and constructive consensus, which is a set of instructions telling the group to reach agreement. Experience consisted of working with a 2 weeks to complete. group on two related but different tasks, each taking Previous researchin the fieldof or~anizational strategic decision making has demonstrated that structured conflict can improve the qualityof decisions (Mason, 1969; ~itroff& Mason, 1981; Schweiger, Sandberg, & Flagan, 1986; Schweiger, Sandberg, & Rechner, 1989; Schwenk, 1990; Tjosvold, 1982) and negativel~affectbothgroupperceptionsandprocessoutcomes (Schweiger et al., 1986; Schweiger et al., 1989; Turoff, 1991). Thetwobasic structured conflict methods are D1 and Devils’s Advocate (DA). Schwenk’s meta-analysis(1990) indicates that for studies that focus on groups, D1 has a slight advantageover theDA. The tasks were unstructured decision-making tasks, with no right or wrong answers; they areType 4 (Planning) based on McGrath’s (1984) task circumplex, and fit Schweiger etal.’s (1986) requirements for strategic decision making. The threespecifictasksused in thisstudyweredeveloped by Chidambaram (1989) andwereweremodifiedandupdatedforuse in an asynchronous communications mode instead aof setof discreteFtF meetings. The Threat of Takeover task was used as a training task for all groups and the Issue of Image and Product Line pansi ion were the experimental tasks. ~ s i ~ f fThe . research designis a 2

x 2 factorial repeated measures design.The factors are decision approach and experience. Groupsin each decision approach were given2 weeks (10 business days) to complete each of two tasks.

HILTZ ET AL.

The 160 subjects used in the study were undergraduate and graduate stuNJlT. dents in computerscienceandmanagementinformationsystemsat They all had some fluency with the use of email and computers and were given course credit and a grade forpa~icipation,All subjects were assigned to groups based upon availability and scheduling constraints. The ideal group size was six subjects per group, but due to the subjects’ scheduling constraints, the actual group sizes ranged from 4 to 7 (Fjermestad, 1994). Ekperimental conditions and taskorders were randomly assigned to the groups, The Dialectical Inquiry Approach (DI) is based on the procedures devel(1986; Schweiger etal., 1989) andTung and Heminger oped by Schweiger et al. (1993), modified to support asynchronous communication and decision making ~jermestad, 1994). TheDl groups were divided into two subgroups, denoted as the Plan and Counterplan subgroups. These groups were in separate conferences on EIES 2. All members of both groups wereto initially develop an individual recommendation (including supporting facts and assumptions) within 2 business days and enter in it a List Activityin the CMC system. ThePlangroupthenhad 2 days to develop a single recommendation. embers read the individual case recommendations and then debated and discussed them in a Question Activity which requires each participant to reply before viewing other’s replies. When complete, a case leader organized and entered the subgroup’s recommendation. This was then submitted 2 days to negate the assumptions to the Counterplan subgroup, which had and develop a counterplan. The moderator then created a new conference for the full group and added the plan and counterplan. Thefull group’s objective was to critically evaluate the plan and counterplan through debate and discussion, and to develop a single final group recommendation.A Voting Activity was available if the group chose to useinit,all conditions. The time limit for this task was 4 business days. The Constructive Consensus Approach (CC) follows the basic method & Turoff, 1991; developed by several researchers (Hall, 1971; Hiltz, Johnson, Nemiroff, Pasmore,& Ford, 1976; Schweigeret al.,1986). The CC groups functioned as one group and were in a single conference for the entire task. Their objective was to reach consensus on a single final recommendation.Each indi~dualgroup member had2 days to develop an individual recommendation. The group then had 8 business daysto examine the case situations systematicallyandlogically, in ordertodevelopa finalrecommendation through debate and discussion.A Voting Activity was availableif the group chose to useit. Based on previous research in FtF conditions cited earlier,it was hypothesized thatDl groups would be superior to consensus-structured groups in terms of effectiveness (decision

15. D I S T ~ ~ U T GROUP E ~ SUPPORTSYSTEMS

quality), but would be less efficient and express less subjective satisfaction than the consensus structured groups.It was also expected that group performancewouldimproveonthesecondreplicationoftheuseofthe assigned decision sturcture, and that furthermore, there would be an interaction.Because D1 isanunfamiliar structure, itwasexpectedthatthe improvement would be more marked for theRI groups. Contrary to these expectations, there were very few differences between Task 1 and Task 2, and no differences betweenD1 and CC groups in terms of group performance. D1 groups required significantly more asynchronous meetingtimeandcommunication to completetheirrecommendations. Depth of evaluation as ratedby judges showed no difference; but perceived depth of evaluation was lowerin D1 than in CC groups. CC groups reported greater decision acceptance and willingness to work together again thanRI groups. ReIativeIy few experiential effects were observed. Thus, no advantageswereobservedforthe D1 approach as compared to a consensus appr~achthat also carefully structured the interaction, but it took more work and produced less participant satisfaction. This study of asynchronous strategic decision making and a study using decision roomGSS by Tung and Heminger (1993) reported thatthere are no differences in effectiveness between constructive consensus and dialectical inquiry groupsin a GSS environment. Perhaps whatis happening is that the GSS techno lo^ is significant~y improving the consensus groups to the point where the outcomesare ashigh as the structured conflict processes in a FtF of environment. Thus,GSS equalizes consensus groups’ performance to that the Rialectical Inquiry groups without affecting decision and process satisfaction and without anyof the process losses.

The experiment conducted by Ocker (1995;see also Ocker, Hiltz, Turoff, Fjermestad,1995) investi~atedtheeffects of distributed as~chronous communication on small groups performing high-level requirements ~ ~ sis and design work. It is the first experiment to examine the so~ware is design processin a fully distributed environment. The experimental task the AutomatedPostOffice (WO); groups are required to develop and reach consensus on the initial requirements and interface design of an W O and to submit these in the formof a written report. TheAPO task, as used in this e~~eriment, is a modification of the task used by the ~ n i v e r sofi ~ Michi~an(Olson & Olson, 1991; Olson, Olson, Carter, & ~torrosten,1992; Olson, Olson,Storrotsten, & Carter, 1993). It is primarily a creativity type task,but also containselements of planninganddecisionmaking

y

-

HILTZ ET AL.

( ~ c ~ r a 1984). t h , The N O task canalso be categorizedas occurring during the early stages of the innovation process (West, 1990). ot~eses. Two variablesweremanipulated in this experiment. The first variable, an imposed process, pertains directly to the degree of coordination required for the effective performance and in the imposed procsatisfactionof groups working ona creative task. Groups ess conditions followeda sequence of steps adopted from research on argumentation and structured communication (IBIS; Kunz & Rittel, 1970; “Design & Moran, 1991). The imposed procSpace Analysis,” MacLean, Young, Bellotti, ess contained three main phases: generation of design alternatives; periodof critical reflection and individual evaluation of alternatives; and group evaluation of alternatives and consensus reaching. The second variable is mode of communication (as~chronouscomputer-mediatedcommunication [CMC] vs. face-to-face). It was expected that as~chronousCMC groups would outperform FtF groups, because of fewer process losses, and the ability of each of the participants to think and work at their own paces. Dependent variables include performance outcome (quality and creativity), and group satisfaction,

H1:Asynchronous CMC groups will produce solutions of higher quality than face-to-face (FtF) groups.

The prob~em-so~~ng structure was chosen due to its capabilityto structure communication and for its fit with the activity of design. It was felt that FtF groupswouldnotneedthisaddedcoordination,becausehigh-level design has its own inherent structure (Olson et al., 1992), but that it would ease the cognitive burden of distributed asynchronous groups. Therefore, an interaction was hypothesized: HZ: CMC structured groups and FtF groups will produce solutions of higher quality than CMC unstructured groups andFtF structured groups.

ased on an analysisof the task requirements (e.g. Guindon, 1990; King & Anderson, 1990; Simon, 1973; West, 1990), it was hypothesized that overall, the solutions produced by the asynchronous groups would be more creative than those producedby the face-to-face groups. H3.CMC groups will produce morecreative solutions than FtF groups.

Again, based on relative coordination requirements, an interaction effect was hypothesized such that the solutions produced by the asynchronous imposed-process groups and face-to-face no-imposed-process groups would be more creative than those produced by asynchronous no-imposed-proc-

15. D I S T ~ ~ U T EGROUP D SUPPORT SYSTEMS

ess groups and face-to-face imposed-process groups. Finally, it was hypothesized that the face-to-face groups would be more satisfied than the asynchronous groups becauseFtF is a “richer,” more personal medium, S. Subjectswereundergraduatestudentsenrolled in an undergraduate upper level systems design course, or graduate students in CIS,MIS, or the MBA program. The majority had coursework and/or job experience directly relevantto systems design. ~roups were requiredto reach consensus on the initial requirementsof the APO and to submit these requirementsin a formal report at the endof the experiment, The asynchronous design groups communicated using the EIES 2 computer conferencing system; eachof these groups communicated in its own computer conference. The experiment lasted2 weeks. The asynchronous groups met together for one %hour training session, whereas the face-to-face groups met twiceafortotalof 6 hours, with the first 2 weeks apart. Both the FtF groups and and second sessions spaced exactly the asynchronous groupsin the imposed process condition were trained on the process using the same script. All as~chronousgroups were trained on the basic useof the EIES 2 system. The FtF groups had a PC and word processor available for creating their finalreports.(Technicaldifficultiesledtwogroupstohandwritetheir report; these were among the longer reports, so it does not seem to have negatively affected them. These were later transcribed.) The computer conferences and FtF meetings were minimally facilitated. The facilitator played the roleof a technical assistant, helping groups with equipment problems and answering questions of a technical nature. All participants completed questionnaires, which was the source of subjective satisfaction data.All groups’ final reports were printed using the same word processing package, to mask indications of mode of communication. Quality and creativity of solution were rated by an expert panel of judges, using procedures and scoring adapted from Olson and Olson (1991).

The overall quality of solution was rated by a panel of expert judgesto be equally good between the as~chronous groups and the face-to-face groups. (Although the asynchronous groups’ were rated as consistently higher, the difference was significant only at the .0’7 level). Contrary to hypotheses, there was no sig~ificantinteraction effect between mode of communication and the presence or absence of an imposed process in relation to qualityof solution. As for creativity, the solutions produced by the as~chronous groups were judged to be significantly more creative than those producedby the

HILTZ ET AL.

face-to-face groups. Again, there was no interaction effect between mode of communication and the presence or absence of an imposed process. Contrary to hypotheses, there were no significant differences between CMC and FtF groups on key measures of subjective satisfaction: perceived depth of analysis, solution satisfaction, and decision scheme satisfaction. ns. The imposed process was hypothesized to benefit in asynchronous groupsby providing the added coordination that is missing this form of communication.There are several possible explanations for why this did not occur. Concerning the design of the APO, a strong metaphor is availablein the formof automatic teller machines. The problem may have been familiar enough to groups, so that the need for coordination might havebeengreatlyreduced;uponenteringthegroup,groupmembers already knew how to approach the solution to this problem. The major finding of this experiment is that groups that communicated asynchronously,whethertheyfollowedastructuredproblem-solving approachdesignedtoenhancecoordination orwereleft to theirown devices to reach a decision, produced significantly more creative results than the face-to-face groups. A tentative conclusion is that asynchronous communication,in and of itself, leads to higher levelsof creativity. Possible explanations for this include a greater amount of communication over an extended period of time, reduced production blocking, and the production of a collectivememory.

The task for this experiment consistedof review and decision on publishability of a manuscript submitted toa refereed journal or conference (Rana, 1995). Contrary to the traditional review process, where two or more expert review and rate the quality of a manuscript individually, the distributed group support system based review process as adoptedin this experiment called uponreviewers to undertake the task as part of a group, or panel. This new mode for conducting a review involved performance processes are that typical of intellective,decision making,and cognitiveconflicttasks (~cGrath,1984). We classify it as primarily intellective, because two criteria for rating the quality of the solution were available: the ratings of the article by the actual reviewers of the paper, and the ratings by a panel of expert judges. The desirabilityof a system that can support group review activities in a different-time different-place mode was evident. This study being the first to investigate the viability of a DGSS based review process, motivated a

15. DIST~~UTED GROUPSUPPORTSYSTEMS

research design that would allow the studyof independent and interactive effects of support tools from withina DCSS. Two support tools, Polland Question activities on EIES2, were made available for this experiment, and utilizedin a 2 x 2 factorial design. Po11 activity, especially developed for this research, allows reviewers to rate the of aquality and enables them to view summary statistics on manuscript on various scales group ratings. Several scales can be grouped into one Poll. Question activity establishes a structured formof group discussion by requesting the provision of justifications for ratings on individual scales and maintaining an independent chain of discussion on each scale. Group members' responses to Question activity are textual items with no limit on size, whereas Poll activity responses consist of numbers representing scale anchors. One important feature, common to both Poll and Question activities, is that group members cannot view others' responses before having provided their own initial responses. ~ r ~ An s . EIES 2 conferencewasestablishedforeach of 33 groups, with30 group. The final dataset for this experiment consisted groups of Size three and 3 groups that began at Size 3 but ended at Size 2. The majority of subjects were graduate students (73%); all subjects were enrolled in courses that required them to read and critique journal articles, and to useEIES 2 as a regular partof coursework.The mean age of subjects was 30 years with an average full-time work experience of slightly over 5 years. Because some subjects were enrolled in distance sectionsof courses, the manuscript and training materials instructing them how to use the tools and procedures for their condition were mailed to all subjects, rather than being explained in a face-to-face training session. Subject groups reviewed a manuscript actually submitted to a refereed source with a pending editorial decision. The selected manuscript met the criteria developed as a result of three roundsof pilot studies and was judged to be commensurate with ability levels of the potential subject population, Croups in all four conditions were to individually evaluate the manuscript, provide ratings and justifications for the ratings on six scales; share their responses with the group, and finally, discuss and reach agreement on ratings. This four-step process was to be completed over aperiod of 2 weeks with Steps 1 through 3 completed by the endof the first week. The identities of group members were concealed through the use of pen names. Three measures for quality of group outcome were adopted: quality of the decision (disposition recommendation); quality of the review;and comprehensiveness of the review.A panel of expert judges independently rated the paper on the same scales as the subjects, and their ratings were compared to those producedby each group. Disposition recommendation categories consisted of Accept as is; Accept with minor revisions; Major revisions;orReject. Thefourexpertjudgeswere evenlydivided onmajor revisions or rejection. Thus, either of the latter two recommendations were

HILTZ ET AL.

considered “correct”in assessing qualityof the decisionin terms of the correctness of the disposition recommendation, The quality of the review was calculated on the basisof the deviation of the group’s decision from the judges’ decisions on the separate aspectsof the manuscript (literature review,me tho do lo^, presentation style, andso on). The comprehensiveness measure consistedof counts of the number of lines of discussion associated with eachof the separate scales: for example, ‘~ubstantiveemphasis” was the amount of attention paid to critiquing the literature review; methodological emphasis, stylistic emphasis, interpreti~e emphasis, and wisdom were lines devoted to methodological critiques, and so forth (~ummings, Frost, & Vakil, 1985). Adaptive structuration was measured by a seriesof questionnaire items. t ~ ~ s Because ~ s . Question

activity imposes a structure for gr S to engage in the group proceedings and coordinate their activities, it was expected that Question activity groups would do better. than ~ ~ ~ u e s t i o n three measures of quality (Easton, Vogel, & Nunamaker, 1989; & DeSanctis,1989). (Poll activity was expected to be primarilya consensus~nhancing tool, rather than a quality~nhancing tool; these resultsare not included here.) There were also expected to be some interactions between Poll and Question.In all cases, a moderating variable must be prediscussion levelof agreement; if most of the individual reviewers agreed on their initial ratings, one would not expect any of the tools used to make much difference. In addition, many hypotheses were developed with the basic premise that positive forms of Adaptive Structurah levels of comfort, consensus, and respect regarding the tools) would be strongly related to favorable outcomes. o f f c l ~ s ~ o fN f so. differences in the quality of decisionweredetectedduetoQuestionorPollactivity. In fact,most groups reacheda decision that the paper could not be accepted; there thus, was very little variance froma correct decisionin ratings for dispositionof the paper, and hence none of the independent and intervening variables were significantly associated with this measure. With respect to the qualityof the review, the results showed that improvement depended upon the level of prediscussion agreement. If groups started with a lower levelof initial agreement, then the quality of review was enhanced by the tools, Specifically, at lower levels of initial agreement, groups with the Question activity producedsignific~tlyhigher quality reviews than No-Question groups.Highly agreed on poorquality ratings before the discussion phase left little or no opportunity for an improvement in the quality of review through discussion. Unexpectedly, the Poll activity showed a marginally significant( p = 0 . 0 7 main ~ ~ effect on the quality of the review. This main effect was attributed to the fact that Poll activity groups had significantly lowerof levels prediscus-

15. D I S T ~ ~ U GROUP ~ D SUPPORTSYSTEMS

sion agreement than No-Poll groups. No significant effects on quality of the review were observed due to the modes of appropriation. In terms of the effectson the measuresof comprehensiveness, relatively few effects of the Question activity were supported.Oneof the significant findings was that groups that used the Question activity had significantly in higher wisdom (concern for the paper’s contribution and significance) their reviews thanN~~uestion groups. ~ e d i a t i effects n ~ of the levelof prediscussion agreement similar to those for qualityof review were observed on the amount of methodological emphasis. An unexpected, although not surprising, result was thatin the absence of the Question activity, the Poll activity had a negative effect on interpretive emphasis. Mediating effects ofthe modesof appropr~ation (DeSan~tis & Poole, 1994; Poole & DeSanctis, 1990) on measures of comprehensiveness wererare. The level of challenge was observed to be one of the stronger mediating factors. Additionally, it was observed that a higher levelof respect for the system did not always lead to an enhanced level of performance. Respite the fact that the majority of the hypothesized effects were not observed, the experimental findings offer important implications for the review process.In summary, it was concluded that the strength of the RCSS based review process liesin its ability to allow for (a) disagreement among (b) the subsequentopportu~ity reviewers before the discussion phase, and for resolution of the disagreements with the use of support tools. These mechanismscombinedwithanonymouscontributionscanbeprofitably used to avoid many commonly noted dissatisfactions with the traditional peer review process (Cole, Rubin, & Cole, 197% ~ahoney,1977, 1978, 1985; Peters & Ceci, 1982; Rana, Hiltz,& Turoff, 1995).

Silver (1990) defined system restrictiveness as the degree to which and the manner in which a system limits its users’ decision-making processes to a subset of all possible processes. The objective of this study is to examine how the use of coordination structures with different degrees of coordinain a distion flexibility, or system restrictiveness? affect group performance tributed asynchronousGSS (or DCSS for short) environment. The investment club task, developed for this study, is classified as primarily an intellective task. A group was asked to select at least one, no butmore in 6 than three stocks froma list of 15 stocks to maximize its portfolio value months. Six monthsafter the experiment, all portfolio values were calculated and ranked to evaluate decision quality.The task also has aspects of a planning task, as the group had to decide what information to gather and

HILTZ ET AL.

how to evaluate this information, in order to reach a decision; and of a preference task, because the group had to reach agreement, and at the oftime the decision, there was no way to actually know what decision would turn out to be best months 6 later.

. Theexperiment (Kim, 1996) was conducted with a2 x 2 factorial design. There were 212 subjects in 47 groups. The subjects were Rutgers, NJIT, and Fairleigh Dickinson University students enrolledin various degree programs. All subjects in all conditions were given the same basic agenda as a coordinating structure, consisting of root comments in their conferences which requested them to define objectives, decide on criteria, review the candidates (stocks, in the case of the experimental task), evaluate the candidates on the criteria, and reach agreement on selection. Training was given to all subjects in the form of a week long as~chronous conference, which included a practice task, the selection of a leaderby following the agenda, ending with the sending of a private mesby each member, giving a rank ordering. sage to the experimenter Four coordination structures were created with two independent variables. The four coordination structures were different in that each structure restricted interaction in a different way. In parallel communication groups, by all memall discussion items were presented, and discussed concurrently bers of the group, throughout the experiment. Sequential communication groups discussed one item on the agenda at a time. Once a group moved to the next item, revisiting the previous items was not allowed. In sequential groups, moving from one discussion item to the next item (in groups with a leader), or by a timetable was made by a leader’s decision (in groups without a leader).In sequential groups without a leader, the discussion deadline for each discussion item was announced to the groups at the beginningof the experiment. In sequential groups with a leader, the leader made the summary of the discussion of the item for the group, and opened the next discussion item.In parallel groups with a leader, the only requirement for a leader was to summarize group discussion in aonce while. Groups without a designated leader heard no more about this topic after the training. For thosein the Leader conditions, the experimenter used the individualrankings toarriveatthe most preferredleader. Thiswas announced in thegroup’sconference,alongwiththeleader’srole. The leader was specifically empowered to assign a division of labor, and requesta ed to trackand summarizethe group’s progress. Leaders also were input “leadership conference” where they could ask questions about the role. r groups that expected It was supportedwithalessrestrictivestructurewouldperformbetterthan groups supported with ahighly restrictive structure. Previous researchon

15. DIST~BUTEDGROUP SUPPORT SYSTEMS

DGSS indicatedthatanimposedcoordinationstructurecanbe overly restrictive due to the limited bandwidth of the interaction medium (Hiltz, Johnson, & Turoff, 1991), and the need to synchronize indi~dualactivities. Previous research also demonstrated that GSS a with a high degree of system restrictiveness hadnegativeimpactsongroupperformance (Chidambaram & Jones, 1993;McLeod & Liker, 1992; Mennecke, Hoffer,& ~ y n n e , 1992). GSS’s with a high degree of system restrictiveness leave no freedom for the group to adaptively structure the system to its own preferable decision strategy (DeSanctis & Poole, 1991; Poole & DeSanctis, 1990). DGSS, in which coordinationof individual activitiesis one of the major requirements, to should not behighly restrictive. Research indicates that individuals come the group with a relatively inflexible preference for a particular decision making strategy (Putnam, 1982). Therefore,DGSS should be flexible enough to allow the individual freedom to concentrate on aspects of the problem to which he or she can best contribute (Turoff, Rao, & Hiltz, 1993). A previous experiment with synchronousCMC (Hiltz, Johnson,& Turoff, 1991) found thata designated leader electedby the group could improve the quality of decision for an intellective task. Therefore, it was hypothesized that this would also be true in a distributed environment. in dependent variables were not signifMany of the observed differences icantly related to experimental condition. However, objective decision quality, evaluated with actual portfolio values 6 months after the experiment, was significantly better for leader than for no-leader conditions. Parallel groups perceived that their decision quality was better than of sequenthat tial groups. Parallel groups also had higher decision qualityas objectively so ( p = J4). The average lengthof comments measured, but not significantly in the Leader conditions was longer than in the No Leader conditions;there were no other differencesin consensus and participation. Satisfaction witha coordination process was higher in sequential groups, a designated leader than groups with a leader. and higherin groups ~ithout (As in many other studies, these subjective satisfaction results run counter to the objective quality results). Satisfaction with the group, however, was higher in parallel groups. Sequential groups reported more improved understanding of the task structure than parallel groups.

It isinterestingtonoticethat se~uentialgroups sfactionwiththecoordinationprocessandmore improved taskunderstand in^ than parall~lgroups. This iscontra~yto hat was expected, butconsi~tentwith some previous research, which indicates that GSS should be designed with some degreeof restrictiveness (Dickson, in groupinteraction obinson, 1993). Toomuchfreedom c~hesiveness.This, in turn, increases the decision cost ing a lower quality decision or taking more time to make a

HILTZ ET AL.

decision. Therefore, a coordination structure in Distributed GDSS (DGSS) should impose some restrictions on interaction to maintain a certain levelof group cohesiveness. In future research, the degree of system restrictivenessof a coordination In this study, sequential coorstructure needs to be defined more precisely. dination was assumedto be morerestrictive than parallel simply because it has more procedural order. However, the findingsof this study can not be generalized, or comparedtothefindings of otherresearch,unlessthe degree of restrictiveness can be objectively determined. Very little is known about what determines the perceived degreeof system restrictiveness. Although leadership is the process of coordinating the activities of group members (Jago, 1982), there were not many significant findings related to the leader variable.One explanation may be that, though research on lead1959), ership explains the varietyof leadership styles (House, 1971; Stogdill, the leader functionin this study was too narrowly defined. All leaders were expected to behave exactly the same as they were instructed, regardless of their natural leadership style. Implementing leader’s functions in DGSS as process structuring toolsis one of the requirementsin designing DGSS. The successfulimplementation of leadershipfunctions in DGSS, however,is dependent on the further understanding of coordination effectivenessof different leadership styles within different contingent factors of DGSS (Turoff et al., 1993). Little research has been done in this area,

On the basisof prior studies, several measures were taken to help to assure the effectivenessof the CMC groups in these experiments, regardlessof the

All experimentalconditionormanipulationwhichtheyrepresented. received substantial training and practice before being left on their own for a week or two to do their experimental task (with the exception of the Peer Review experiment, in which all subjects were already system users). All were at least technically facilitated, with the facilitator checking in daily to see if there were any problems requiring assistance (some also had designated group leaders).All had a clearly stated task, objective, and deadline, and all subjects considered the task at least minimally important, as it was a graded assignment.If any of these conditions were omitted, we suspect that the resultswould be negative (Hiltz& Turoff, 1991). There are three basically different conceptualizations about the nature of C, and of asynchronous CMC in particular. One point of view, becoming less prevalentnow that millionsof people are spending hundredsof millions of hours surfing the net for fun, is that it is a “poverty stricken”and “cold”

15. ~ I S T ~ B U T EGROUP D SUPPORT SYSTEMS

medium. This point of view focuses on what it is not: it does not have some of the channels of communication of the face-to-face medium. Mostof those scholars whohavespenttimedevelopingandstudying CMC as a support for group interaction share the assumption that it can be an effective and sociable form of communication, but they differ on how this can best come about. One group views such systems essentially as a techCMC must be built intoa featurenological mechanism, feeling that effective rich andhighly structured and restricteden~ronment.Groups need to have in what are seen as effecthe technology essentially “force” them to behave tive ways to use the medium,in order to minimizeprocess lossesand maximize process gains (Johnson-Lenz& Johnson-Lenz,1991). An example of this approach is theCoordinator(e.g.,Flores,Graves,Hartfield, 1988), or software to forcea completely sequential mode of coordination of interaction. The second approach tobuilding CMC systems conceives them asa context for interaction, “containers” so to speak, just as rooms are. This conception is based on a social theory thathuman systems are self~rganizing and arise outofthe unrestricte~interactionofautonomousindividua~s. From this perspective, the role of the computer system is to provide a place for people to meet and self-organize (Johnson-Lenz& Johnson-Lenz, 1991). CMC is a very different form of communication than face-to-face meetings, and it takes some time for individuals to learn to usethe bothmechanics and the social dynamics of such systems effectively. All of the experiments prein which groups used asynsented here have included at least one condition C, but without time pressure, They had adequate training and k to complete their discussions and produce their group product or decision. Under these conditions, it appears that groups do not need a restrictive,‘~echanistic” approach toco~rdinatingtheir interaction. They are capable of organizing themselves and will tend to feel frustrated by overly restrictive structures or procedures, and/or to become more inefficient. Almost all of our attempts at a mechanistic process intervention hadno significant positive effects on outcomes. For example, there were nosi~nificantdifferences in themajordependentvariablesmeas between the “imposed sequentia~process and the no-proc process groups for the preference task, or for the investment for a creati~tytask, therewas no difference between group an imposedproced~re, and those thatdid not; and fora plan was no difference between groups that used ~ialecticalIn that used a consensus approach. On the other hand, the ~resenceof “tools” tha ready to, does S

HILTZ ET AL.

experimentsincludetheabilityto build a commonlist, a set of voting options, the ‘~uestio~-response activity” that structures the exchange of ideasandopinionssimilartoNominalGroupProcess,thepossibility of anonymity, and a “polling” tool which can allow a group to construct any sort of ~uestionnairetype item, and display results of the polling. One must choose the tools made available to the group very carefully, to match them to the natureof the task and the size of the group, we suspect, though we have not experimented with the option of just “throwing~ all the available tools at a group and lettingit decide on its own what might be appropriate and how to use it. We suspect that even 2 weeks is too short a time to expect a group to deal effectively with this much complexity, but that very longs that interact for months toyears, would do perfectly well with chest at their disposal. The results of these e~perimentssupport the assertion that asynchroC is not like any other formof group communication; not only is it f a c ~ t ~ f aunsupported ce meetings, but it also has very different dynamics thana computer-supported meetingin a “decision room.” Coordi-

15. DIST~~UTED GROUPSUPPORTSYSTEMS

cross-media comparisonsin the future, particularly if we are successful in our quest to obtain the necessary equipment for a stateof the art “Recision Ocker research has been Room”en~ronment.As a startin this direction, the estended thus far to examining two additional communication conditions, synchronousCMC groups, and ”mised mode” groups that have 2 hours of FtF meeting, 2 weeks of aysnchronousCMC,and a final face-to-face meeting. We have concluded that the use of anas~chronousCMC system for GS allows for a much wider range of possible coordination modes and toolsup port thanis effective fors~chronous meetings. All of the esperimentsto date have con fir me^ that even the most estreme asynchronous structures do not reduce the quality of the solutions when compared to more classical coordination and group approaches. The reasons for this cannot as yet be confirmed by any of the experiments, but they are hinted at from some of the results: *

All individuals are free to participate as theyfitsee an

desire to.

* * *

m of partici~ationas an indi~dualseems to encourage: expression of ideas more reflection less inhi~itionof ideas consi~erationof more options

HILTZ ET AL.

may connect through the networks and participate any time, day or night, seven days a week. The software activities developed for this application stress collaborative learning approaches. Field trials of various types with collaborative learning have been takin placeat NJIT since 1980. Currently NJIT offerscompleteundergraduate in Computer Science and many degree programsin Information Systems and additional graduate courses through a remote learning program utilizing asynchronous group communications. What this example and others have taught us, when combined with the experimental workin GSS, is that the key to successful systems is to discard manyof thebiasesthatcomefrommakingcomparisons to face-to-face approaches and trying to adapt an approach of automating the face-to-face environment. Rather, the factors that seem to be crucial to enhancing our understanding of this area and in the future design'of the functionality for such systems include: @

@

@

Providing a "nonlinear agenda" that allows the individual members of indethe group to focus on the contributions that each can best make, pendent of the work of other membersof the group at that momentin time (Turoff,1991). Allowing a group to tailor the relationships structure of comments to fit the application domain as they perceive it. This can only be done by freeing fixed comment relationship structures to provide a full collabo& Hiltz, 1991). rative Hypertext capability (Turoff, Rao, Providing"reciprocal"coordinationstructures(Hiltz & Turoff,1978) that will be able to check on consistency and agreement at the group level and inform participants when they need to reconsider their inputs based on more recent contributions of others. The reason why these factors have not playeda significant rolein most current GSS work has been the typical lackof complexity of the problem being examined.

The other area that our research is focusing on is the software development process and tools to support that task. Initially, the primary objective oing project is to increase knowledge about how to create more roductive systems to support distributed, collaborative groups, particularly for complex software design and planning type tasks. The subtasks in the software developmentarea span a wide range of critical problemare~s:

@

a

The need for enhanced creativity in the design process. Greater understanding of requirements between U ers and ~ e s i ~ n ~ r s .,experts who sometimes speak different languages). The planning of projects and efforts.

15. ~ I S SUPPORT T~~UT GROUP ED

SYSnMS

I

Complex project management, which includes the tracking and monitoring of what has been accomplished, the detection of potential problems and the handoff and coordination of work between different individuals and subgroups. Within systems development, it is recognized that the stages of requirements definition and high-level design are important, and even crucial to the development of effective software. Collaborative designers work to achieve some consensus on the general characteristicsof the new system in question (Olson & Olson, 1991). Ineffective communi~ation during the requirements definition process is consistently associated with user dissatisfaction and lower quality systems, whereas effective communication is associated with improved productivity and higher quality systems (Curtis, Krasner,& Iscoe, 1988). Additionally, it has been increasingly suggested that the development of information systems and the definition of high-level requirements and design, could benefit from the infusion of creative and innovative soluer, Higgins, C% McIntyre, 1993; Telem, 1988). Particularly critical to this area will be the adding of additionaltools and processes (such as group hypertext/ hypermedia authorin Daft and Lengel(l986) were certainly right when they point objective of most work meetings is to reduce both uncertainty and equivocality in unstructured problem solving. Although most work in the Hypertextarea appreciates the utility of nonlinear relationships in the content of the material to reduce uncertainty; however, it also seems self-evident that the problem of equivocality can only be handled by allowing people to perceive one another's reactions to the information. This has always been clear in the contextof asynchronous comto know thestatmunications, where itis critical that each participant needs usof the other members and the roup as a whole. Within the contextof a collaborative H~ertext environment, it becomes necessary for the individuals to be ableto perceive how others traverse the network and how they modify it in a thought process typeof temporal sequence. The conceptof utilizin~Hypertext to support individualsto integrate the differentdomainssupportingsoftwareengineeringanalysis,designand development is not new (Isakowitz, 1993). However, the equally important 1991) has concept of supporting group processes andco~laboration (Turoff, receivedonlyalimitedamount of attention. Thespecificgoalofour research will be tofocusonalltheprocessesassociatedwithsoftware ay be aided by Collaborative Hypertext Systems( 5). There have only been a few specific syst rshali & §hipman, 1993). The currentemergence of a wholenewgenerationof impleme~t tools means thatin the future itwill be much easierto develop specific

HILTZ ET AL.

sion Support functionality and Hypertext capabilities. It also meansthat there will be a major shift back to more internal development of tailored user software within organizations, rather than the current emphasis on purchased software.An objective of future researchin the Croup Support Systems area should be to provide a kind of "checklist"of what kinds of tools and procedures are likely to be helpful for different typesof tasks, so that organi~ationscanbeguided in their self-tailoring of software to fit their needs.

This research was supported by grants from the National Science Foundation program on Coordination Theory snd ~ollaborationTechnology (NSF IF3 901~236and NSF-1~-940~0~). The opinions expressed do not necessarily represent thoseof the National Science Foundation. Among the many people who have contributed to the program of research, in addition to the coa~thors,are Raquel en bun an, RobertCzech,Kenneth Johnson, Cesar Perez, Ronald Rice, Scott Poole, James Witescarver, and William Worrell.

Balasubramanian, V. and Turoff, M.(1995). A Systematic Approachto User Interface Design for Hypertext Systems,~ ~ e e d i28th ~ s"S, , Vol. W., 241-250,1995. Chidambaram, L. (1989). An ~mpiricalInvest~ationof the Impact of ~omputerSupport on Group doctoral dissertation,Indiana ~ e v e l o ~ ~and e n ~t e c ~ i o n -~~~ fao ~ r mi a~n cUnpublished e, Univers~ty. Chidambaram, L.,Bostrom, R. P., and Wynne, B.E. (1990). A Longitudinal Study of the Impact of Group Decision Support Systems on Group Development, Journal o f ~ a n a ~ e m e n t I n f o r ~ ~ t i o n Systems, 7(3),7-25. Chidambaram, L., and Jones, B. (1993). "Impact of Communication Medium and Computer Support on Group Perceptions and Performance: A Comparison of Face-to-Face and Dispersed ~eetings, MIS quarter^, December, 465-491. Cole, S.,Rubin, L.,and Cole, J.R. (1977). Peer Review andthe Support of Science. Scienri~cAmerican, Vol. 237, pp. 34-41. Couger, J. D., Higgins, L. F,, and McIntyre, S. C. (1993). ~n)structuredcreativity in information MIS Q u a ~ e rDecember, ~, 375-397. systems organ~zations. Cummings, L. L., Frost, P. J., and Vakil, T. F. (1985). The manuscript review process: A view from the inside oncoaches, critics,and special cases.In L. L. Cummings andP. J. Frost (Eds.),Pub l ~ hinithe ~ O~an~ational Sciences (pp. 469-508). Homewood, I L Richard D. Irwin. Curtis, B., Krasner, H., and Iscoe, N.(1988). A field studyof the software designprocess for large systems. C4CM, 31, 1268-1287. Daft, R. andLengel, R. (1986). Organizational information requirements, media richness and structural design.Manageme~tScience, 32(5), 554-571. DeSanctis, G. and Gallupe,R. B. (1987). A foundation for the study of group decision support systems. ~ a n a ~ e ~Science, e n t 33(5), 5 8 ~ 9 .

15. DISTRIBUTEDGROUPSUPPORTSYSTEMS

DeSanctis, G. and Poole,M. S. (1991). Understanding the Difference in Collaborative System Use Through Appropriatation Analysis. Proceedings ofthe 24th Hawaii International Conference on System Sciences, 750-757.

DeSanctis, G. and Poole, M. S, (1994). Capturing the complexity in advanced technology use: Adaptive structuration theory.O ~ a n ~ a t i Science, on Vol. 5, No. 2, pp. 121-147. Dickson, G., Partridge, J. L., and Robinson,L. (1993). Exploring Modesof Facilitative Support for GDSS Technology,MIS Q u a ~ e 17(2), r ~ June, 173-194. Dufner, D. K.(1989). Replication of the Wutson Research: ~perimentalDesign and Pre-ResearchTrials Using Five Studenr Groups at~ JUnpublished ~ . research report,NJIT, Newark, NJ, Dufner, D. K.(1995). The Effectsof Croup Support~ ~ i s t i and n g Voting Tools) and Sequential Procedures on Group Decision Making Using Asynchronous Computer Conferences,Doctoral thesis, Rutgers University Graduate School of Management. Dufner, D. K., Hiltz, S. R., and Turoff, M. (1994). Distributed Group Support: A.Preliminary Analysis of the Effects of the Use of Voting Tools and Sequential Procedures. Proceedings of the Hawaii, 114-123. 27th Annual Conference on System Sciences, Dufner, D. K,,Hiltz, S. R, Johnson, K., and Czech, R. M. (1995). Distributed Group Support: The Effects of Voting Tools on GroupPerceptions of Media Richness. Group Decision and NegotiationJournal, Special Issue on Distributed Communicat~on Systems (Vol. 4, No. 3), pp. 235-250.

Easton, A. C., Vogel, D. R., and Nunamaker,J.F. (1989). Stakeholder identif~cationand assump tion surfacing in small groups: an experimental study. Proceedings of the 22nd Hawuii Internutjona~Conference on System Sciences(Vol. 3), pp. 344-352. Ellis, C. A., Gibbs, S. J., and Rein, G. L. (1991). Groupware: Some issues and experiences.Communications o f the ACM, 34(1), 39-58.

Fjermestad, J. (1994). Croup Strategic Decision ~ Environment:AComparison

ain a Computer-Mediated€ommunicatio~ ~ i ~ of Dialectical I n q u i ~andConstructiveConsensus A p p r ~ c h e s .

Unpublished doctoraldissertation, Rutgers University. Fjermestad, J. and Hiltz, S. R. (1996). An assessment of~perimental Studies of Group Decision Support Systems. Manuscript submitted for publication. Fjermestad, J., Hiltz, S. R., and Turoff, M. (1993). An Integrated Theoretical Framework forthe Study of Group Decision Support Systems, H I B '93,W, 179-188. Fjermestad, J., Hiltz, S. R., Turoff, M., Ford, C., Johnson, K., Czech, B., Ferront, F., Ocker, R., and Worrell, M. (1995). Group Strategic Decision Making: Asynchronous GSSUsing Structured Conflict & Consensus Approaches.Proceedings of the 28th Annual ~ a w a i~nternarjona~ i Conference on System Sciences (Vol. IV), pp. 222-231. Los Alamitos, C A IEEE Computer Society Press. of Fiores, F., Graves, M,, Hartfield, B., and Winograd,T. (1988). Computer systems and the design organizational interaction.ACM Transactions onOffce ~nformation Systems, April, 153-172. Guindon, R. (1990). Designing the design process: Exploiting opportunistic thoughts. HumanComputer Interaction, 5, 135-344.

Hall, J.(1971). Decision, Decisions, Decisions, Psychology Today,5, pp. 51-54,86,88. Hiltz, S. R. (1994). The Virtual Classroom: Learning Without Limits Via Computer Networks. Norwood, NJ: Ablex. Hiltz, S, R., Dufner, D. K, Holmes, M. E., and Poole,S. M. (1991). Distributed Group Support Systems: Social Dynamics and Design Dilemmas. Journal of O ~ a n ~ a t i ~ a l C o m 2(1 p )u, 135-159, ti~ Hiltz, S. R., Johnson, K., and Turoff, M. (1991). Group Decision Support: The Effects of Designated Human Leaders and Statistical Feedbackin Computerized Conferences.Journal of Management Information Systems,KZ), 81-108.

Hiltz, S. R. and Turoff,M. (1978). The etw work ati ion: Human Co~municationvia Computer.Reading, M A Addison Wesley Advanced Book Program. Revised editionMIT Press, 1993. Hiltz, S. R. and Turoff, M, (19%). Structuring Computer-Mediated Communication Systemsto avoid Information Overload.CACM, 28(7),682-689.

HILTZ ET AL. Hiltz, S. R. and Turoff,M. (1991). Computer Networking Among Executives:A Case Study. In J. F. Nunam~er,Jr., and R. H. Sprague, Jr. (eds.),Proceedings HICSS, pp. 758-769. House, R.J. (1971). A Path Goal Theory of Leader Effectiveness.Admin~~ative Science quarter^ 16,321-338. Hsu, E. Y. P,, and Hiltz, S. R. (1991). Management Gaming ona Computer Mediated Conferencing System: A Case of Collaborative Learning through Computer Conferencing. In J. F. Nunamaker, Jr., andR. H. Sprague, Jr. (eds,),Proceedings HICSS, Vol.IV, 367-371. Isakowitz, T. (1993). Hypermedia, Information Systems, & Organizations: A Research Agenda, Proceedings of the 26thH i m , Vol. W. Jago, A. G. (1982). Leadership: PerspectivesIn Theory and Research.~ a n ~ e m eScience, nt 28, 315-336. Johnson-Lenz, P. and Johnson-Lenz, T. (1982). Groupware: The process and impacts of design choices. In E. B. Kerr and S. R. Hiltz (eds.), Computer-~ediate~ Communicat~on Systems: Status and valuation (pp. 45-55). New York AcademicPress. Johnson-Lenz, P. and Johnson-Lenz,T.(1991). Post mechanistic groupware primitives: Rhythms, boundaries and containers.Int. J.OF~an-~achine Studies, 34,395-417. King, N. and Anderson, N. (1990). Innovation in working groups. In M. A. West and J. L. Farr (eds.), innovation and CreQtivi~ at Work. Chichester: Wiley, Kim, Y. J.(1996). C~rdinationin ~ i s ~ i b u t eGroup d Support Systems: A Con~olled~ p e r i m e nComt Doctoral dissertation, RutgersUniversity Graduate par@ Parallel and Sequent~al Processes. School of Management. Kunz, W. and Rittel,H. (1970). Issues as elements of inFormation systems (Working paper no.131). Institute of Urban and Regional Development, University of California, Berkeley. MacLean, A., Young, R.,Bellotti, V., and Moran, T. (1991). Questions, options, and criteria:Elements of design space analysis. Human~omputer interac~on: 6(4), 201-250. Mahoney, M. J. (1977). Publication Prejudices:An Experimental Studyof Confirmatory Bias in the Peer Review System.Cognitive ~herapyand Research,Vol. 1, pp. 161-175. Mahoney, M. J. (1978). Publish and Perish.H u ~ a nBehavior,Vol. 7, pp. 38-41. Mahoney, M. J.(1985). Open Exchange and Epistemic Progress. American P s y c h o l ~ ~Vol. t , 40, pp. 29-39. Malone, T. W. and Crowston,K. (1990). What is coordination theory and how can it help design cooperative work systems?C!XW90 ~ e e d i n g s357-370. , Marshall, C. C. and Shipman, F. M. (1993). Searching for the missing link Discovering implicit structure in spatial hypertext. Proceedings ofHypert~t,ACM Press, 212-230. Mason, R. 0.(1969). A dialectical approach to strategic planning. ~anagementScience, 15(8), B403-B414. McGrath, J.E. (1984). Groups: interaction andPe~ormance.Englewood Cliffs, NJ:Prentice- all. McLeod, P. L. and Liker,J.K (1992). Electronic Meeting Systems: Evidence from Low a Structure Research, 3(3), 195-223. En~ronment.inFo~ma~on Sytems Mennecke, B. E.,Hoffer, H. A., and Wynne, B.E. (1992). The Implications of Group Development and History for Group Support System Theory and Practice. Small Group Research,23(4), 525-572. Mitroff, I. I. and Mason, R. 0.(1981). Creating A Dialectical Social Science: Concepts,~ e t h ~and s: ~ ~ e Boston, l s . M A D. Reidel. Nemiroff, P. M,, Pasmore, W. A., and Ford, D. L, (19'76). The Effectsof Two Normative Structural Interventions on Established and Ad Hoc Groups: Implications for Improving Decision MakEffectiveness. Decision Sciences, 7,841-855. Nunamaker, J. F., Dennis, A.R., Valacich, J. S., Vogel, D. R,, and George, J. F. (1991). Electronic meeting systems to support group work. Communjcat~ons of the ACM, 34(7),40-61. Ocker, R. (1995). Computer SupportForD ~ ~ i b u tAsynchronous ed Software Des&nTeams. Doctoral thesis, Rutgers University Graduate Schoalof ~anagement, Ocker, R., Hiltz, S. R., Turoff, M., and Fjermestad, J. (1995). Computer Support for Distributed Software Design Teams: Preliminary Experimental Results. ~ o c e e d i of ~ sthe 28th Annual

15. DIST~BUTED GROUPSUPPORTSYSTEMS

Hawaii Intemational Conference on System Sciences,Vol. W,pp. 4-13. Los Alamitos, C A IEEE Computer Society Press,1995. S. (1991). User-centered designof collaboration technology. Journal of Olson, G. M., and Olson, J. O~an~ational Computin~l, 61-83. Olson, G. M,, Olson,J. S., Carter, M,, and Storrosten,M. (1992). Small group design meetings:An uter 7(4), 347-374. analysis of co~laboration.~ ~ m a n ~ o ~ pInteract~on, Olson, J.S., Olson, G.M,, Storrosten, M., and Carter,M. (1993). Groupwork close up:A comparison of the group design process with and without a simple group editor. ACM Transactions on Oftice Information Systems,21(4), 321-348, Osborne, A. F. (1957). Applied Im~ination(2nd ed.). New York Charles Scribner’s Sons. Peters, D. P. and Ceci, S. J. (1982). Peer-review practices of psychological journals: Thefate of published articles, submitted again.The ~ehauioraland Brain Sciences, 5, 187-255. Poole, M. S. and DeSanctis,G. (1990). Understanding the use of group decisionsupport systems: ns The theory of adaptive structuration. In C. Steinfield and J. Fulk (Eds.), O ~ a n ~ a t i o and communication techno lo^ (pp. 173-193). Beverly Hills: Sage. Poole, M. S. and DeSanctis, G. (1992). Microlevel Structuration in Computer-Supported Group Research, 19(1), 5-49. Decisfon Making.Human Communica~on Futnam, L. L. (1982). Procedural Messages and Small Group Work Climates: A Lag Sequential Analysis. Communication Earbook, 5,331-350. Rana, A. (1995). Peer Review Process and Croup Support Systems: Theory Development and Experimental Validation.Doctoral thesis, Rutgers University Graduate School of Management. Rana, A. R., Hiltz, S. R,, and Turoff, M. (1995). Peer Review Process: A Group Support Systems Approach ork king Paper, NJIT). Newark, NJ. Rice, R. E. (1984). The New Me~ia:Communication, Research andT e c h n o l ~Beverly . Hills, Rice, R. E.(1993). Media appropriateness: Using social presence theoryto compare tradi and new organizational media.Human communication Research,1!?(4), 451-484. Sambamurthy,V. and DeSanctis,G. (1989). An experimental evaluationof DGSS effects on group Proceedings of the 23rd Annual Hawaii Internationperformance during stakeholder analysis. alConferenceon System Sciences,Vol. 3, pp. 79-88. Schweiger, D. M,, Sandberg, W. R.,and Ragan, J. W. (1986). Group Approaches for Improving Strategic DecisionMaking: A Comparative Analysisof Dialectical Inquiry, Devil’s Advocacy, and Consensus.Academy ofMa~agementJournal, 29(1), 51-71. Schweiger, D. M,, Sandberg, W. R., and Rechner, P, L. (1989). Experiential Effects of Dialectical Inquiry, Devil’s Advocacy, and Consensus Approaches to Strategic DecisionMaking. Academy of Manage~ent Journal, 32(4), 745-772, Schwenk, C. R. (1990). Effects of Devil’s Advocacyand Dialectical Inquiry on Decision Making: A Meta Analysis.O ~ a n ~ a t j o n~ehaujor af and ~ u m a nDecision Processes,47, 161-176. Silver, M. S. (1990). Decision Support Systems: Directed and Nondirected Change.~n~ormatioR Systems Research,l( l), 47-70. Simon, H. A. (1973). The structure of illstructured problems.Arti~cial Intell~ence, 4,145-180, New York Oxford UniversiStogdill, R. M. (1959). individual Behavior and Group Achievement. ty Press. Tjosvold, D. (1982). Effects of Approach on Superiors’ Incorporationof Subordinates’ Information in Decision Making.Journal ofApplied psycho^^? 67,189-193, Telem, M. (1988).Information requirements specification brainstormin Processing & Man~ement,24(5), 549-566. making approach, Info~ation Tung, L. and L. Heminger,A. R. (1993). The Effects of Dialectical Inquiry, Devil’s Advocacy, and Consensus Inquiry Methodsin a GSS Environment. Information & Management?25, 33-41. Journal o f O ~ f f R ~ a r j o n Turoff, M. (1991). Computer-Mediated Requirements for Group Support. a1 Computing, I, 85-113. Turoff, M., Hiltz, S. R., Bahgat, A. N. F.,and Rana, A. (1993). Distributed Group Support Systems. MIS Quarterly, December, 399-417.

HILTZ ET AL. Turoff, M., Rao, U,,and Hiltz, S. R (1991). Collaborative Hypertext in Computer Mediated Communications. Proceedings ofthe 24th HICSS, Vol. W, 357-366. Van de Ven, A. H. and Delbecq, A. (1974). The effectiveness of nominal, delphi and interacting Academy of~ a n a g e ~ e~ournal, nt 17,605-621. group decision making processes. Watson, R T. (1987). A Study of Group Decision Support System Use in Three~nd- our Person Groups for a Preference Allocation Decision, Unpublished doctoral dissertation, University of Minnesota. West, M. A. (1990). The social psychology of innovation in groups, In M. A. West and J. L. Farr (eds), Innovation and Creativity at Work.Chichester:Wiley. Worrell, W., Hiltz, S. R., Turoff, M., and Fjermestad,J.(1995). An experiment in collabortive learning using a game and a computer-mediated conferencein accounting games.~ ~ e e ~ofi R ~ s the 28th Annual H ~ ~ aInternational ii Conference on System Sciences, Vol. IV, pp. 63-71. Los Alamitos, C A IEEE Computer Society Press. Zmud, R.,Lind, M,, and Young, F. (1990). An Attribute Space for Organizational~ommunication 1(4), 440-457, Channels. Information Systems Research

C H A P T E R

e e

Indiana University

Unjve~ity of California, lwine

Purdue University

New York University

University of California, lwine

KLING ET AL.

TheAdvancedIntegrated Manufacturin~ Environments(AIME) project has been a multiyear study of coordination changes in US. manufacturing firms implementing new information technologies. AIME project research has confirmed the need for a~ e ~ a ~ i oas r awell Z as an information-pr~essin~ view of how IT changes coordinationin practice. Thereis a long tradition of organizational analysesof information systemsin organizations-how social forcesinfluencetheirselectiveadoption,shapetheirconfigurations, enhance or undermine their implementation, and influence their subsequent uses (e.g.,Kling, 1980, 198’7). Information-processing viewsof coordination change show how structurby simplifying al featuresof IT directly improve organizational performance keycoordinationproblems of scheduling,synchronizing,andallocating (e.g., Malone & Crowston, 1994). The information-processing approach has special appeal because it offers a way to think about optimizing organizational structures to reduce coordination costs. Information-processing formulations, such as Malone and Crowston’s, emphasize siatic, relatively optimal, solutionsto organizational problems.This information-processing view, however, gives us an incomplete understanding of how to copewith ~ y n a ~ ic organizational problems that arise from changing coordination practices within a world of powerful social and economic logics. The AIME project has used behavioral theories from organizational sociology and institutional economics to create anunderstanding of IT as a shifter of potentials and constraints in a world ofexistingeconomicand socialcoordinationprocesses. This chapteridentifiessomekey findings from the project. At the project’s inception, the dominant discourse about coordination and information technology was framedin terms of informationprocessingtheories of coordination.Althoughweanticipatedthat behavioral analyses would add depth to the information-processing analyses, wedid not know how much these alternative approaches would be complementary, conflicting, or synergistic. Many observations in this chapter came from detailed empirical field studies of the useof IT to coordinate manufacturing activities. We did find that IT can sometimes be usedto solvemany existing coordination problems without any substantial side effects. This is especially the case when the problems and technologies are simple and straightforward. But as the coordination problems become more organi~ationally complex and interdependent,so often do the information technologies intended to solveIn them. such instances, itis more accurate to speak of the useof IT as ~ansformin~ one set of coordination problems into another setof coordination problems. The new set of coordination problems may be more or less tractable for the organization. The “irony” ofIT and coordination is that the new kinds of interdependencies created by the sustained use of IT may, in some circumstances, be more difficultto coordinate than the original problems IT use was supposed to

16. PROMISE AND PROBLEMS OF IT IN COO~INATION

address.Much of the difficulty is due to the relative inexperience organizations have in dealing with these new coordination problems. New design techniques and new institutional~rangements for organizational usability have the potential to make these coordination problems much less severe.

Although a distinction is often made between coordination activity and productionactivity in organizations(e.g.,ScottMorton,1991),coordination It is usually defined ata very abstract level, itself is an extremely broad term. as the alignmentof distinct but interdependent activities (Malone& Crowston, 1994). Everything from human communication, to factory scheduling algorithms, to an international currency market can be conceptualizedas a coordination problem.To give more concreteness to the kindsof organizational Coordination issues the AIME project has focused on, we briefly discuss examplesof the difficultyof coordination changes throughIT. One kind of organizational coordination problem arises when the value of asharedinformationsystemdependsonhowdifferentindividualsand groups use the system jointly. For example, managers who have acquired group calendar systems that help their subordinates automatically schedule meeting times,or at least be aware of each others schedules, have faced significantorganizationaldifficulties(Bullen & Bennett, 1996; Grudin,1994). Each person must maintain an accurate, up-to-date personal calendar that publicly defines their appointments and "free time," Maintaining personal calendars on a computer system isa si~nificantamount of work-work done largely for the benefit of other group members, and the clerical staff that schedules meetings. Group calendars-in practice-havea political economy of effort that can make it hard for those who do most of the record keeping to feel that they have gained proportional value (Grudin, 1994; Kling, 1980). There is also a politics to allocating time and having one's time commitments be publicly visible. The men and women who use the system have to agree on the meaning of free time.Can a person have no free time? Is a person allowedto block off time on his or her calendar for anyreason, or only for official company events? The coordination problems of organizational systems such as group calendars are not limited to providing electronic communications and scheduling. Wagner (1996) found one intriguing challenge whenshe tried to designa surgical calendar system fora surgical teams that were composedof (typically) male surgeons and (typically female nurses. She noted that: If . .women's (nurses) voices are not heard, the resulting rules and regulations concerning "privacy versus transparency" will reflect a one-sided tendency to protect primarily the needs and interestsof surgeons, hereby disre-

KLING ET AL.

garding the organizational problemscreated when in an organization characterized by strong dependencies and frequent emergencies one group enjoys a high level of time autonomy (and privacy). (pp.891-892)

A coordination solution, such as using a temporal database-calendars-to

managecommitmentsand to moreefficientlyschedulepeopleand resources can be very appealing, especially when itis abstracted from conit rapidly becomes a crete working conditions and social relationships. But problemofmanagingincentives to keeppersonalcalendarsup-to-date, agreeing on the meaning of free time, giving different workers effective voice in scheduling major events, and facing the local politics of temporal privacy. Another example of organizational coordination difficulties comes from the use of massive, technically complex computer systems that span an entire organization. Whereas a complex system may improve aspects of a firm's coordination, making these systems run smoothly on a daily is abasis hugecoordinationchallenge of itsown. In manufacturing,forexample, RPII (~anufacturingResource Planning) systems have faced significant implementationdifficulties(Hayes, ~heelwright9 & Clark, 1988; Warner, 1987). As different groupsare tightly linked together, the new dependencies betweengroupshave to becoordinated(Attewell,1991; IUing & Iacono, .The technical capabilitiesof the system, and any modifications, have negotiated by all of the groups relying on the system. The organizational complexityof using MW11 for coordination is shownby the fact that internal politics have been a better predictor of the extentof MRPII use than purely technical factors (Cooper& Zmud, 1990). These two examples illustrate the kinds of issues organizations face in translating the potentialof IT into improved organizational coordination. IT, because of itsinherentcapability tostore,process, andtransmitvast amounts of information, has rightly been seen as a powerful enabler of new forms of organizational coordination (e.g., Scott Morton, 1991). However, the specific ways thatIT changes organizational coordination in practice cannot be fully described by inherent technological capabilities such as "reducing timeandspace to zero."Theseexamples areconsistentwithprevious research on computing and organizations, which shows that the use of IT may lead to different coordination outcomes, depending on existing social and economic logics (e.g.,fling, 1996). The actual coordination changes that takeplace in the presence of IT isaquestionthathastobeanswered throu~h empirical, behavioral study. In the AIME project, we have studied coordinationin manufacturing, for both practical and theoretical reasons. ~anufacturingis of undeniable prac' importance to the U.S. economy, and thus is important to S sman, 1987). From a theoretical~ e w ~ o i however, nt9 studyi tion in a domain suchas manufacturingis particularly interesting because of

16. PROMISE AND PROBLEMS OF IT INCOORDINATION

the harsh technical demands of manufacturing coordination, the rich institutionalenvironment that bringsmanydifferentgroupsandactivities together, and the constant experimentation taking place with new forms of coordination.

When individuals and groups specialize, by concentrating their expertisewithin a narrow range of activities,there is a need for coor~ination.Coordination among specialized individuals and groups takes place by ordering and arranging the interdependencies among their separate activities (fling et al., 1992). Manufacturing firms, like all firms, coordinate extensively at many levels.Different groups, such as engineering, marketing, production, and materials each have deep specialized knowledge of their own domains. But their decisions and behavior are frequently interdependent with other activities outside of of these different groups must be their domain. Althoughthe skill and attention focused and choreographed for good organizational performance, conflicts of perspective and practice between any of these groupsis common. To explain this diversityof organizational action,the AIME project used multiple theoretical perspectivesto account for actual coordination behavior in IT-using manufacturing firms. This following sectiondescribes those theoretical perspectives. It begins with the initial theoretical critiques of a view of manufacturing coordinationthat depends too heavily on the inherent attributes of the technology, a view that is widespread in the literature on computer-integrated manufacturing.Next, it reviews the theories from organizational sociologythe project found useful for studying coordinatio~ behavior. Finally, the work on theoretical perspectives from institutional economics is presented as another useful lens for viewingthe role ofIT in manufacturing coordination.

Since the 1980s, discussions of IT’S potential for changing manufacturing coordination have taken place under the banner of Computer-Integrated Manufacturing (CIM). CIM is a strongly information-processing centered vision that emphasizes the need for greater computerization, and greater data integration, arounda single, enterprise-wide database (e.g., Warrington, 1973; Melnyk &C Narasimhan, 1992). The CIM vision argues that the path to more effective coordination isto tightly link togetherseparate areas of the factory through databases andcomputer-mediatedcommunication-integrating the “islands of automation”to achieve globaloptimi~ation.

KLING ET Ai.,.

Since the 199Os, practical problems with the CIM vision have been discussed in the research literature,and the popular press.A common obserof CIM are "more organizationvation in popular articles is that the problem al (or cultural) than technical" (e.g., Sheridan, 1992). Social researchers also note the importanceof organizational issues, and their elusiveness.In writing about less ambitious incremental changes in manufactuirng systems, Shani, Grant, Krishan, and Thompson (1992) note: One result that is abundantly clear is that critical management problemsarise not in the adjustment of the technical system, but in the adjustment of the social system. Notonly are the time frames required for adjustment much longer (for example, in employee training and in gaining the commitment of managers at different levels and in different functions), but the problems of interpersonal relations and organizational structure are far less transparent and much less easyto define than those of technology.(p. 108)

Analysts who have critiquedUS.manufacturing computerization make a distinction between computerizing direct (production) activities versus indirect (coordination) activities. AlthoughUS. manufacturers are seen as not making enough appropriate use of computerization for direct production activities (e.g., Jaikumar, 1986), they have been criticized for excessively computerizing, and overcomplicating, indirect activities such as scheduling, ~roductionplanning,andoductioncontrol(e.g.,Dertouzos,Lester, Solow,1989;Hayeset al., 19 Roven & Pass,1992). The AIME project began its conceptual work by identifying the kinds of anizational issues that were repeatedly being observed in real coordination behavior, but that were not being addressed by the information-processing focus ofCIM. From existingresearch on the impacts of computer techis not nology,the AIME projectwasawarethatorganizationalchange dictated solelyby the inherent capabilities of a new technology. When an IT system becomes sufficiently largein scope (involving numerous groups), it can be seen as a social and economic institution (IUing 8r Iacono,1989) shaped by behavioral as well as technological forces.As a recent National Research Council(1994) study concluded: IT alone does not create impacts; its effects reflect a host of decisions made and actions taken-wisely or not-by a range of stakeholders including senior managers, technical professionals, andusers. (p. 161)

To understand how manufacturing firms actually use IT to facilitate coordination, the first task of the analyst isto define a setof theoretical concepts that can account for changes in coordination technology, but within a framework of human decision, behavior, belief, and history.

16. PROMISE AND PROBLEMS OF IT IN C O O ~ I N A ~ O N

There has been a substantial bodyof research on common organizational problems with computer systems (Gasser, 1986; Knights & Murray, 1994; Laudon, 197% Orlikowski, 1993). Our reading of this literature indicates that the appropriate theories would have to account for: the social relationships betweenparticipantswhoinfluencetheadoptionanduse of computerbasedtechnologies,theinfrastructuresforsupportingsystemsdevelopmentanduse,andthe history of localcomputingdevelopments (Kling, 1987). Special attention would have to be paid to information-processing views of coordination that assume harmony and cooperation, rather than the possibilityof partially conflicting preferences, interests, or values(Kling, 1991; Orlikowski, 1993). Purely technological theories of coordination also tend to overestimate the ability of different subgroups to coordinate quickly and smoothly. The capacity to coordinate can be limited by organizational processes (Beuschel & ing, 1992; Kling, 1992b; Kling, 1993). ~oordinatingmanufacturingoperationsthrough IT raisesimportant, recurring organizational issues. Addressing these issues requires the ofuse social and economic perspectives, that assume actors will behave as groups in a social context, or as economic agents (Kling et al.,1992).

The first set of behavioral perspectives used by the AIME project comes from organizational sociology (Perrow,1986; Pfeffer, 1982; Scott, 1992). Sociological theoriesof coordination assume that groupswill conflict over goals and interests. They identify social bases for group differences and interests, such as status, power, and social identif~cation. Sociological theories are pertinent to understanding IT and coordination in manufacturing because such systems tie together organizational units with different occupational cultures and work practices(Kling et al., 1992). The AIME project used three theoretical perspectives from organizational sociology, each with their own strengths and weaknesses: institutional theory, structural contingency theory, and resource dependency theory. of Each of these perspectives has its own language for describing the nature "alignment" betweenseparate activities.The main concepts from each theoretical perspective, along with key examples from our research, are presented next. More detail on each of these perspectives can be foundin Kling et al.,1992. l ~ s t j t u t j o ~eo^. ~l Institutional theory views organizations as groups of people who embody and enact loosely coupled standardized packages of rules, procedures, and beliefs (Powell & ~iMaggio, 1991). These standardized

KLING ET AL.

packages, or “rationalized myths” (Meyer& Rowan, 1977), are adopted priin the eyesof powerful external marily to maintain organizational legitimacy actors and belief systems.~aintainingorganizational legitimacyin the eyes of outside institutions, such as government regulators, professional organizations, and powerful clients, contributes to the survival of the organization. “taken-for-granted”~rgaOver time, the institutionalized packages become nizational actors can no longer think of legitimate alternatives, and the packages become extremely difficult to change. Our research found institutional theory especially useful for explaining coordination changesin situations where manufacturers face strong external legitimacy demands, and cope with complexof sets technologies thatare sensitivetotheorganizationalassumptionsembeddedwithinthem. We in boththe foundinstitutionalforcesshapingcoordinationoutcomes AIRTECH andDISKCO cases. We conducted a case study at the Wing Control Division(WCD) of AIRTECH, a Southern California aerospace manufacturer. WCD produces sophisticated control equipment for airplanes, helicopters, and missiles that requires theintegration of mechanical,hydraulic,andelectronictechnologies. WCD’s 10-12 product lines are evenly split between commercial and military markets. In terms of market positioning,WCD has a reputation forhightech design skill, and high prices. As one design engineer said, ‘ke’ll win [the contract]on technology if the price doesn’t kickus out.”Our data collection over the 18-month period consisted of three waves of 22 total individual and team interviews.DISKCOis one division of a multinational computer manufacturing company thatmanufacturesdisk drives for mainframes, minicomputers, and workstations. We studied the efforts of DISKCO’s manufacturing engineers and IS specialists to develop an effective CIM system to support a new assembly line that manufacturesl-2GB disk drives for workstations. Oneof the most potent examples of the power of strong external belief systems comes from the AIRTECH case. AIRTECH used a complex computerized scheduling and logistics system to tightly couple many different factory activities. This Manufacturing Resource Planning( M ~ I I system ) used assumptions about how long it takes to build certain parts, how long it takes to move parts between areas, and how many usable parts are output to tightly coordinate activities. One set of mid-career operations managers, freshfromtheirprofessionalseminarson“just-in-time”manufacturing, reducedmanyof these systems assumptions to overly optimistic levelswhat their professional ideology told them should be the case. (For example they reduced the parameters for the times to move materials between work centers to zero.) Short-term schedule improvements turned into long-term chaos,andtheoperationsmanagerswereeventuallydismissed(Allen, Bakos, & Hing, 1994).

16. PRO~ISEAND P R O B L E ~ SOF IT IN COO~lNATlON

AI~TEC offers ~ an extreme exampleof how institutionalized beliefs about ideal forms of managing that are legitimate in managerial worlds can cause coordination difficulty, Anotherkind of institutional problem is seen in the DISKCO case. A shop floor control system designed to be used by skilled workers who needed little monitoring caused new coordination problems for another DISKCO division that adopted the system (Allen,Kling, & Elliott, 1994). Complex computer systems for manufacturing coordination tightly link groups together. These tight linkages contributeto inertia (Beuschel& Kling, 1992), somewhat reduce local experim~ntation (Allen, 1992), and create a new set of horizontal coordination and control needs thatis open to institutional clash (Allen, 1994b). These horizontal linkages are complicated by the multiple institutional forces that ~anufacturershave to answer to, or try to take advantage of, simultaneously. ~anufacturersanswer to multiplesets of government regulators for labor issues, safety issues, and business law; multiple customers, each with their own requirements for quality and product flexibility; and multiple professional organizations, each with their own growing body of dogma. We found these partially competing logics permeating our two major case sites (Allen,Bakos, & Kling, 1994; Allen, Kling, C% Elliott, 1994). Also, as organizations face more intense time and cost pressures, they tend to make more useof available institutionalized packages such as temporary workers, or standardized software (Allen,Kling, & Elliott, 1994). This process leadsto a new institutional challenge: tryingto effectively fit together fixed packages of organizational procedures from the outside world. In our research, we also found institutional theory useful for describing how organizational actors initially choose a legitimate new formof coordination. Of the many alternatives, organizational actors tend to repeatedly select from those few options that are publicly visible, and seen as legitimate. Teams are chosen as a popular coordination mechanism, even if their primarypurposeisforcostreductionratherthaninvolvement(Allen, Bakos, & Kling, 1994). Groups choose to coordinate others through a cornputer system, rather than engagingin a process of organizational redesign (Allen,1992;Allen,Bakos, & Kling, 1994). ~nderstandingtheprocess of choosing coordination methodsis an important part of accounting for IT'S role in manufacturing coordination. In sum, institutionaltheory allowed usto account for how external belief systems could become a significant force in coordination change, the tensions between different external logics that permeated the organizations, andtheparticularmechanismsused tochoose newcoordinationtechniques. Institutional theories expanded our ability to account for coordination issues that lie outside organizational boundaries. Institutional theories also gave us a wayof understanding how different groupsrely on different unifying logic. logics of coordination, rather than sharing a common

KLING ET AL.

onting~ncyT ~ e o ~Structural . contingency theoryviews organizationsas bureaucracies designed to complete tasks. The structure of anyparticularbureaucracyisdeterminedlargely by theuncertainty involved in their formal tasks (Calbraith,1977). The greater the task uncertainty, the greater the amount of information processing required. Each kind of coordinationstructure,fromstandardizedrulestocross-functional teams, can cope witha different levelof uncertainty at a specific cost to the organization. Structural contingency theory is the behavioral perspective most closely associated with an information-processing view of coordination. In our research, we found structural contingencytheory most useful for characterizing the internal technical needs of manufacturers for coordination. Although institutionaltheory provided a better explanationof why particular coordination methods were selected and usedin the AIRTECH case (Allen,Bakos, tk Kling, 1994),structuralcontingencytheorywasableto why certain coordination choices were able to explain the technical reasons persist or perish. Structural contingency theory provides the vocabularytask uncertainty, interdependence, and complexity-for discussing generic technical needs for coordination (Allen, Bakos, tk Kling, 1994). It explains the most whenthere are strong technical demands on interdependent tasks, but (perhaps surprisingly) many manufacturing activities do not have strong in the technicaldemands.Withitsemphasisoninformationprocessing organization, however, structural contingency theory views can lead to an overemphasis on coordination as formal information exchange, instead of the interconnectionof distinct groups (Beuscheltk Kling, 1992). en~encyT ~ e o ~Finally, . resourcedependency theory holds that organizations obtain resources from their environments for survival (Pfeffer, 1982). According to resource dependency theory, organizations respond most readily to the demands of outside organizations that control critical resources. Groups within organizations who manage relations with powerful external organizations gain internal influence. Organizations strive to increase their autonomy relative to powerful organizations in their environment, and organizational subunits seek autonomy from each other. Resource dependency theory, despite its early promise, was not used much by AIME research. AIME researchers did see some early examples of manufacturing coordination where resource dependency would appear to be an issue. The question of standardized computer systems is a resource dependency issue, both within and between firms. For example, one powerful customer demands the use of a standard CAD package, while another customerdemandsanentirelydifferentpackagewithdifferentsystems needs. Internal groups can be reluctant to become dependent ona shared,

16. PROMISE AND P R O ~ L EOF ~ SIT IN COO~INATION

centralized system under the control of other groups. Despite this seemingly natural desire for autonomy, what remains to be explained is the incredible extent to which computerized coordination is creating new interdependencies between groups in manufacturing (Allen, 1994a), and the extent to which separate groups have been receptive to this linkage (Allen, 1994b). Autonomy-seeking does not seemto be a powerful explanatory toolacross our multiple cases.

Theoreticalperspectivesfrominstitutionaleconomicsformstheother, complementary set of theories for AIME project research. Economic theories examineoptimalways toallocateresourcesunderuncertaintyand under the assumption that actors are individual utility maximizers. Institutional economic theories seek to identify effective toways coordination and governgroups ofeconomicagents in their transactions with each other (Kling et al.,1992). Economic perspectives generally assume that agents behave opportunistically and rationally in theirowninterests.Organizations,markets,and institutions provide incentive and enforcement mechanis~s for governance. The choice among governance mechanisms, as well as their structureand effectiveness, are dependent on the costs of the underlying processes. To the extent thatIT affects the governance processesin organization, instituIT influences changesin orgational economics perspectives can show how nizational structure and performance. GurbaxaniandShi(1992)developedacomprehensive theory of the impact of IT on coordination in manufacturing organizations. This theory, based on the institutional economics perspectives of agency theory and transaction cost economics, provides set a of hypotheses aboutthe impact of advanced manufacturing information technologies on coordination, and its resulting influence on organizational structure, processes, and performance. In the Gurbaxani and Shi framework, manufacturing firms strive to select the incentive and governance structures that maximize economic returns. ~anufacturingcosts are deter~inedby the sum of internal coordination costs, external coordination costs, and operations costs. Internal coordination costs are the ~o~bination costsincurreddue to goaldifferences between economic principals and the agents they hire (agency costs), and the costs of making decisions with less than perfect information (decision information costs), Accordingto agency theory, decision rightsin an organization should be located where the total internal coordination costs are

KLING ET AL.

minimized. External coordination costs are thesum of costs associated with establishing and maintaining contractual relationships with other parties (contractual costs), and the costs resulting from losses of operational efficiency (operational costs). Operations costs refer to all other noncoordination costs, such as production. The Gurbaxani and Shi framework predicts the following organizational outcomes from the use of IT for coordination. Coordination throughIT will lead to the use of more performanc~basedcompensation schemes. IT use will also lead to a flattening of organizational hierarchy. The reduction in internal coordinationcosts leads to a larger firm size, especially horizontal firm size. Effects on the location of decision rights are more complicated, because IT-based coordination tends to reduce both agency and decision information costs, potentially leading to increased decentralization or centralization, depending onother contextual factors. Changesin vertical firm size are also contextually dependent, because IT-based coordination leads to a simultaneous reductionin internal coordinationcosts, and the external coordination costs of using market-based mechanisms. In practice, however,IT use has been broadly correlated with a decrease in firm size(Brynjolfsson,Malone,Gurbaxani, & Kanbil,in press).The decrease in firm size across a variety of measures-number of employees, revenues, and value-added per firm-suggests a decrease in both horizontal and vertical firm size. Accordingto the Gurbaxani and Shi framework, this appears to make reductions in external coordinationcosts a more powerful explanation of economy-wide changesin firm size than reductions in internal coordination costs. However, as ~rynjolfsson etal. argue, even if both internal and external coordination costs decrease relative to production costs, firms will favor the use of external marketsto coordinate rather than their own internal hierarchies. MME project research has developed new institutional economytheory to better explain the complexities of reduced coordinationcosts. For examto increase ple, a reduction in coordination costs due to IT should lead firms the number of suppliers they use. Althoughthere is evidence of increases in outsourcin~,we find that leading firms in many industries are using fewer suppliers (Bakos & Brynjolfsson, 1993a). The keyto understanding this anomaly is to add the problem of incentive to a theory of coordination costs, particularly the supplier’s incentive to invest in activities whichimprovequality(Bakos & Brynjolfsson, 1993a, 1993b, 1993~).By decreasing the number of suppliers, the buyer makes the relationship more permanentby making it more difficultto switch to alternative suppliers. The stability in the relations~ipgives the supplier the i~centiveto make ‘~oncontractible”investments in quality, or investments that are difficult to specify andverifyin contracts. Because the use of IT has increased the importance of product quality, thistheory predicts that many

16. PROMISE AND PROBLEMS OF IT IN C O O ~ I N A T I O N

firms will use fewer suppliers even when search and coordination costsare low. This same analysisof the new importance of reward and incentive can be applied within the firm as well, to the relationship between managers and operators. The initial survey research indicates that the extremely dynamic computer disk drive industry has reduced the relative importance of individual production quantities, and has begun to reward behaviors which are difficult to quantify. As jobdesignmovesfromindividual operators to a moreteamandprocess-centeredmodel,firms are mostinterested in rewarding skill acquisition and retention. This is despite the fact that C1 technologies are providinggreaterinformationthan ever before on the details of the production process. Firms in these dynamic industries are more likely to make noncontractible investments in their labor force, such as training, and treat operators as less of a commodity.

AlME project research has used the theoretical research dkscribed above to guide its empirical investigations of how manufacturers useIT to coordinate in practice. The empirical research, which integrates the results from the case studies and the pilot surveys, documents both the potential for IT to

solve existing coordination problems, and to create new coordination problems that are sometimes easier, sometimes harder to solve.

There is widespread agreement thatIT has the potential tosolve im~ortant coordinationproblems.Theeconomy-widereductions in U.S. firmsize (Brynjolfsson et al., in press) suggest that IT use is already enabling new ways of organizing that emphasize the use of network and market based coordination mechanisms. The ways that IT is used to solve coordination problemsin m~ufacturing areillustrated by the AIME project’slong-termstudyof AIRTECH. In the AIRTECH case, we sought to explain the adoption of new coordination practices that fall under the label “World Class ~anufacturing” (Allen, Bakos, & Kling, 1994). Manyof these world class manufacturing techniques were being proposed as solutions to the problemsof poor coordination in US.manufacturing, which had become a leading explanation of poor US. manufacturing performancein the 1980s. The ideal descriptions of these techniques, from justin-time inventory control to concurrent engineering, suggest that they increase the qualityof coordination between value-adding production activities.

KLING ET AL.

The case study looked at the adoption of three coordination reformsleadtimereductionthrough MW11 systems, core competencies through manufacturing cells, and cross-functional teaming. Each of these coordination reforms solved key coordination problems for the organization (Allen, Bakos, & Kling, 1994). The MW11 computer system allowed operations management to quickly modifythe assumptions built into the scheduling model across the entire organization, The manufacturing cells allowedAIRTECH to easily identify parts that fell outside of their core competencies, and thus could be outsourced.And the cross-functional teams enabled the early and continuous involvementof many functional areas in group projects. Coordination reforms that help the organization allocate and schedule, identify key products and processes for further improvement, or enhance communications are likely to solve some existing coordination problem. Clearly, information technology has the ability to contribute to all threeof these possible kinds of solutions. In another extended case studyof a disk drive manufacturer,DISKCO, IT was able to help the organization coordinate in a different way, DISKCO required substantial improvementsin both its production and coordination capabilities because of severe new market demands. The lifetime of products in their industry was being reduced from 5-7 years to 12-18 months. DISKCO’s existing means of coordinating design, production, sales and were not intendedto deal with thiskind of time pressure, and a solution had to be found quickly, under severe resource constraints and a shrinking profit margin. Much like a design engineer might turn to an industry standard part, rather than custom designing in-house, when istime short and costs must be low, DISKCO changed its policy to buy as much standardized software and automated tooling from the outside world, rather than designing them inhouse(Allen, Kling, & Elliott,1994).Thesestandardsolutionsallowed DISKCO to resolve an important, recurring coordination problem-ho~ to bring up a new assembly linein a fraction of the time and costof its traditional methods, involving many different functional areas and activities.

Each of the preceding cases revealed a number of fundamental strategies for coordination improvement, the possible roleof IT in those improvements, gnd the actual useof IT for such improvementsin selected firms. Although IT use solves some coordination problems, it also raises other new coordinationproblems.Thesenewcoordinationproblems are sometimesless important than the ones they help solve. However, sometimes theyare so new or unconventionalthattheymaketheoriginalcoordinationreform unsustainable (Allen, Bakos,& Kling, 1994). What makes some of these new coordination problems especially difficult is that are theyoften new typesof

16. PROMISE AND P R O ~ L E OF ~ SIT IN COO~INATION

I

problems, which conventionally organized manufacturing firms have little experience coping with. Although specific coordination reforms are often described as changing the amountor quality of coordination, it may be more useful to view as them transformations, or transforming the set of coordination problems facedby the organization (Allen, Bakos, & ISling, 1994). Particularly as coordination problems become more complex and interdependent, transformations from in a new set of issues one setof coordination problems to another may result that may be less tractable for the organization. In the work on supplier relationships described earlier, the ability of IT use to reduce the coordination problems of outsourcing brings an entirely new kind of coordination dilemma to the surface. Howcanwe ensure that suppliers will participate in necessary, but difficult to verify, mutually beneficial investments? (Bakos & Brynjolfsson, 1993a). The problem of coordinating incentives to invest appears to be more challenging than the old coordination problems of finding suppliers in directories and paying the bills,if for no other reason than firms have less experience with managing this problem. In the AIRTECH case, eachof the coordination reforms created its own new set of coordination problems.In the case of cross-functional teaming, career paths and job performance evaluation were more problematic because both the cross-functional team and the traditional functional area were involved, When conflictsarose between different cross-functional teams, there was not a clear hierarchy for resolving disputes. However, these new problems were less critical than the gains from early cross-functional communication. In the case of the MW11 scheduling system,AIRTECH faced the difficult taskof how to coordinate belief systems. The belief systems of operations management worked to unilaterally change the assumptions built into the factory scheduling model, with ultimately disastrous results (Allen, Bakos, & ISling, 1994). Operations management, fresh from a just-in-time seminar, decidedto reduce the move and queue times in the scheduling model to zero, in accordance with whattheysaw as goodjust-in-timepractice.Instead ofencouraging reducedcycletimes,thechanges to themodel,withoutcorresponding changes in shop floor practice, made scheduling priorities even more unstable. AIRTECH fell further and further behind schedule. It was clear thatAI~TECHhad no developed meansof discussing or challenging these assumptions-they could only wait until the daily production situation deteriorated to the point where new operations management people were brought in. Despite the coordination gains of instantaneous, uniformupdatingofschedulesandschedulingassumptions, AIRTECH was much less experienced with the problemof reconciling strong world views, and the effort to reduce lead time by modifyingthe MW11 systems was scaled back considerably.

KLING ET AL.

TheDISKCO case illustrates a new setof coordination dilemmas created

by what we refer to as the of~t~e-s~elf organization (Allen, Kling, & Elliott,

.

1994). When the technical demands of a market or key customers increase dramatically,organizationsturntopreexistingpieces of institutionalized practices (such as temporary workers) and technologies (such as standardized software packages). Theset of coordination problemsin this case shifts in emphasis fram coordinating traditional production and administration activities to coordinating the combining and fittingof standardized organizational parts, imported from the outside world, that may have partially competing logics (Allen,Kling, & Elliott, 1994). At DISKCO, both temporary workers and the purchase of sophisticated automated tooling from the outside world were pursued as strategies for ameliorating tough coordination problems, yet their presence together created predictable tensions. A new software package, brought in from another division, made the skills developed on an internally developed system obsolete. DISKCO employeeshadspentyearslearninghowtodoad-hocdata queries with their homegrown system. The new system, however, used an industry-standard database that DISKCO workers were unfamiliar with. The lackofprogrammersand users withtheskillstousethenewsystem reducedtheiraccesstoimportantproductiondata.Although DISKCO reaped tremendous coordination gains from using these standardized organizational pieces, by reducing the time to bring new products to market, they are still inexperienced with the new problemsof coordinating the different pieces. Other new coordination problems associated with IT relate to the infrastructure and skills required to make technology-centered visionsof coordination work smoothly. Although a vision such as Computer-Integrated anufacturingmightemphasizecoordinationthroughcross-functional database linkage, this linkage requires significant amounts of resources and attention that often are not planned for, talked about, or sometimes even possible in an era where all organizational activities labeled sas~ p p are o~~ being cut frommanufacturin~ budgets.A simple viewof computerized coordination as a lower cost replacement for other organizational means of Coordination, asin the exampleof a configuration management committee in the AI~TECHcase, unrealistically discounts the amount of continuing human effort needed to coordinate(Beuschel&Kling, 1992). What kinds of coordination are viabledependsontheexistinginstitutionalizedsocial arrangements? An increasingly centralized vision of IT-enabled coordination also contributes to inertia and reduced organizational experimentation, when more functional areas have to approve of all changes (Allen, 1992). Changing a particular technology is much faster than changing the - skill and experience base that makes a technology useful for the manufacturing organization.

16. PROMISE AND PROBLEMS OF IT IN COO~INATION

Some of the new coordination dilemmas posedby IT, such asa society-wide lack of key technical skills or an increase in quality awareness, cannot be solved by the individual manufacturing organization. However, other coordination dilemmas can be anticipated, given an understanding of the organization’s history and configuration. Some of IT’S coordination dilemmas can be partially managed by involving key organizational actors in a joint process of organizational and technological design, informedby behavioral theory. One method the AIME project has explored to address these problems is “design for organizational usability.” Systems usabiZi~refers tohow well people can actually exploit a computer system’sint~ndedfuncti~nality.Usability can characterize any aspect of the ways that people interact witha system, even its installation and maintenance. Thereare two aspectsof IT usability: interface and organizational, Interface usability is centered around an individual’s effective adaptation to a user interface, whereas organizational usability is concerned with how computer systems oan be effectively integrated into work practicesof specific organizations. Although while the Human-Computer-Interaction (HCI) research community has helped pioneer design principles to improve interface usability, organizational usability is less well understood. Design for o ~ a n ~ a ~ usabiZi~ o n a ~ is a new term that refers to a process of of designing computer systemsso that organizational usability is the key focus design (Elliott& Kling, 1997; IUing & Elliott, 1994). It includes, but goes beyond, the focus on user interfaces which is the subjectof “design for usability” as currently understood in the HCI community. Design for organizational usability includes designing the infrast~ctureof computing resources that are necessary for supporting and helping people learn to effectively use systems. It encourages system designers either to accommodate to people’sofmix skills, work practices, and resources, or try to to systematica~~y them. alter Design for organizational usability can be applied to the selection and integration of existing computer systems, or to the design of new systems, to improve the likelihood that people will use them effectively. Coordination issues within an organization’s various departments are considered when designing for organizational usability, including: the design of the infrastructure of computing resources thatare needed to support and coordinate various groups of users; the appropriate “fit”of computer systems into workers’ mix of skills, work practices and resources; and the compatibility of data linkages and architectures between groups within an organization. The motivation for design for organizational usability comes from the technical and organizational complexity of manufactu~ingfirmswehave of computing sysobserved in our AIME project case studies. The collection

KLING ET AL.

tems in a medium to large-scale manufacturing firm are likely to be complex, both individually and as linked together. It is common for such systems to be ineffectively used by an organizationin a way that does not realize the system’s full potential. For example, CIM software may include an end-user database reporting package, butif it physical locationin an organization is inaccessible to most employees, then many people who might benefit from this reporting facilityare unable to doso. Reasons for ineffective use include poor user interface design; lack of adequate training; missing or unnecessary functionality; and/or a lack of coordination of systems usageby varying groups within an organization such as marketing, engineering, information systems (IS), manufacturing, distribution, and sales. The first two reasons In contrast, for ineffective useare examples of traditional interface usability. the second two are concerned with organizational usability. They involve training, and the facilitation of effective systems use in real working environments. The techniques associated with design for organizational usability are described in more detailin Elliott, Kling, and Allen,1994.

AIME project research has identified many specific kinds of coordination changes in manufacturing firms usingIT, and has exploreda setof econom-

In this section, we ic and sociological concepts for explaining these changes. summarize two main themes of our research. First, the explanations we have found most useful regarding how changes in coordination are actually taking place, and the role of IT in those changes. Second, describing the balance of new coordination opportunities and new coordination problems thatare commonly found in IT-using organizations. emands l n ~ ~ e n c e

The starting pointof AIME research was that information technology could play an important role in changing coordination, a role that needs to be investigated and understood. However, changes in coordination behavior are heavily dependent on existing featuresof the organization, and its environment.Inherenttechnologicalcapabilities,we find, are selectively invoked and maintained by social and economic logicsin the organization. Our research resultsare consistent with the claim that behavioral theories of coordination activityare needed to cope adequately with the new organizational challengesof coordination change.

16. PROMISE AND PROBLEMS OF IT IN COO~INATION

AIME project research supports the contention that theories which take seriously the open systems nature of organizations (Kling & Jewett, 1994; Scott, 1992) are indispensable for describing changesin coordination. Both

the institutional economics and organizational sociology perspectives foregroundthedilemmas of coordinatingmultiplestreams of activity,performedbyindividualsandgroupswithconflictingpreferences.Despite theirdifferences,institutionaleconomicsandorganizationalsociology share a fundamentally human concern with incentive and payoff, obligation and reciprocity. Of all the open systems perspectives described, which ones best answer the questionof how coordination changes? The answer depends,of course, on which question you most want to answer. Each of the theoretical perof spectives usedby the AIME project focuses on onlya few significant parts the larger coordination picture. The institutional economics perspectives used here are particularly appropriate for questionsof incentive, monitoring, and contract enforcement. The organizational sociology perspectives are more useful for questions of group belief and power struggles. The key observation here is that “coordination” is defined so broadly that manydifferent questions can be asked. Even questions thatare largely unrelated. The fundamental assumptions of eachtype of theory, however, define the limits of their practical usefulness.If a situation can be adequately described by utility maximizing individuals pursuing defined costs and be.nefits, economic perspectives have insight. If group phenomena, or group membership, is important, a sociological theory opens the possibility forkind thatof analysis. These theoretical assumptions also suggest limits in practical use. In the AIRTECH case, the focus on taken-for-granted beliefsin institutional theory was more useful for describing the process of group selection and consensus arounda particular reform than the more task-minded structural contingencytheory;Structuralcontingency theorybetterexplainedthe recurring technical barriers to the sustainability of some of the new coordination reforms (Allen, Bakos,& Kling, 1994). This observation leads us to the most important tool we have discovered for evaluating the relative explanatory power of these theories: the nature of the environmental demands. Scott and Meyer (1991) defined two different dimensions to environmental demands: technical demands, and institutional demands (Allen, Bakos, & Kling, 1994). All organizations face technical and institutionaldemandsfromtheirenvironments,although to varying organizationsare rewarded for effective degrees. In technical en~ronments, and efficient control of their production systems as their products or services are exchanged in a market. In institutional environments, organizations must conformto an elaborate setof rules and requirementsif they are to receive support andlegitimacy.lnstitutionalrequirementsmaycome from regulatory agencies, professionalor trade associations, or from gener-

KLING ET AL.

a1 belief systems held by society. A computer chip manufacturer in a commodity market may face only strong technical demands. A public school may faceonlystronginstitutional(regulatoryandformaleducation) demands. A bank in a highly competitive market may face both strong technicaldemands(fromcustomers)andstronginstitutionaldemands(from government regulators). To the extent that organizations face strong technical demands, rational perspectives on organizations (such as agency theory, transaction cost theory, and structuralcontingency theory) willhave the mostexplanatory value (Scott,1992). To the extent that organizations face strong institutional demands, natural perspectives on organizations (such as institutional theory, and resource dependencytheory) will have the most explanatory value. This hypothesis is consistent with the results reportedin Allen, Bakos, and Kli

uring firms are typically seen as having strong technical environments, and relatively weak institutional requirements. Thus, rational theories should have the most explanatory power. AIME research suggests, however, that many manufacturers, particularly high tech manufacturersin industries such as aerospace and healthcare, alsoface very strong institutional demands (Allen, Bakos, & Kling, 1994). These surprisingly strong institutional demands, whichare increasing, imply that accounting for coordination changes will increasingly require natural, as well as rational, perspectives on organizations, Evenin manufacturing industries with only strong technical demands, however, we have found that large increases in the severityof technical demands force organizationsto turn to institutional~zed coordination methods from the outside world, in a process that is best understood through natural perspectives (Allen,Kling, & Elliott, 1994). The special role of IT in these coordination choices is best understood in terms of how they tend to shift key parameters in the existing economic and social logics of the situation.In the Gurbaxani and Shi (1992) framework,IT plays arole in shifting internal and external coordination costs. Some of the impacts will likely be unidirectional, such as the increase in output-based compensation schemes. Others, such as the relative decreasein internal vs. external coordination costs, are more dependent on the particular set of choices, previous commitments, and features of the environmentin any particular situation.In institutional theory, therole of IT as an embodiment of a particular belief system is essential in describing its coordination impacts. Institutional theory highlights the importanceof the inherent attributes of a technology, but in this case it is the abilityof IT to embody a particularset of values, and a definition of reality, rather than a generic ability to store, process, and transmit more information (Allen, Bakos, & Kling, 1994). The role of IT in coordination change is a tendency to shift key parameters in important preexisting behavioral logics,

16. PROMISE AND PROBLEMSOF IT IN ~ O O ~ I N A T I O N

In evaluating the costs and benefits ofusing IT to change coordination,

researchers have understandably emphasized the obvious potential benefits. Many problems of coordination can be framed in terms of forma1 information exchange, and formal information processing: scheduling, communication, and simulation are a few examples. We have seen in our cases a significant numberof opportunities for improving coordination through the application of IT’S increased storage, processing, and networking power. Information technologyby its very nature, however,is a technology that opens up significant new coordination challenges, Unfortunately, these costs are often harder to see for those analyzing the problem than the benefits. They are also difficult to manage, because manufacturers typically have less experience with these coordination problems. The practical challengeof IT is to ensure that itsolves more important coordination problems than it creates. The coordination challenges most often mentioned in the context of IT use have to do with issues of infrastructure and skill(Kling, 1987,1992a). The use of IT, particularly complex setsof multiple ITS joined together, requires a massive infrastructure of support and services that must be coordinated and maintained. Computer hardware and software costs are only a small fraction of the total “costs”of keeping IT running smoothly, andare increasing. Complex IT tends to demand new skills, both conceptual and technical, that are difficult to acquire and maintain. Because these coordination activities are often seenas “indirect,” or “support” activities, they are particularly difficult to maintain and coordinate. ~anufacturing is no exceptionto this. Indeed, the hostility toward support activities is probably even more intense than in other economic sectors. Perhaps a more important new coordination challenge found in AI project research is the problemof coordinating ‘”worldviews.” Coordination takes place between different groups, and individuals (Beuschel & Kling, 1992). IT has the potentialto embed particular organizationalvalues, both in terms of the resources and skills it requires to be maintained, in the and very design assumptions used inthe definitionof data models, access rights, and data policies. Through the viewpoint of ~nstitutional theory,many groups tightly coordinated through IT must also align the wor~ingassumptions built into the system. Through agency theory, the emphasis on tighter output mon~toring begs the q~estionof what exactly should bemonitor~d, and how, since p c j ~ p khave a strong tendency to work to what is ~easured, rather than what ~nanag~me~t intends (e.g., Grant,~iggins,& Irvi The coordinati~nof assumptions is aparticularlydifficultcoo p r ~ ~ ) l ~ ~it explicitly ~ e c focuses ~ u s eon differencesin purpose. When com-

KLING ET AL.

bined with a dependence on distant technical personnel, and a tight interlinkagewith other groups that makes agreement on change difficult, the strong embedding of organizational assumptions makes for a particularly troublesomenewcategory of coordinationproblems.The AIME project techniques of “design for organizational usability’’ are intended to address some of these new coordination dilemmas (Elliott&C Kling, 1997; Elliott et al., 1994).

The AIME project has engagedin a multiyear research study of the roleof IT in manufacturing coordination. The role of IT is best seen as introducing powerful new capabilitiesand constraints to an existing world of strong economicandsociallogics.Understandingthe organi~ationalchallenges of these changes is possible with the ofuse behavioral theories from the social sciences. Observers are often enthusiastic about the tremendous potential of IT to change manufacturing coordination for the better. This enthusiasm for new technologica~ capabilities is understandable, but history suggests that the design, imp~ementation, use, and impact of IT is shaped in important ways by established patternsof institutional behavior.AIME project research has investigated this process through the useof theoretical perspectivesfrom institutional economics and organ~zational sociology, The AIME project researchon the role of IT in manufacturing coordination has been exploratory. However, the results from our multiyear study have been strongly consistent with the following claims: 1. Efforts to improve coordination through ITneed for a beha~ioralas well as an information-pr~essing view ofhow changes coordination in practice.

Information-processing viewsof coordination change show how inherent attributes of IT directly improve organizational performance by solving key coordination problems of scheduling, synchronizing, and allocating.Thisinformation-processingview,however,givesusan incomplete understanding ofhow to cope with the chronic organizational problems involved in changing coordination practice within a world of powerful social and economic logics.

2. The explanato~value of differentbehavioraltheories of coordination

depends on the natureof the environmental demandsfaced by an o ~ ~ a n i zation. Rational systems perspectives on organizations, such as agency

theory and structural contingency theory, explain coordination behavior in the face of strong technical demands. Natural systems perspectives on coordination, such as institutional theory, explain coordina-

16. PROMISE AND P R O ~OF~IT~IN S COO~INATION

tion behaviorin the faceof strong institutional demands. We find institutional demandsto be surprisingly strong in manufacturing firms.

3. IT can be used to solve many existing coordination problems without any s~bstantialside effects. This is especially the case when the problems

and technologiesare simple and straightforward. For example, IT has played a useful role in communicating production schedules, crossfunctional team communications, andthe abilityto link organizational groupstogethertoreducenewproductandprocessintroduction to dramaticaltimes. ~tandardizedIT systems also allow or~anizations ly change production processes quickly, at a low up-front cost. 4. As c~rdinationproblems become more complex and interdepende~t,so often do theinformationtechnologiesintended to solvethem. In such instances, it is more accurate to speakof the use of IT as transforming oneset of coordinationproblemsintoanotherset of coordination problems. The newset of coordination problems may be more or less tractable for the organization, The “irony”of IT and coordination is that the new kindsof interdependencies createdby the sustained use of IT may, in some circumstances, be more difficult to coordinate than the original problemsIT use was supposedto address. 5. The implementation of IT-based coordination technologies is easier hen the new c~rdinationproblems do not face strong institu~onaldemands.

IT-based coordination can bringto the surface the difficult problem of coordinating different ‘tivorldviews” and incentives, both within the organization and from outside professional and regulatory bodies. For example, centralized databases and close IT-mediated sup~lierrelationships reveal differencesin fundamental assumptions that have to be coordinated. Muchof the difficulty is due to the relative inexperience organizations havein deaiing with these new coordination problems. New design techniques and new institutional arrangements for organizational usability have the potential to make these coordinat~on problems much less severe. 6. The theo~eti~al ~ e ~ i b of ilI iT~makes it especially at~activewhenthe ~ ~ dynamism of changin~o ~ a n ~ a t i o npractices al can b e n e ~ from t changes in information formuts and information fiows.ow ever, the actu-

i

al implementationof 1T-based coordination technologies locks in many specific design choices which can require substantial skilled labor time to renovate. Thus, IT use for coordination is more smooth when the formal, technical demands of production are clear, and do not create fundamentally new kinds of social and economic interdependencies. Even in these cases, however, a lackof infrastructural resources for skill-buil~ingand support can, and do, hamper technical implementation, given the severe resource constraints manufacturers are facing.

c

KLING ET AL.

Changing a particular technology is much faster than changing the skill and experience basethat makes a technology useful forthe manufacturing organization. 7. The use o f I ~ f ocoordination r is simplest in stableen~ironments,but much more ~hallenging in fast-movingindus~ies.The more rapidly information changes, the more that manufacturers turn to IT as a coordination solution. However, the time and discipline required for computerization and automation conflictswith the need for short time to market and fre~uentproduct changes. Firms in fast-moving industries,such as disk drive manufacturing,are focusing on non-ITchanges to cope with the pace of the industry. Specifically,they are changing their reward and incentive systems to encourage skill acquisition, retention, and effective use for continuous improvement. Despite the increase in detailed process information, reward systemsare increasingly concerned with encouraging behaviorsthat are difficult to quantify.

E project research has begun the important work of identifying and ~tudyingthese practical opportunities, and problems. These results have the follo~ingimplications formanufacturin~practice: 1. ~fforts to implement IT for coordination are more likely to succeed if they consi~ersocial and economic aspects. What must be carefully consid-

is the extent to which proposed information technology fits the nization’s coordination problems, solvesthose problems, and/or creates new coordination problems. . T ~less e the degreeof change r e ~ ~ ibyr the e ~i~~lementation of new inforati ion technolo~ies,the greater the l i k e l i h o ~ successful of i~ple~entation of the t e c ~ n o l o ~ .

s ~ i aand l instit~tionalS n tot ~ e s e c ~ a nChief g~s.

16.PROMISE AND PROBLEMS OF IT IN COORDINATION

Research for this chapter was partially supported by NSF grant I R I 9015 497. We alsobenefitedsignificantlyfromourdiscussionsaboutcoordinatiQn early in the project with our colleague Professor JohnKing, and with later discussions about manufacturing with Dr. Werner Beuschel. We also appreciate the commentsof reviewers, includin~Gary Olson. Thisproject, as well as NSF's ~oordinationTheory focus, was enabledby Dr. Larry Rosenber

Alien, J. P.(1992). Enabling Participatory Design in a Hierarchical, TightlyIntegrated Setting.In PDC'92: ~ ~ e e d j n of g sthe ParticjpatoryDesign Conference, M. J. Muller, S. Kuhn, & J. A. Meskill (Eds,). Cambridge, M A Computer Professionals for Social Responsibility. Allen, J. P.(1994a). Aduanced Manufacturing Technolo~/ComputerIntegrated ~anufacturjng of Organ~ationalImpacts, Working Paper. Irvine, C A Center for ~A~/CIM AReview ~: Research on Information Technology and Organizations, University of California. Allen, J. P. (1994b). Mutual Control inthe Newly Integrated Work Environments. The Information Society, 10(2), 129-138. Allen, J. P., Bakos, Y., & Kling, R (1994). Sustaining New C~rdinatjonMethods: The Case of World Class ~anufacturing.Working Paper. Irvine, C A Center for Research on Information Technology and Organizations, University of California. Allen, J. P., Kling,R., &Elliott, N.(1994). Technol~andSkillsin Of~The-Shelf Organ~ations. Working Paper. Irvine, C A Center for Research on Information Technology and Organizations, University of California. Attewell, P. (1994). Skill and occupational changes in US. manufacturing.In P. S. Adler (Ed.), Technology and the Future of Work (pp. 46-88). New York Oxford University Press. Bakos, Y., & Brynjolfsson, E.(1993a). Why Information Technology Hasn't Increased the Optimal f Hawaii International Conference on System SciNumber of Suppliers. ~ ~ e e d i n g stheo26th ences, January. Bakos, Y., & Brynjolfsson, E. (1993b). Information Technology, Incentives andthe Optimal Number of Suppliers. Journal of manage men^ Information Systems, 10(2), 37-53, Fall. Bakos, Y., & Brynjolfsson, E. (1993~).From Vendors to Partners: Information Technology and Incomplete Contractsin Buyer-Supplier Relationships.Journal o f O ~ a n ~ a t ~Computin~, onal 3(3), 301-328, Beuschel, W., & Kling, R. (1992).How Coordination Processes Influence CIM Development. In CS of Hy6rid

Automation Systems-II~~ ~ ~ e d ~ n g s Aspects o f Advanced ~anufacturing and Hybrid

oes Information Technology Bullen, C. V., & Bennett, J. L.(1996). Groupware in practice: An interpretation of work experi-

(Ed.), Com~uter~ation and Con~oversy:Value Confljcts and Social Choices

(2nd edition). San Diego: AcademicPress. Cohen, S. S., & Zysman, J.(1987). ~anufacturing ~atters: The Myth of the P ~ t - I n d u s ~ ~ a l ~ c o n o m y . Information technology implementation research: A technagemen?Science, 36(2), 123-139.

KLlNG ET AL.

Dertouzos, M. L., Lester, R, K., & Solow, R. M. (1989). Made in America: Regaining the Productive Edge. Cambridge,M A MIT Press. Elliott, M., & Kling, R. (1997). Organizational usability of digital libraries: Case study of legal research in civil and criminal courts.Journal of the American Society for Information Science, 48(11),1023-1035.

Elliott, M., Kling, R., & Allen, J. (1994). O ~ a n ~ a t i o n a l U s a b i Information li~ of Systems: Strategies for Assessment. Working Paper. Irvine, C A Center for Research on Information Technology and Organizations, Universityof California, Galbraith,J. (1977). O ~ a n ~ a t i o nDesi'gn. al Reading,M A Addison-Wesley. Grant, R. A., Higgins, A., & Irving, R. H. (1988). Computerized performance monitors: Are they costing you customers?Sloan Man~ement Revieq 29(3), 39-45. Gasser, L. (1986). The Integration of Computing and Routine Work.ACM Transactions on ON'ice Information Systems, 4(3), 205-225. Grudin, J. (1994). Groupware and social dynamics: Eight challenges for developers. Communications ofthe ACM, 37(1), 92-105. Gurbaxani, V.,& Shi, E. (1992). Computers and Coordinationin Manufacturing. Journal of Organ~ationalC o m p ~ t2i(1 ~), 27-46. Harrington,J. (1973). Computer IntegratedManufac~ring.Huntington, Ny: Krieger. Hayes, R.H.,Wheelwright,S. C., & Clark, K. B. (1988). Dynamic Manufacturi~:Creating the Learning O ~ a n ~ a t i oNew n . York Free Press. Jaikumar, R. (1986). Post-industrial manufacturing.Harvard Business Review,November-December, 69-76. Kling, R. (1980). Social Analyses of Computing: Theoretical Orientations in Recent Empirical Research. Computing Surveys, 12(1), 61-110. Kling, R. (1987). Defining the Boundaries of ~omputing Across Complex Or~anizations.In R. Boland & R. Hirschheim (Eds.), Critical Issues in Information Systems(pp. 307-362). London: Wiley. Kling, R. (1991). Cooperation, Coordination, and Controlin Computer-Supported Work.Communications o f the ACM, 34(12), 83-88. Kling, R. (1992a). Behind the Terminal: The Critical Role of Computing lnfrastructure In Effective Information Systems' Development and Use. In W. Cotterman & J, Senn (Eds.), Challenges and Strategies For Research in Systems Development (pp. 153-201). London: Wiiley. Kling, R. (1992b). The AIME Project. C~rdinationTheory and ~ollaborationtechno lo^ Workshop Proceedings. Washington,DC: National Science Foundation. Kling, R. (1993). The AIME Project. C~rdinationTheory and Collaboration T e c ~ n oWorkshop l~ Proceedings. Washington,DC: National Science Foundation. Kling, R. (Ed). (1996). Computer~ationand Controversy: Value Con~ictsand Social Choices (2nd ed.). San Diego: Academic Press. Kling, R., & Elliott, M. (1994). Digital design for organizational usability. Proceedings o f Digital Libraries '94 Conference (pp. 146-155). J, L. Schnase, J, J. Leggett, R. K. Furuta, & T. Metcalfe (Eds.). College Station, Kling, R., & Iacono, S. (1984). Computing as an occasion forsocial control.Journal ofsocial Issues, 40(3), 77-96.

Kling, R., & Iacono, S. (1989). The institutional character of computerized information systems. Office:Techn~ogy and People, 5(1), 7-28, Kling, R., & Jewett, T. (1994). The Social Design of Worklife With Computers and Networks: An Open Natural Systems Perspective.In M. C. Yovits (Ed,),Advances in Computers vol.39, pp, 239-293). Orlando, F L Academic Press. Kling, R., Kraemer,K. L., Allen, J., Bakos,Y., Gurbaxani,V., & King, J. (1992). Information Systems in Manufacturing Coordination: Economic and Social Perspectives. Proceedings of the Thirteenth International Conference on Information Systems,Dallas, TX. Knights, D., & Murray, F. (1994). M a n ~ e r Divided. s Organ~ationPolitics and Information Technology Management, London: Wiley.

16.PROMISE AND PROBLEMSOFIT IN COO~INATION Laudon, K.(1974). Computers and~ureaucraticReform.~lley-Interscience:New York. of coordination. ACM ComputMalone, T. W,, & Crowston, K. (1994). The interdisciplinary study ing Surveys, Z6( l), 87-119. Melnyk, S. A., & Narasimhan, R. (1992). ComputerInte~atedManufacturing: Guidelines and Applications from Industrial Leaders. Homewood, I L Business One Irwin. Meyer, J. W., & Rowan, B. (1977). Institutionalized organizations: Formalstructure asmyth and ceremony.American Journalof Sociolo~,83(2), 340-63. National Research Council, Computer Science and Telecommunications Board (1994).Information T e c h n in~ the ~ Service Sector: AT w e n ~ ~ i rCentury st Lever.Washington, DC: National Academy Press. Orlikowski, W. J. (1993). Learning from Notes: Organ~zational Issues in Groupware Implementation. The Information Society, 9(3), 237-250. Perrow, C. (1986). Complex O ~ a n ~ a t i o nAsCritical : Essay(3rd ed.). New York Random House, Pfeffer, J. (1982). O ~ a n ~ a t i oand n s O ~ a n ~ a t i Theory. on Boston: Pitman. Powell, W. W., & DiMaggio, P. J.(Eds,). (1991). The New Institutionalism O in~ a n ~ a t i o nAnalysis. al Chicago: Universityof Chicago Press. al Powell, W. W., & Dimaggio, P. J. (Eds,). (1991).The New Institutionalism inO ~ a n ~ a t i o nAnalysis. Chicago: Universityof Chicago Press. Roven, B., & Pass, S. (1992). Manufacturing m~agementinformation systems require simplifi~, 50-53, cation.IndustrialE ~ i n e e r iFebruary, s: Natural, and Open Systems (3rd ed.). Englewood Cliffs, Scott,W. R.(1992). O ~ a n ~ a t i o nrational^ NJ: Prentice Hall. Scott, W. R., & Meyer, J.W. (1991). The organizationof societal sectors: Propositions and early evidence. In W.W. Powell & P, J. DiMaggio (Eds.), The New Inst~tutionalism in O~an~ational Analys~@p. 108-142). Chicago: University of Chicago Press, Scott Morton, M. S. (Ed.). (1991). The Corporutionofthe 1990s: In~ormationT e c h n o l ~and Oqpn~ationalT r a ~ f o r m u ~New o ~ .York Oxford UniversityPress, Shani, A. B., Grant, R., Krishnan, R., & Thompson, E. (1992). Advanced Manufacturin~Systems California~anagementReview, and Org~izationalChoice: Sociotechnical System Approach. 34(4), 91-111. Sheridan, J.W. (1992). The CIM evolution: Bringing people back into the equation.Industry Week, April 20,29-51. Wagner, I. (1996). Confronting ethical issues of systems designin a web of social relationships. In R. Kling (Ed.), Computer~ationand Controversy: ValueCon~ictsand Social Choices(2nd ed., pp. 889-902). San Diego: AcademicPress. Warner, T. N. (1987).Informationtechnology as acompetitiveburden. Slmn Management Review, Fall, 55-61.

This Page Intentionally Left Blank

C H A P T E R

hristine M. ~ e u w i ~ h David S. Kaufer Ravinder Chandhok James H.Morris

Camegie Mellon University

The goal of our research is to provide computer support for distributed collaborative writing. Writers can be said to be distributed when they have disin tributed knowledge and skill, and they share that knowledge and skill order to develop a draft; or, even when they have significant overlap in knowledge and skill, they distribute the work of producing the draft itself In the among them. Butin this sense, all collaborative writing is distributed. sense we use the termhere, distributed collaborative writing refers to, additionally, situationsin which the writers are distributed in time (Le., they do not work on the artifact at the same time) or place (Le., they do not meet face-to-face).Thecentralresearchquestions in distributedcollaborative writing are: What does the processof producing a written product look like when it is divided among writers who coordinate to produce it over time and space? and Whatis the relationshipof these processes to success? When the process includes “active agents,” the scope of the first question shifts slightly to include not only people, but computers as well. This question is, of course, the central questionof ‘~istributedcognition” or‘~oordinationscience,” applied to collaborative writing, Analogous with the way cognitive scientists (psychologists,AI researchers, and so on) are interested in identifying strategies and representations involved in individual cognition, coor-

N E ~ R ET T AL. ~

dination scientistsare interested in identifying the strategies and representations that groups of %gents”-people and computers-use to coordinate their activities (Malone, 19%). The central question of computer support for distributed collaborative writing is: What are the requirements for supporting distributed collaborative writing processes? This question is related to the previous questions: in order to produce knowledge that is usefulin designing computer support, the description of strategies and representations needs to be sufficiently detailed that it yields answers to a set of related questions, including (a) What problems do such writers have and are there ways computers can mitigate them?(b) Are there ways computers can augment the processes? (cf. Olson & Olson, 1991). There are properties of the processof writing that makes it an interesting, although challenging, application domain for coordination science. Writing is an open-ended design process.A design process is one that involves the creation ofan artifact. An o ~ e n ~design n ~ eprocess ~ is one in which any existing specifications for the artifact leave many design decisions open, in order to create it. Moredesign decisions that nevertheless must be made over, any specifications that might exist are often open to interpretation, and the more heterogeneous the background knowledge and skills of members of the collaborative writing group, the more likely that differences in interpretation will arise(Gabarro,1987).As a result of these properties, there may be situationsin which members of thecollaborative writing group do not have shared knowledge, shared goals or criteria, or even a shared representation of how best to proceed with the task (cf. Hewitt, 1986). Y

Our research strategy can be outlined by the following steps (Neuwirth& bufer, 1992): Identi~ingwriters (e.g., novices, experts) and a writing task (e.g., coauthoring). In our research, we have not focused on collaborations in which coauthors or comment~rsinteract face-to-face, although systems that support researchintotheissuessuchcollaborationsraiseareclearlyvaluable ~ ~ c ~ u g Hymes, h l i n & Olson, 1992; Olson, Olson, Storrrasten, & Carter, 1992). * ~uilding a theory- and research-based model of the task, with a focus on user-centered design(Could, 1988). Understanding howwriters function and hypothesizing the sources of their successes and failures is vital building to tools to support writers. The model draws on techniques, both cognitive and social, for building models of composing processes,but focuses onproblems that writers-even experienced ones-have with the task. This model informs the design of technology.

*

17. DISTMBUTED C O L L ~ O ~ T I V WMTING E

Designing technology to alleviate these problems. This step involves building a theory of the prima facie ways computers can augment writers’ performance of the task by drawing on a theory of the role of external representations and atheory oftaskactivity.Thetechnology representsa hypothesis about a solution, perhaps a partial solution, to some needs or problemsidentifiedby thetheory- and research-based model. We have embodied our theories of distributed collaborative writing into a ”work in preparation”(PREP) editor, a multiuser environment to support a varietyof collaborative and, in particular, coauthoring and commenting relationships for scholarly co~munication. * Studying the techno lo^ in use, with the aim of building knowledge that will help to refine the model of the task and the design of the technology.

*

These steps are interconnected and often recursive. For example, studying in actual use led us to refine one of our software tools, the Comments program, our model of the task and to design a new software tool, the P W P Editor (Neuwirth, Kaufer, Keim, & Gillespie, 1988; Neuwirth, Kaufer, Chandhok, & Morris, 1990).Indeed,giventhatallwritinginvolvestechnology(e.g.,penand paper), the last step can be thoughtof as the second step repeated.It is not the same group of researchers.A necessary for the steps to be carriedbyout study by a group of empirical researchers observing a software tool in use may be relevant to researchers working at other steps, perhaps ontheory the of composing or on the design of software. For example, the work of Haas and Hayes (waas, 1989a, 1989b; Haas& Hayes, 1986), which identified the problems writers have of getting a “sense” of their texts when using word processors, added to our theoretical understanding of the processof composing by identifjhg an additional subprocess, the subprocess of reading one’s own writing, and highlighting its importance. This result stimulated further research into the roleof reading during writing, both in print and hypertexten~ronments, and has been usedby our research group to inform software design. Although it is not necessary, then, for the steps to be carried outby the same group of researchers, it is necessary, or at least desirable, that researchersunderstandtheinterconnectedness of thesteps in orderto increase the likelihood that results they produce at onewillstep be relevant to other steps.

in the follow in^, we outline our model of collaborative writing. In its major

outlines, we have drawn heavily upon the process model of writing developed by Flower and Hayes (1981a; Hayes &C Flower, 1980). Although developed for single authors, the empirically based model isa useful also starting point for characterizing the cognitive processes involved in collabor~tiv~

N E ~ R ET T AL. ~

writing and, supplementedby observations about actual collaborative writing groups, in deriving design requirements for computer support for those processes. We next introduce the major components of the model: planning, drafting, and reviewing, discuss some implications of the model for design requirements, and provide examples of how our prototype attempts to meet those goals. In order to make general observations about collaborative writing concrete and to gain further insight into how to support groups, we observed groups of writers working underthe following conditions: able to meet faceto-face; ableto work at the same timebut not face-to-face; and neither able to meet at the same time nor place. The groups, consisting of three writers each, ere asked to write a press release, respond to two letters, and write In the following, we describe PREP Editor's a brief report on their activities.' support for collaborative writing by drawing on observational data from one of these groups, the group that was neither ableto meet at the same time nor place. lanning refers to processesof gene~ating (a) criteria for the text (e.g., the in order tomeet the needsof its purpose of the text, features the text needs audience, andso on), (b) ideas for the contentof the text, (c) plans for how to organize that content, and (d) plans for howto proceed with the process of writing itself (e.g., deciding that particular people will write particular arts of the document or do particular tasks such as review for technical accuracy or style). one, they may not need to articulate the constraints and the goals they have set(alt~oughstudies of experi-

'We studied the group's activities and interaction behaviorby observing and videotapin S asking subjects working alone to think-aloud, The group members were PhD. students in English.

17. D I S T R I ~ ~ ECOLLABORATIVE D WRITING

ing coauthors and commenters from having to infer the

other’s plans. If other coaut~orsunderstand the goal, they may be more likely to be to able produce revisions to the draft that are compatible with the first author’s goals, and the draft they produce is more likely to be seen as useful by the other author, Or, if another author has a differeslt pointof view, it may be more likely to surface and be resolved. Of course, unnecessary communicationcan alsobedistracting,leadingto a degradation in performance. Research we are currently conducting attempts to relate such patterns of communication to measures suchas thetime to completea project and the quality of the product. When collaborative writers are able to meet face-to-face, they can communicate about plans relatively easily: Face-to-face communication is both highly interactive (e.g., requests for clarification can be answered immediately) and expressive (e.g., facial cues and gestures can also betoused cornmunicate).Evenface-to-facecommunication,however,isnotwithout its problems,and,interestingly,maybeaugmented by computersupport (Olson et al., 1992). But writers seem to experience more difficulties when working over distances.In a studyof groups of writers working face-to-face versus at a distance, for example, Galegher and Kraut (1994) reported that 0 writers, w o r ~ i nat~a distance with traditional computer-mediated ~ nication tools (e.g., ernail and conferencing tools) and phone, needed to spend more time to achieve the same quality of result and repo~edless sattho§e isfaction with their work and with other members of the group t h ~ n working face-t~face. Two recent linesof research represent attempts to mit-

~

phone) and video links, allow writers to communicate at the same time over distances, supportedby the ability to see an evolving draft. The parameters that allow writers to control how quic~ly transmitted to others (~euwirth,Chandhok, Charney,

discuss initial

NE~RTH ET AL..

Here aresome questions fw you. --John ................................................... What should the press release be ~ b o u ~

that the press to know e major's ofwce is eonsiderin~bids but notyet formally chosen ways to reorganize handicapped,

................................................... Is there a document we should

I think our job is simply to report what seems

'

dedriont to hire a fleet of minibuses, 1 suppose that means finessing the drawbacks with the kind of defense that mody gave major, Or maybe not, Look at the r#tionate 1 wrote (I drafted from what Kim wrote),

1 worked on the l

rationale but work on first? Should we pick what d~umentwe want to write didn't work on anything else, or should we just pick each other's The problem is brains about the task? that I don't have the info

FIG.17.1. ~ommunicationabout initial design decisions in the PREP Editor.

tion remainsto be seen. The groups weobserved werenot shown this possible use of the tool, although weare conducting an observational study in which this usewill be demonstrated. Groups of collaborative writers using theP W P Editor over distances use the tool frequently and spontaneously to communicate about plans, goals, and constraintsin a way that is groundedin an evolving draft.As an example of such communication,Fig. 17.2depicts the draft of a letter, together with the author's own comments on the draft, intended for the other writers and explaining the goalof the paragraph. (Note also, the author indicating a jud ment about the completeness of, and confidence in, parts of the document). Although there is a tendency to equate the actof writing with producing the contentof the written draft, studies show that experienced writers typically engage in many acts of writing (e.g., jotting down ideas, drawing) that bear no direct relation to the text product, butserve as inexpensive, intermediate external representations to remind writers of their plans for audience, purpose, and procedure, as well as content (Flower & Hayes, l ~ ~ l b ;

17.

. . . . . . . . . . . . . . . . . . . * . . . . . . . * . . . . . . . ‘ . . . . . . . . . . I . . . . . , . . 1 . . . . . . . . . . I . . .

Nexxus decided to support the minibus service because that option is the only customized transportation service that meets the special needs of the d i ~ r e n t i a l l yabled. Although problems, such as those you outlined in your letter, will have to beaddressed as this new system is put into place, we believe that many of these problemscan be anticipated and

I’m not sure this paragraph iseven worth keeping,but I

i

~

~

g

for proposing minibuses, But I’mnotadverseto zapping this if i t j u s t soundslike rehashing the same

~

~

~

2:

FIG. 17.2. Communicat~on aboutplans,goals and constraints in the P E P Editor.

Flower, Schriver, Carey, Haas, & Hayes, 1989; Haas, 1990). When working with computer environments that do not support the creation of arrows, boxes, or other diagrams for displaying conceptual relationships among and ideas the suppressionof detail, writers report frustration (Bridwell-Bowles, Johnson, & Brehe,1987)andimportantplanningactivityiscurtailed(Haas, 1989b). Thus,observationsof expert writers at work suggest that supporting the jotting, drawing and note-taking that writers engage in as they writeare especially importantin writing and that cognitive aspects must be taken into account when designing computer support for coauthoring and commenting tools. There have been some attempts to understand the task-specific activities (e.g., jotting, drawing, writing, gesturing) that occur in collaborative tasks in order to informthedesignofspecialized tools to support those tasks (Stefik,et al., 1981; Tang & Leifer, 1988). But becausethere is a tendency to equate the substantive workof writing with a written draft, most text annotators support only communication about the working draft or outlines of a draft.The P W P Editor containsa drawing tool and the objects produced in the drawing can be annotated, but it has not been used extensively to support planning by groups we have observed.An outstanding research questhis may be due to the fact that the very rough s~etchtion is to whate~tent es andprivatejottingsthatwritersworkingaloneproduce are less “sharable” artifacts andto what extent it isa deficiency in the toolsto facilitate such informal sketching.An interesting research approach might be to provide tools for graphically based idea generation, roughly corres~on~ing to the “network mode” in Smith, Weiss, & Ferguson’s Writing Environ~ent (1987), along with en-based input.

~

N E ~ R T ET H AL.

Drafting isthe processof producing text. Studies of experienced writersindicate that they often set new goals for themselves as they draft, that is, they discover what it is they want to say in the process of saying it (Hayes & Flower, 1980).As a resultof this propertyof writing, a collaborative writer’s knowledgeofotherparticipantsandtheiractionsmaybeuncertain or changing. This has been confirmedby case studiesof collaborative writers at work.For example,b y e (1993) observed: “Regardlessof the levelof detail in a . . specification. .,more often than not, the author’s ideas only become clear when the extended ,..draft is being written. As a result, colleagues’ perceptions of what is being written (based on an earlier draft) may not be an accurate reflection of what any individual team memberis in fact writing. Riley(1984) referred to thisas the “out-of-step” phenomenon... Various informal tacticsare used to minimize the likelihood of [integration] problems arising, and this is obviously fairly straightforward when ..colleagues work in adjacent offices and see each other regularly, both formally and informally,in between the ...meetings” (pp. 47-48).

Any system to support collaborative writing needs to accept that any plansmade in advance of draftingdonotcompletelycontroldrafting; indeed, that plans will not be made completely in advance of writing and must support communication about changes in plans. Groups using P E P Editor use the interface to discuss evolving plans frequently. A second property of drafting is that the partially completed product plays an important rolein open-ended design processes: The partially completed product becomes part of the task environment and constrains the subsequent courseof the design. Writers frequently re-read portions of the text they have produced to provide constraints for another segment of text that they wantto produce (Flower & Hayes, 1981a; Kaufer, Hayes,& Flower, 1986). From a coordination science perspective, the draft itself is a “shared resource” that all writers may benefit from accessing, even if they have agreed to work on a particular part. These two properties of drafting suggest that there will be situationsin which collaborative writers could benefit from having parts of the document that others are working on available. This observation must be tempered, however, by noting the need to take the wide variability of writing groups into account. Some writing groups want updates to drafts to be available immediately to all membersof the group. Other groups want information about changes delayed until the source has been to check able them for correctness and ‘~ommit” to them.Forexample,NewmanandNewman (1993) ne scribed a case studyof a large groupof writers working on a document to support budget allocation decisions within their or~anization.Dif-

17. D I S T ~ B ~ ~ E DO

L

~

OWRITING ~ T I

~

ferent departments had responsibility for different parts of the document, and the outcome of the budget allocation would affect their respective budgets. In this case, departments concealed early drafts from members of other departments,so that the text would not be available to writers outside the department until the political issues had been thought through by writers inside the department. Likewise, individual authors vary considerably in their stylesof working (Posner& Baecker, 1992), with some authors anxious to receive immediate feedback on even “half-baked” ideas, whereas other authors, as b y e (1993) observed, do their writing in concentrated burstsof activity prior to previously agreed deadlines,in and, any case,may not wish to make their developing drafts public(p. 48). These observations suggest thata system to support collaborative writers should provide authors with the ability to be flexible in making parts of documents visible.We have defineda setof parameters of interaction for the networked versionof our prototype, parameters that allow writers flexibility in sharing partial results along the dimensions of who to share with, and what, when, and how quickly to share (Neuwirth et al., 1994).Each user can set these parameters to define his or her own ofpattern data exchange. For example, setting parameters for data to flow automaticallyaat grain size of a keystroke with fast transmission speed approximates synchronous communication. On the other end, setting parameters for data to flow only upon explicit request at a grain size of the column models the situation in which a coauthor requests to see the latest version of anothercoauthor. By settingthese parameters, users can adjust the characteristics of the information flow. A good deal of the work has been directed at mapping flexible social protocols onto practical communication protocols. Clearly, if social protocols are to be flexible, data exchange protocols (withina system) and network protocols (across systems) must be as well.As far as the flexibilityof data exchangewithin a systemisconcerned,DewanandChoudhary(1991) argued that systems must be flexiblein their assumptions about data interaction. They proposeda set of system primitives that would allow usersto calibrate their assumptions about the exchange of information (not simply data, but views, formats, and windows as well) to other users in a flexible fashion. Our work builds on some of their primitives by allowing for the incremental versus complete exchange of data, but extends it for collaborative writing applications. Future work needs to extend this to sharing views and windowsas well. The fact that different writing situations require integration of asynchronous and synchronous styles of work has also been noted (Dourish& Belotti, a sys1992; Posner & Baecker, 1992). Minor and Magnusson (1993) presented tem to support an integration of writers’ asynchronous and synchronous work strategies. Their model is similar to the one described here, in that it is based on working with copies of a document rather than actually sharinga

~

E

~ ETRAL. T

~

document. It differs in that, whena user opens a versionof a document that is currently being editedby another user, the system attempts to make users aware of each other’s activities by showing the differences between two versions of the document. Other parameters of interaction (grain size, flow, and so on) are not defined. The underl~ngmodel relies on system knowledge of the hierarchical nature of documents (e.g., sections, subsections). In contrast, the model described here requires minimal interpretation of the document structure per se (import/export algorithms that support reading documents produced in other word processors interpret paragraph breaks as chunks). Up to this point, we have only discussed collaborators ~ i esections ~ i of~ ~ documents thatare being worked onby others. There is reason to suppose, however, that it can also be useful for collaborators to to bechange able sections of documents being worked on by others. Because texts have “texin one section may ture,’’ that is, coherence relations throughout, a change require changes in another. That is, although writers may decompose writing the document into subtasks, the decomposition almost always entails interdependencies among subtasks. Although it is possible for an author to suggest a change to an authorof another section, it is sometimes more effia noncient to simply “do it.”In our model, coauthors can choose to accept conflicting change automatically or to receive notification of the change with final approval residing with the coauthor responsible for the section.

The process of reviewing consistsof two subprocesses: evaluating text and revising text. Often this evaluation and review process takes the form of comments on the text. The problems with comments, that is, critical notes on texts, are well-known and legion: Writers don’t understand comments, they think the comments reflect confused readings rather than problems in their texts, they are frustrated by perceived lackof consistencyin comments and contradictory comments (~euwirthetal.,1988).Theproblems in author-commenterrelationshipsbecomeevenmorepressing if authors solicit Comments from multiple readers. From a coordination science perspective, there is also a consumer-producer relationship that must be managed in the review process. When collaborative writerscan meet face-to-face, this relationship can be in an obsermanaged by communication about the comments. For example, vational studyof a group of physicists working to produce an article over a period of months, Blakeslee (1992) observed face-to-face meetings in which members of the group discussed and clarified comments that they had made on drafts. This suggests that computer support for distributed collaborative writing should support discussion, not only about plans and drafts, but also about the comments themselves.

17. DISTRIB~ED~ O L ~ B OWRITING ~ T I ~

A second form that evaluation and review takes, at least among coauthors, is actually making changesto partsof a document that someone else has written. A principal difficulty coauthors face is in copingwiththose changes, especially understanding why the other person made them. For example, in a studyof eight writers’ productionof an insurance company’s two-page annual report, Cross (1990) observed that each writer ‘‘omitted, added, highlight or modified the text to agree with his or her preconception, with unexplained changes causing “considerable frustration” for other writers and an undetected change causinga major problem(p. 193). This suggests thata system to support collaborative writing should support the detection of changes from one version to another, along with supporting communication about those changes (e.g., annotating the changes with questions about the decision and explanations of changes). With paper documents, even reviewers often make “changes” in content by marking up the draft. This phenomenon may be due to the fact that many significantproblems in texts(e.g.,voice,persuasiveness,organization), although easyfor anexperiencedwritertodetect,cannotbeeasily described. For such problems, rewriting is often a more efficient strategy than trying to describe the problem, and writers often choose this strategy &C Carey, when revising others’ texts (Hayes, Flower, Schriver; Stratman, 1987). Some earlysystemstosupportcollaborativewriting(Comments, ~uilt)restricted reviewers to the role of attaching comments to the base document. Although this increases the usabilityof the commenters’ activities from the point of view of the author, it seems to increase the difficulty of the task for reviewers. In our observations of reviewers working with the Comments prototype, writers in the role of commentersoftencopied a region of the base document into a commentingboxandproceeded to in this fashion, however, reporteddifrewrite the copy. Writers who worked ficulties in revising because their revisions were physically separated from the larger bodyof text. More specifically, they reported needing a “senseof the whole text” even when commenting on a part. One exasperated commenter went so far as to copy anentire document into a comment box and to reviseit from there. Whethera commenter is able to modify the base document or not should certainly depend on his or her rightful relationship (coauthor, commenter) to the text. Despite potential problems, role specification is likelyto be a useful strategy for managing some coordination problems; our design, however, allows new ways of dealing with this interdeoiew of the pendency by giving commentersthe ability to rewrite his or her textandsupportingwaysforauthorsto seethechangesasproposed changes to the original base document. Figure 17.3 depicts an interface for detecting changes from one version of in the PREP Editor. The comparison interface produces its a draft to another report in a new column, with the differences linked to the original column for

N E ~ R T ET H AL.

two years ago. ~ c c o ~toi one n ~

the only one customized to the about theproble~s that existed in the resolved?'

FIG.17.3. Communicating about changes in drafts,

easy, side-by-side evaluation.To illustrate,Fig. 17.3 depicts four columns: an original draft, its revision, the comparison report, and an evaluation column An author can "push" particular revisions across the link, as in InterNote (Catlin, Bush, 4k Yankelovich, 1989). The evaluation column in Fig. 17.3 consists of annotations to the comparison report thata coauthor producedin order to explain some of the changes or to solicit advice about them. An important featureof the interface is the abilityof users to annotate changes withexplanations of thechangeorquestionstocoauthors.Ourgroup's experience with this feature suggests that reviewerswill annotate changes selectively in order to draw their coauthors' attention to changes they want to discuss or explain. Likewise,a coauthor can ask a reviewer to explain a change. Our hypothesis is that the ability to annotate changes will greatly alleviate writers' frustrations with undetected and unexplained changes that Cross (1990) observed. We have experimented with heuristics for automatically generating comparison reports, depending on role relationships among writers. For example, if the annotated draft is from a coauthor, then display changes upon request; if from a reviewer, then display all changes automatically. Apart from genera tin^ the Comparison before returning the revision (which as we, coauthors, currently sometimes do), the revision's author has little control

17. DISTRIBUTED C O L L A B O ~ T WRITING ~

over how the comparison might be done. As this information might leadto a more productive exchange, we plan to experiment with adding “comparison settings” information to revisions that would serve as hints from the coauthor to anyone who would generatea difference report.

of our theory of collaboraThe PFEP editor’ prototype, then, embodies part tive writing.Itapproachesrequirements by supporting com~unication about plans, constraints, drafts, comments, andso on, and by providing a flexible set of parameters for interaction, Central to the PREP editor is a focus on providing a usable, visual representation of the information that will allow new ways of managing interdependencies in open-ended design tasks, in addition to supporting existing patterns. Much of our work has focused on the interface, specifically on the visual representation of the draft and an optimized action grammar. For therep visual resentation, we have pursued a path that could be called “dynamic glossing,” because we support annotation in a style similar to old, glossed scholarly texts. Although in some sense this means that we are mimicking the static annotation process, weare alsotaking advantage of the dynamic natureof the computer to use visual cues such as shading and spatial relationship to show the interconnections among chunksin the system.To create a visual system thatwill lend itself to providing and accessing comments easily, the visualgrammarmustbecapableofsupporting writers’ needs. We have found, for example, that visual alignment of comments isa useful feature for allowing collaborators to see comments ‘ht a glance” (see Fig. l?.l), but in a flexible system, the general case requires a constraint-based layout algorithmthatcanhandle arbitraryshapes andcomplexinterconnections amongdynamicallyselecteditems(Smolensky,Bell,Fox, King, & Lewis, 198’7). We have also worked on the action grammar, optimizing actions that are used frequently. For example, to create a comment, a writer need only click and drag the mouse. The cognitive needs of collaborative writers are too numerous to detail here. We focus, therefore,on one: accessing comments. Most text annotation systemsare based on a hypermedia model and the primary method for 2TheprototyperunsonMacsandisavailablebyanonymousftp at http://eserver.org/ for a commercial product, Common softwarelprep. See alsohttp://~*si~hfloor.com/CSl.html Space, based on this research.

~

E ET AL. ~

accessing information in hypermedia systems is following link icons from node to node. Typically the user brings a node (e.g., a text node) onto the screen, reads its contents and notes any links, then chooses to traverse some of the links. Such localized link following is adequate for browsing tasks but has been problematic for others (Halasz, 1987). For example, we have found that coauthors and commenters want to visually scan a set of comments quickly and resent the time required by the “search and click interfacetocall up eachcomment,inspectitandputitaway.Some researchers have worked to tailor the navi~ationallinking system of hypermedia systems to meet user’s writing needs (Catlin et al., 1989; Edwards, Levine, & Kurland,1986;Fish,Kraut,Leland, & Kurland,1988;Neuwirth, ufer, Chimera, & Cillespie, 1987), but the access problem remains to be addressed. Our approach calls for a tailoring the program to match user’s cognitive activities (Norman,1986). We have previously analyzed the design features of lftthcentury glossed bibles in Cavalier, Chandhok, Morris, Kaufer, & Newirth (1991) and the match to the cognitive needs of commenters. The analysis suggests the following requirements (seeFig. 17.1): The prima^ text is easilyd~tinguishablefrom the annotation text.This requirement allows readers, who may not have seen either the original text or the annotations, to orient themselves to the texts quickly. In glossed bibles, this distinction was usually madeby varying type size: the primary text is several points larger than textin the annotations.Of course, other typographic signals such as colorcould also be used. The annotations are visible ‘kt a glance” while reading the prima^ text, This requirement minimizes the problems readers havein accessing annotations. Glossedbibleswereusuallytheresult of calligraphic as well as scholarly effort; the annotations were packed in an aesthetically pleasing fashion onto a page, so that all annotations, no matter how dense,are visible.As the corpus of annotations increased over time, the books were recopied with more space by expanding the leading between for annotations, preserving the easy access the lines in the primary text as needed to insure the visual alignment of all annotations. +

+

+

The relationsh~between the ~rimarytext and the annotations is easy to see.

This requirement insures that readers will be able to see which annotations refer to particular portions of text. In glossed bibles, the annotations were typso it is possibleto scan from the ically aligned horizontally to the primary text, primary text across to the annotation rapidly. Moreover, the of thescope annotation was usually indicated by graphic symbols in the primary text. ~ i ~ e r econ~ibutors nt are readily disting~~hable. This requirement aids readers in interpreting annotations by different commentators. The different annotators of glossed biblesare easily distinguishable because each has his or her own column. +

R

The result of all these features of glossed bibles was that access to the annotations was superior to most electronic annotation systems. The reader could quickly skim the set of annotations "ata glance." For the scholar, the assignment of marginal "real estate" allowed for quick and easy annotation. A comment could be made as quickly as moving the pen to the adjacent margin. In addition, several scholars often annotated a document side by side, leading to an easy to follow parallel discussion that was the synthesis of in the PREP Editor both sets of comments.We have embodied these features prototype. The PREP Editor utilizes an underlying node-link architecture. This section describes those mechanisms and discusses the features that underlying hypertext engines need to provide to support annotations as theyare implemented in PREP. Although theP E P Editor prototype restricts itself to linear texts, themodel employed can be applied in those portions of a hypertext app~ications that provide design objects to supporta linear layout of nodes and links, for example,GUIDE, the ~hetoricalSpace in SEPIA (Haake & Wilson, 1992). To describe thePREP architecture and interface, we use terms drawn from the Dexter hypertext reference model (Halasz & Schwartz, 1990). The Dexter model divides hypertext functionality into three layers: the storage layer-the node-link network structure; the r ~ n t i layer-the ~e mechanisms supporting the user's interaction with the hypertext (including presentation); and the ~ i t h i ~ ~ o layer-the ~ ~ o ~ content e n t and structureith hi^ nodes and links. The fundamental entity in the Dexter storage layer model isa component, either an atomic component(or node), alink Component (or link), or a composite component (or composite node) composed of other components. P E P Editor definesa C O Z to ~ be ~ a~ composite node consistingof atomic or

composite nodes with "path" links between them, forming the nodes into a connected graph. The nodes of a column are further constrainedin that the "path" links, together with a traversalmechanism defined for them, must allow the hypertext runtime layer to construct a linear ordering for the nodes (i.e., to display the nodes linearly). The structure of links among the nodes does not haw to be restricted to a directed-acyclic graph (DAG), but the traversal mechanism must include a decision rule for finitely terminating any cycles. The ~ ~ ~ o t a t ~i~~ i o A~ i ~ a ~~,i r e c t i o ~~ f f ~ ,y

~~~~

~

notation e ~

links" are binary, directional, typed links froma source node to an annotation node. ThePREP Editor allows users to create annotation links between

N E ~ R ET T AL. ~

columns (i.e., composite nodes)? For example,in Fig. 17.1, the two columns on the right are linked to the leftmost column. Such links define a tree of linked columns (althougha single PREP document can holda forestof these trees). Links between columns allow users to create annotations rapidly:A user only has to locate where in the primary text to make an annotation and in the linked columnto make both a link and new click next to that location node: This user interface allows us to approximate the ease of paper-based annotating by proximal writing in a margin. However, PREP goes one step furtherthan a paper-basedmetaphor in that PREP annotationsremain aligned to their source even as the source text is edited. ~ j t In ~ most ~ . hypertext systems, a link, regardless of its type, is representedin the runtime layer layout asa line connecting two squares (in a graph view), or is representedby an icon that is given a "folin the PREP applilow it" interpretation when the user clicks on it. The links cation, however, carry different implications for the runtime-presentation layer. Path links between nodesin a column result in a linear, scrollable disin a word processor. play of the nodes that look like an ordinary document s notation links between columns resultin the linked column being allocated (by default)anarrowerdisplayandsmallerfont;unlinkedcolumns, which usually containa primary text,are allocated (by default) a wider display and larger fonts. Annotation links between nodes result in the nodes being displayed in a side-by-side, horizontal alignment. An object-oriented constraint-based layout algorithm is at the heart of determining andmaintain in^ side-by-side layoutof annotations. Whenever a user changes the screen through some operationaon column or node-creating, moving, deleting, linking, adding to, and so on-constraints from the local objects on the screen whose display is affected by the change are placed in a queue. A constraint solver satisfies the constraints; their satisfaction often leads to the propagation and satisfactionof more constraints. The process of constraint maintenance and propagation continues cyclically until the display is no longer affected. The PREP editor requires dynamic communication between the within-component layer (the content and structure within nodes and links) and the runtime-layer. In particular, the PREP editor requires~ o ~ p o n e n t i ~ and a t i osize n infor~ation, and size and i ~ ~ t i o n A i ~ n ~ o p t orequires To~ the chan~e events. For example,a constraint such as

'When a user creates a new column, if a column is selected, then the new column will be linked to the selected column. If no column is selected, the new columnwill be unlinked.To link two existingcolumns, a userselects the "from"column, chooses "Link" froma "Column"menu, and chooses the "to"column. 41f the user actually makes aselection in the primary text and clicks in the linked column, then the system puts a link anchor around the selection; otherwise, the system puts a zerolength link anchor in the same line as the primary text.

constraint solver to be able to query an object about its position on the screen. This, in effect, results in querying a link about the location of the component at the other end. el. The storage model presented to users is a datain a file as an uninterpreted byte base. Whereas a file system presents data sequence, a database encapsulates substantial information about the types and logical relationships of data items stored within it. A database model in it (e.g., chunks, linked allows users to exploit knowledge of the data stored columns, andso forth) to achieve better performance.

ure. The abstractsystemarchitecture consists of three logical units:a source, a filter, and the receiver(s). In this system, a source agent consults the filter to determine what information to send informato which receivers. The work has significant similarities withinwork tion filtering (cf. ACM Special Issue,1992). The differences residein the information source having human agency and thus being ableto offer information based on nonformalized models of receivers’ interestsand states. Information transport can be done in a varietyof ways. We have chosen to emphasize scalability and independence from details ofanindigenous network in our implementation. We want the system to support small user groups of two to four individuals, as it has in the past. But wealso want it to A scale upwardracef fully to much larger groups across wide-area networks. goal, for example,is to havePREP support usersof major national research and information networks. As a result, we have chosen stochastic (rather than deterministic) algorithms, of the kind represented in epidemic algorithms (Demers et al., 1987), as having desirable properties of scaleability and independence from many of the details of an indigenous network. The highly replicatedmodelwe are employinghassignificantimplications for consistency. Issues of concurrency control for advanced applications suchas cooperation and coordination are just beginningto be studied (Barghouti & Kaiser, 1991). Most work has been done in the contextof software development environments and some in CAD systems. It is important, however, to pursue development of protocols for collaborative writing as well, because the consequencesof inconsistency are quite different fordifferentdomains(e.g.,theobjectcodeproduced by compilationmaybe invalid, whereasa writermay be ableto regain consistencyby reversing the effects of some operations explicitly). Thus, we believe that writers may be willing to trade “high availability” for ”accuracy,” an issue we intend to explore through providing users with a setof parameters that they control. This strategy turns “consistency” from an accuracy issue into a performance issue. An example of a system that implements a similar policy is the Coda file system (ICistler& Satyanarayanan, 1992). Ekcept for a small number .

N E ~ I R T HET AL.

of files that must remain consistent, Coda’s strategy is to provide the highest availability at the best performance: The most recent copy that is physically accessible is always used to satisfya file request. Coda’s view is that inconsistency is tolerable if it is rare, occurs only under conditions of failure, is always detected, and is allowed to propagate as little as possible. Kistler and Satyanara~anan(1991) noted that it is the relative infrequencyof simultaneous writ~sharingof files by multiple usersin most file system environments of infrequent that makes this a viable policy. Unfortunately, the assumption simultaneous writ~sharingof files sometimes fails to hold for collaborative writing groups. Several users may want to share of parts a document in order in a tight deadline, one person may be workto be able to work. For example, ing five paragraphs behind another, the first person draftin technical details, the second, adding support from empirical research results. Although this situation could be accommodated by implementing the artifact as a large number of elements implemented as small files, it is easy to imagine situations in which this solution breaks down: For example, a person may, while drafting, discover that he or she needs to amake change toa part of the document currently being worked on by someone else.On the other hand, collaborative writing groups spend large parts of their time engagedin less synchronousactivities. A fixedpolicythatpenalizessuchgroups to accommodate the simultaneous~ite-sharingcase seems inappropriate.

To examinethe P W P Editorprototype’ssupportforvoicemodality,we undertook a study with two goals: to compare the nature and quantity of voice and written comments, and to evaluate how writers responded to comments produced in eachmode(Neuwirthet al., 1994).Writerswerepairedwith reviewers who made either keyboarded or spoken annotations from which thewritersrevised. Thestudyprovidesdirectevidencethatthegreater expressivity of the voice modality, which previous research suggested benefits reviewers, produces annotations that writersfind alsousable. Interactions of modality with the typeof annotation suggest specific advantagesof each mode for enhancing the processes of reviewand revision. This study adds to the previous pictureof the utilityof the voice modality for supporting collaborative writing activities. The results can be summarized as follows: 1. The mode of production (keyboarded vs. spoken) affected the typeof problem that reviewers communicated: Although all reviewersin the study produced more comments on problems of substance than any other typeof problem, reviewers in voice mode were likely to produce more comments about purpose and audience than reviewers in keyboard mode, whereas

17. DISTRIB~ED COLLABORATIVEWRITING

reviewers in keyboard mode were likely to produce more comments about substance. It may bethat the written text, which more readily permits review of what has been written, reflection upon it, and revision, may facilitate comments that involve complex substantive issues.If production modality doesinfluence the typesof problems communicated, then writing tools offering both modes may need to provide guidelines for choosing the most appropriate mode to work in for encouraging evaluation at the appropriate level. 2. The mode of production affected how reviewers characterized problems. Althoughreviewers in both modalities produced about the same number of annotations overall, the numberofwordsperannotationwasfar greater in speech. This difference can be accounted for, in part, by the greater frequency of reasons and by the greater number of words used to produce mitigated statements. A higher proportion of the annotations produced in voice contained reasonswhy the reviewers thought something was a problem and polite language that mitigated the problem. 3. The mode of production affected how writers perceived their reviewers, Writers’ evaluations of their reviewers were likely to be less positive when reviewers produced written, rather than spoken, annotations. 4. The study failed tofind an overall differencein reviewers’ assessments in the of how responsive writers were to annotations produced or received two modalities. Future analyses are planned to examine whether the nature of the annotations and writers’ perceptions of reviewers interacted with responsiveness. 5. Despite the previous research findings that spoken annotations would likely be tedious to listen to and more difficult to process, writers using the PREP Editor interface for voice annotations were generally favorably disposed or neutral to voice annotations for most typesof comments, except low-level mechanical ones. In this study, authors chose their reviewers and reviewers were conin only one modality. Moreresearch is needstrained to produce comments ed that varies both conditionsof producing annotations and the social relations between the writer and reviewer and looks at annotation interfaces for other sortsof documents (e.g.,CAR drawings, blueprints, videos). NCL

Our approach has been to draw on the social and cognitive research literature in writing and on our experience with prototype tools to identify social, cognitive and practical issues that we are attempting to address with a formativ~evaluation-based prototype.

N E ~ R ET T AL. ~

If we believe that our tool allows writers to create new forms of interaction, we needto understand the possibilities better. What new kinds of coordination structureswill emerge? Are these new structures desirable? What is necessary for them to work well? We are conducting studies that chart how the prototype is used, as work teams make progress through realistic document coauthoring andcom~entingtasks. This should help us come up with better descriptive theories that go beyond the normative theory that currently prevails. McCrath (1990,p. 54) noted how an increasein the volumeof information is related to a decrease in the ability to control and structure it. Similarly, collaborative technologies that increasingly relax the boundaries of who, where,when,andwhatinformation will flow across a group network is bound to increase uncertainty.We see our work as split between increasing the technological potentialof group interaction and harnessing this potential to satisfactorycommunication outcomes. Down this second branch, we expect to find some technological solutions, but many social ones as well. Cultures define hundredsof regulatory devicesin face-to-face interaction to of establishing culmonitor social behavior. We are stillin the earliest stages tures for group exchange over networks.

The work reported here has been supported by NSF under grant numberIRI8902891. We thank Dale Miller and Paul Erion for work on programming the PREP editor prototype,and Todd Cavalier for work on graphic design for the PWP editor interface.

ACM Special Issueon Informat~onFiltering. (1992). Communicat~onsof the ACM, 35(12), 27-%l. Barghouti, N. S., & Kaiser, G.E. (1991). Concurrency controlin advanced database applications. ~omputingSurveys,23(3), 269-317. Blakeslee, A. (1992). Inventing scientific discourse: dimensions of rhetorical knowledge in physics. Unpublished doctoral dissertation. Carnegie Mellon University, Pittsburgh, PA Bridwell-Bowles,L.S., Johnson, P., & Brehe, S. (1987). Computers and composing: Case studies of experienced writers. InA. Matsuhashi(Ed.), Writinginrealtime: odel ling production processes (pp. 81-107). Norwood, N J Ablex. Catlin, T., Bush, P.,& Yankelovich, N. (1989). InterNote: Extending a hypermedia framework to ~9 (pp. 365-37%). Pittsburgh, PA. support annotative collaboration. In H y p e r t ~ t ~oceedings Cavalier, T., Chandhok,R., Morris, J., Kaufer,D., & Neuwirth, C. M. (1991). A visual designfor collaborative work Columns for commenting and annotation. J. F. Nunamaker, Jr., (Ed.), P m ceedings of the24thHawaiiInternationalConferenceonSystemSciences (HICSS-24) @p. 729-738). IEEE Press.

17. DIST~BUTEDCOLLABO~TIVE WRITING of an Cross, G. A. (1990). A Bakhtinian explorationof factors affecting the collaborative writing executive letter of an annual report.Research in the Teaching ofEnglish' 24(2), 173-203. Demers, A., Greene, D., Hauser, C., Irish, W., & Larson, J.(1987). Epidemic algorithms for repliProceedings of the Sixth Annual AC~Symposiumon Principlesof cated database maintenance. dk~ibutedcomputing (pp. 1-12). ACM Press. Dewan, P., 8t Choudhary, R, (1991). Flexible user interface couplingin a collaborative system.In Proceedings ofthe CN1'92 Conference (pp. 41-49). New Orleans, ACM SIGCHI. In ProceedDourish, P,,& Bellotti, V. (1992). Awareness and coordination in shared workspaces. ings B C W '92Conference onCo~puter~upported Cooperative Work@p. 107-114). ACM SIGCHI & SIGOIS, Toronto, Canada. Edwards, M. R., Levine, J. A., & Kurland, D. M. (1986). ForComment.Novato, C A Broderbund. Fish, R. S,, Kraut, R, E., Leland, M. D.P., & Cohen, M. (1988). Quilt: A collaborative tool forcoop erative writing. In Proceedings of COIS '88 Conference on OfficeInformation Systems (pp. 30-37). ACM SIGOIS. Flower, L., & Hayes, J. R, (198la). A cognitive process of theory of writing. College Comp~ition and Communication,32,365-387. Flower, L., & Hayes, J. R. (1981b). The pregnant pause:An inquiry into the nature of planning. Research in the TeachingofEnglkh, 15, 229-243. Flower, L., Schriver, K. A., Carey, L., Haas, C.,& Hayes, J.R. (1989). P~anningin writing: The cognition of a cons~uctiveprocess. Technical Report34, Center for the Studyof Writing, Carnegie Mellon University. Gabarro, J.J.(1987). The developmentof working relationships.In J. W. Lorsch (Ed.), H a n d ~ ~ k of o~an~ational behavior (pp, 172-189). Englewood Cliffs, NJ:Prentice Hall, Galegher, J., &Kraut, R. E. (1994). Computer-mediated communication for intellectual teamwork An experiment in group writing.information Systems Research5(2), 110-138. Gould, J.D. (1988). How to design usable systems. In M. Helander (Ed.),~ a n d o bf ~h u~m a n ~ o ~ puter interaction(pp. 757-789). Amsterdam, North-Holland: Elsevier. Haake, J.,&Wilson, B. (1992). Supporting collaborative writingof hyperdocuments in SEPIA, In Proceedings of CSCW92 Conference on Computer~uppo~ed Cooperative Work (pp. 138-146). Toronto, Canada. Haas, C. (1989a). Seeing it on the screen isn't really seeing it: Computer writers' reading problems. In G. E.Hawisher & C. L. Selfe (Eds.),Critical Perspectives on Computers and Comp~ition (pp. 16-29). New York Teachers College Press. Haas, C. (1989b). How the Writing Medium Shapes the Writing Process: Effects of Word Processing on Planning. Research in the Teaching~ E ~ l i 23(2), s h 181-207. Haas, C, (1990). Composing in technological contexts:A study of not-making. Written Communication 7(4), 512-547. Haas, C., & Hayes, J. R. (1986). What did I just say? Reading problems in writing with the machine. Research in the TeachingofEnglish 20,22-35. Halasz, F. G. (1987). Reflections on Notecards: Seven issues for the next generation of hypermedia systems.In ~ y p e r t ~ t ' 8 7 P r o c e e d(pp. i ~ s 345-365). Chapel Hill,NC. Halasz, F.G., & Schwartz, M. (1990). The Dexter hypertext reference model. In ~ ~ e e d i n of g the s Hypert~tStandard~ation Workshop, ~ationalinstitute of Standardsand T e c h n o l ~@p. 95-133). NIST Special Publication500-178. Washington, DC: US. Government PrintingOffice. Hayes, J. R, & Flower, L. (1980). Identifying the organization of writing processes. In L. Gregg & E. Steinberg (Eds.), Cognitive processes in writing: An interdisciplina~a p p r ~ c h@p. 3-30). Hillsdale, NJ: Lawrence Erlbaum Associates. Hayes, J. R., Flower, L., Schriver, K. A., Stratman, J., & Carey, L. (1987). Cognitive processes in revision. In S. Rosenberg (Ed.), Aduances in appiied psychoii~u~ficsJ VolumeIi: R e a d i ~writ(pp. 176-240). Cambridge, England: Cambridge University Press. ing, andl a ~ uprocessing ~ e Hewitt, C. E. (1986). Offices are open systems. ACM ~ransactionson OfficeInformation Systems 4(3), 271-281.

N E ~ R T HET AL. Kaufer, D. S., & Carley, C. (1993). Communication ata distance: The inffuenceofprint on sociocultural o ~ a n ~ a t i oand n change.Hillsdale, NJ: Lawrence Erlbaum Associates. Kaufer, D. S., Hayes, J. R, & Flower, L. (1986). Composing written sentences. Research in the Teaching of~nglish20(2), 121-140. b y e , T.(1993). Computernetworkingfordevelopment of distance education courses. In M. Sharples (Ed.),Computer supportedc ~ l a ~ r a t iwri~ng v e (pp. 41-69). London:Springer-Verla~. Kistler, J. J., & Satyanarayanan,M. (1992). Disconnected operationin the Coda FileSystem.ACM Transactions on Computer Systems,IU(l), 3-25. Malone, T. W. (1988, February). What is coordination theory? In Coordination Theory Workshop. National Science Foundation, Washington, DC. Malone, T. W., & Crowston, K. (1994). The interdiscipl~narystudy of coordination. ACM Computi~Surveys,26(1), 87-119. McGrath, J.E. (1990). Time matters in groups. In J. Galegher,R E. Kraut, & C. Egido (Eds.), Intellectual teamwork (pp. 23-61). Hillsdale, NJ: Lawrence Erlbaum Associates. McLaughlin Hymes, C., & Olson, G. M. (1992). Unblocking brainstorming through the use of a simple group editor.In Proceedings CSCW '92 Conferenceon Computer~upportedCooperative Work (pp. 99-106). ACM SIGCHI & SIGOIS, Toronto. MinBr, S,, & Magnusson, B. (1993). A model for semi(a)synchronous collaborative editing. In 13.0ceedings of theThird ~ u ~ p e aConference n on Computer~upportedC ~ ~ r a t i vWork e (pp. 219-231). Milan, Italy. Neuwirth, C. M,, Chandhok, R, Charney, D., Wojahn, P., Kim,L. (1994). Distributed collaborative writing: A comparison of spoken and written modalitiesfor reviewing and revising documents. In Proceedings of the GW94 Conference on Computer-~umanInteraction (pp. 51-57), Boston, M A Association for Computing Machinery. Neuwirth, C. M., & Kaufer, D. S, (1992). Computers and composition studies:Arti~ulatinga pat-. tern of discovery. In G. E. Hawisher & P. LeBlanc (Eds,),R e i m ~ i n compute^ i~ and composition: Teaching and researchin the virtual age(pp. 173-190). Portsmouth, NH: Boynton/Cook. Neuwirth, C. M., Kaufer, D. S., Chandhok, R, & Morris, J. H. (1990). Issues in the design of computer-support for co-authoring and commenting. Proceedings of theThirdConferenceon Computer~upportedCooperative Work (CSCW '90) (pp. 183-195). Baltimore, MD: Association for Computing Machinery. Neuwirth, C. M., Kaufer, D. S., Chimera, R., & Gillespie, T. (1987). The Notes program:A hypertext d i ~ s345-365). Chapel application for writing from source texts. In ~ y p e r t ~ t ' 8 7 P r o c e e(pp. Hill, NC. Neuwirth, C. M., Kaufer, D. S,, Keim, G., & Gillespie, T. (1988). The Comments program: Computer support for response to writing.CECE-TR-3, Center for Educational Computing in English, English Department, Carnegie Mellon University. Newman, R., & Newman, J. (1993). Social writing: Premises and practices in computerized contexts. In M. Sharples (Ed.), Computer supported c~laborativewriting (pp. 29-40). London: Springer-Verlag. Norman, D. A. (1986). Cognitiveen~ineering.In D. A. Norman &S, W. Draper (Eds.),~ s e r ~ e n ~ e r e d system design (pp. 31-61). Hillsdale, NJ: Lawrence Erlbaum Associates. Olson, G. M., & Olson, J. S. (1991). User-centered design of collaboration technology.~ o u ~ofa l O~an~ational C ~ p u t I,i 61-83. ~ Olson, J. S., Olson, G. M., Stormsten, M., & Carter, M. (1992). How a group-editor changes the character of a design meetingas well as its outcome. In Proceedings CSW '92 Conferenceon Computer-Supported Cooperative Work(pp. 91-98). ACM SIGCHI & SIGOIS, Toronto. Posner, I. R., &Baecker, R. M, (1992). How people write together. In Proceedings ofthe25th ~ a w a i i International Conference on System Sciences(Vol. 4,pp. 127-138). Hawaii. Smith, J.B.,Weiss, S. F., & Ferguson, G. J, (1987).A hypertext writing en~ronmentand its cognitive basis.In ~ y p e ~ ~ t ~ 7 P r o c e e d(pp. i n g345-365). s Chapel Hill,NC.

17. DIST~BUTEDCOLLABORATIVEWRITING

Smolensky,P,,Bell, B., Fox, B., King, R.,& Lewis, C. (1987). Constraint-based hypertext for argumentation. In ~ y p e r t ~ t ’ 8 7 ~ ~ e e d(pp. i n g215-246). s Chapel Hill, NC. Stefik, M,, Foster, C., Bobrow, D. G., Kahn, K.,Lanning, S.,&Suchman, L. (1987). Beyond the chalkboard: Computer support for collaboration and problem solvingin meetings. Communications of the ACM, 30(1), 32-47. Tang, J.C., & Leifer, L. J.(1988). A framework for understanding the workspace activityof design i W ~ ’88 s Conl‘erence on ~omputer~upported Cooperative Work (pp. teams. In ~ ~ e e dB C 244-249). ACM SIGCHI & SIGOIS, Portland, OR.

This Page Intentionally Left Blank

C H A P T E R

Gary N. Olson Judith S. Olson

University of Michigan

Contemporary organizations are increasingly characterizedby more flexible structures thatmake use of continuing or ad hoc workgroups for accomplishing goals (Peters & at er man, 1982). A representative workgroup has fewer than 10 members, and works on their task for a period of weeks or months. Although the relationship between workgroup effectiveness and the overall productivi~of the organization is complex, it is likely that what makes workgroups effectivewill contribute to overall organizational effectiveness. Thus, we feel that one promising strategy for exploring the relationship between information technology and the productivity of knowledge workersis to focus on factors thatmight enhance the effectiveness of workgroups. There are both practical and theoretical reasons for this focus. First, workgroups are organizationally interesting as well as important. They constitute a key aspectof what have been referred to as “adhocracies” (Bennis, Toffler, 1980), a more flexible form of organization that is contrasted bureaucracies. Second, their members are often organizationally and ~eographicallydispersed, and represent several disciplines, meaning that problems of coordination, representation, and communication are centralproblems that new collaborationtechnol~gymight help solve. Third, workup activity is both natural and easy to study (McGrath, research feasible.Many workgroups last onlyas long as to complete, making it possible to study their entire life cycle. Small wor~grou~s can carry out a wide range of tasks. For instance,they might beasked to investigatea setof issues,to set policyon some matter, to

OLSON AND OLSON

formulate a plan for a project or activity,to design something, to implement a plan or design, andso forth. Tools to support these activities differ depending on who is in the group as wellas the different kindsof tasks the groups perform. Because we could not possibly take on this wholeofrange tasks in our research, we have confined our attention to theoftask design, in particular the designof software systems. We have focused on theearly stagesof system design and specification for several reasons.Much software is designed collaboratively, andat least since Brooks (:975), it has been widely recognizedthere that are large coora rich domain for dination costsin software designby teams. This seems like the investi~at~on of collaboration technology. Additionally, there are currently very few tools for this stageof software development, even though it is widely recognized as botha difficult and an extremely important stagein the software life cycle. We are engaged in a programofresearchtounderstandhow to build appropriate technology to support the work of small groups; We have adoptedtheapproach of user-centereddesignofthesecollaborativesystems (Olson & Olson, 1991).We use a number of empirical methodsto work toward in principles and theories that guide can our efforts.We study groups working their natural environments (e.g., small groups of professionals in companies) andgroupsworking in the laboratory (e.g., observing many small groups a meaningful is replica of real work). engaged in the same task, where the task Our empirical work is guided by a conceptual framework in which we systhe task, and the technology supportin stantematically describe the group, dard terms, varying some of these features, keeping others constant. The goal in which the relationships among these factors can be is to develop a theory explained and thus usedto guide the design of collaboration technology. ~ e w Groupwork o ~ ~ is the resultofmany factors, all of which interact to influence the behavior of the group and the content and form of the eventual product. Many researchers of group work (e.g., Forsyth, 1990;Hackman,1981;McGrath,1984,1990; ~~nsonneault & Kraemer,1989) agree that the major factors include:

featuresof the group’sc~aracteristics, such as size, history, norms, and member characteristics; the local si~ationincluding the technologies available from which the group gets support, the task they have to perform, including its ~ o ~ ainl which they work, including nature and difficulty; theo ~ a n ~ f f t context the useof a structured process like those brought in by facilitators.

Figure 18.1, based on this literature, illustrates the range of factors consi~ered im~ortant in determinin~group behavior and the group’s eventual success in doing a task as well as how they feel about the experience.

18, TECHNOLOGYSUPPORTFOR C O L L ~ O ~ T I V WORKGROUPS E

physical ~rangement technology suppon access 10 RSWrceS

x of activities needed attitude tow& gnxrp work attitude tow& this g m ~ p

FIG. 18.1. Conceptual framework for collaborative work.

What Fig. 18.1 shows is that conclusions about the behaviorsof groups in any particular situation, theoriesof group behavior, and the implications for technology design, must take into account the constellation of interacting factors. Because the range of factors is large, a research program must be explicit in its focus. Reading down the left-hand side of the figure, we focus on small groups(3-7) who know eachother and have worked together, and are of nearly equal status, working on a common design task, where the organizareal time. tional characteristics are held constant and the groups in work

On the other hand, we vary the technology.To date, we have studied people using such tools as whiteboards, paper and pencil, and a group editor called ShrEdit. Thecommunicationchannelshavevariedaswell. We have observed groups whoare face-to-face, distributed with audio only, or distri~utedwith audio plushigh quality video. Finally, we focus on the following measures: the quality of the work; the satisfaction with the process and the product; and the process by which the groups do their work. Our near term plans include co~lparin~ these results with behavior in large amounts of conflictful negotiation and training-mentoring. ~e are also in the midst of a study comparing people using a pen-

OLSON AND OLSON

based shared drawing tool within Aspectsto people using the LiveBoard, a pen based electronicwhiteboard-fli~hart.Our next study will include people work in^ with commercially available desktop video systems that show the other group members in less than life-size windows that are digitally transmitted, resultingin a delay and with fuzzy refresh. Our investigations draw from several disciplines. From the traditional literature on s ~ a i i g r o ~ ~ ~wee ~focus ~ v i primarily or, on how behavior is affected by the task itself. This is reflectedin our work in several ways. First, we chose for our initial focus a complex task, software system design. Because it embodies a wide range of specific activities, it gives us a good ofview collaborative problem solving. Second, influenced by the earlier work of Steiner (1912), Mc~rath(1984) and others, we have developed a taxonomy of tasks andactivitiesthatprovidesacriticalframeworkfor guiding ourwork (Teasley, Olson,& Meader, 1995) and leadingto theorydevelopment. ~ehavesupplementedthetraditionalideasfromthestudy of small groups with some related, but newer ideas. The offirst these is the view that ~ ~ t e ~ small groups at work must be looked at as a systemof ~ i s ~ ico~nitio~ ~utchins, 1990, 1991,1995;Norman, 1993; Olson & Olson, 1991). ~e focus on intellect~alwork that draws from the minds of the me~bersof the group andfrom the artifacts they use. Thesehelpthemremember,focustheir attention9and represent the ideas they discuss. Thisviewpoi~tleads us to analyze the detailed interactions of the group's joint problem solving and their useof the t e ~ h n o las o ~a coordinating artifact. ,where appropriate, we draw from traditionalcogniti~epsycholore to help us understand such factors as the attentional and working memory limits that can come into play as the representational and municative circ~msta~ces o heir work vary (Barnard, 1981;~ i ~ k e n s , r the limits of cognition, in looking bo port~niti~s for more communication among members through technolo-

'In all of our studieswe allow roups to engagein unconstrained verbalinteractions.

18. TEC~~OLOGY SUPPORT FOR C O L L A B O ~ T I ~ ~ O ~ G R O U P S

We also have adopted an overall research strategy that we think is criticalfordoingcarefulscientificworkonprocessesthat are found in real organi~ations.As shown in Fig. 18.2, we feel it is critical to study such phenomena through a linked approach using both field and laboratory work, Although this strategy is infeasible at the level of entire organizations, it is appropriate for studying workgroups. A key piece of this strategy is demonin the field strating that thereis a homology between work processes found We construct laboand those foundin the equivalent laboratory situations. ratory versionsof the tasks thatare done in the field, and study aspectsof of thesamebehavior in morecontrolledsettings,usinglargenumbers groups. This allows us to not only compare quality across conditions of technical support and communication channels, but then to correlate various aspects of the processof group behavior to satisfaction and quality. In both ourfieldandour laboratory workwehavemade a major investment in developing and using process measures so we can understand not only that work has changed, but how. In the remaining sections of this chapter we briefly su~marizesome of our major findings from our program of work studying technology su~port for collaborative workgroups. First, we describe some of the key findingsof our fieldwork on software system design. Second, we describe some of our laboratory work on design, first pointing out its homologies with the field data. ~e then describe our findingsin two studies, onevar support in face-to-face meetings, and onein which distribu techno lo^ work with open audio or audio plus video. Third, we describe enerally some ~ualitativeresults about a philosophyof tec t thatmay bear significantly on the prospects for enhancin~ ductivity of work~roups.

FIG. 18.2. Overall research strategy,

OLSON AND OLSON

E

We have carried out extensive analysesof data collected from naturalistic observations of system design being done in field settings.We have reported a number of results of these analyses elsewhere (Olson, Olson, Carter, & Storrgsten, 1992; Olsonet al., 1995;Olson,Herbsleb, & Rueter,1995),and interested readers can pursue detailsin these other sources.Here we focus in on some aspectsof the field data that allow us to discuss the homologies between field and laboratory settings. The field data were collected at Andersen Consulting (AC) and Microelec(MCC). We sampled 10 meettronics and Computer Technology Corporation ings from four different projects for intensive analysis. The meetings ranged in size fromthree to seven participants, and lasted 1 to 2 hours. Softwaresystem design was the principal topic of all of these sessions. The problems were large, requiring many indi~dualsmany monthsto design and code. The systems were both internal products (e.g., AC an systems analysis tool) and prototype ideas for future systems (e.g., a generic architecture for future In all cases applications and an exploratory system to edit knowledge bases). the problem specification was vague at best. The principal focus of the projects during the phase we studied was further development and refinement of the requirements. The major issues discussed included deciding both what features to offer the client and how to implement those features. Because the systems being designed were reasonably large and complex, a variety of expertises were required. Thus, the designers usually had different skills, including in many cases expertise on systems that were similar in some respect to the target system, oron the languages or architectures that the new system built upon. Further details about the groups and the projects are (1992). found in Olson, Olson, Carter, and Storrgsten Each of the 10meeting’s videotape was transcribed and marked with time stamps. Because we were interested in the moment-by-moment activity in these meetings, we developeda new coding scheme that wasa blend of the ~ e s ~ ~ ~scheme ~ u ~in eDesign n tRationale aclean, Young, & Moran, c t i ~ i ~ of Putnam (1981), Poole and Hirokawa 1989), the~ e c u t i ~ e a categories (1986) and Poole and Roth(1989), as wellas more standard measures of partici~ation(Mc~rath,1984). The analysis we focus on here was based on coding the transcripts into 22 categories that captured the general nature of the design discussions. One set of categories focused on the direct design discussions themselves: Various issues are raised either implicitly or explicitly, for each issue various alternati~edesignpossibilities arepresented.Eventuallyadecisionamong these possibilities must be made by applying various criteria that help select the preferred alternatives.

18. TECHNOLOGY SUPPORT FOR C O L L ~ O ~ T I V WEO ~ G R O U P S

Other categories dealt more with how the groups organized their work. These include such coordination activities as discussion of goals,project mana~ement,and meeting management as well as summaries, walkthroughs, and ~ ~ r e s s i o n s . We also found substantial amounts of clari~ca~on in these discussions. This category was defined very strictly as answers to explicit questions, Wherever possible we associated this clarification activity with the other activities, leading to such coding categories as clari~cationof issues, clarification ofgoals,and so forth. There were two special clarification categories: Clari~cationar~ifact refers to those occasions whena meeting participant tried to explain some aspect of a drawing, a list, or other meeting artifact, and clari~cationgeneral refers to those remaining clarification episodes that could not be associated with any of the other specific clarification categories. Periods of activity thatdid not fit into any of our categories were putin a catch-all class calledother. We achieved satisfactory degrees of interrater reliability in coding our transcriptswith these categories. (See details in Olson, Olson, Carter, & Storrarsten, 1992.) in our sample One of our more striking findings was that the 10 meetings were allvery similar to each other, both in the amount of time spent in these categories andin the patternsof transitions among them. The only obvious difference that emerged was that the meetings varied in the amountof project m a n ~ e m e ntime t they contained, ranging from none as to much as 40%. of Figure 18.3 shows a composite picture based on all 10 design meetings both the use of time and the pattern of transitions among the coding categories. For each category, the size of the circle represents the amount of in the cattime spentin that category. The whiteareas represent direct time egory; the black areas represent the associated clarification time. For the transitions, the thickness of the arrows corresponds to the frequency of the transitions. We only show those transitions above a certain minimal threshold in order tokeep the diagram simple. The locations of the heads and tails in of the transition arrows show explicitly what coding category is involved the transition. These discussions were clearly anchored in the design categories. For instance, there were frequent transitions among core the design categories, are especially in andoutofCriteria.CriteriaandCriteriaClarification involved in 46.1%of all transitions among the 22 categories. Alternatives and their Clarifications account for another large portion of transitio~s:40.9% (this includes the transitions with Criteria, so this figure overlaps with the previous one). Together, Alternative and Criteria along with their associated Clarifications accounted for 68.2% of all transitions. In other words, these four categories from the total set of 22 categories are involved in over two thirds of the transitions. This is in contrast to time use: These four cate-

OLSON AND OLSON

FIG. 18.3. Time use and transitions among coding categories of design activity. From "Small Group Design Meetings:An Analysis of collaboration^ by J. S. Olson, G. M, Olson, M. Carter, and M.Storr~sten,1992,~ u ~ a n ~ o ~Inter~ u t e r action, 7, pp. 347-374. Copyright by Lawrence Erlbaum Associates. Reprinted with permission.

gories only occupied36.5% of the total time. Thus,the design meetings we observed had frequent although brief excursions into discussions bearing on designalternatives and criteria for selecting among them. ~e discovered throughthe use of lag sequential analysis(see Olson et al., the dashed line in Fig. 18.3and the 1995) that the six design categories above remaining 16 categories below the line constit~tetwo functionalls distinct categories. ~e examined the statistically significant sequences in our data for lags up to 5, and we found that there were 25 si~nificanttransitions among the 6 design categoriesof issue, alternative,criterion, and their associated clarifications, but only 2 from these categories to the remaining 16.

18. TEC~NOLO~Y SUPPORTFOR C O L L A B O ~ T~I ~O ~ ~ R O U P S

Similarly, the remaining 16 categories had 79 significant transitions among themselves but only 11 with the 6 design categories. To put this finding a bit less technically,it means that these design sessions consisted ofinterleavedepisodesofdesigndiscussion(the 6 cateWe gories)and of coordinationormanagementdiscussion(theothers). attempted to characterize the sequential structure of the design discussions by collapsing all 16 nondesign categories into a single category called manage~ent,andthenexploring in detailthesequentialorganizationofthe resulting encoded transcripts. Again, to our surprise, wefoundmuchsimilarityamong the 10design meetings. Figure 18.4 shows the significant conditional probab~litiesam on^ the seven categories used in this analysis, and suggests the orderliness of the design discussions,In order to capture higher order sequential dependencies we used an iterative process of rewriting strings of category syrnbols into higher order categories, writingif you will a grammar of design activity. Interestingly,thesamegrammarprovidedanequallygoodfitof the 10 design meetings. The main differences among the meetings were parametric, in that the NICC meetings tended to have longer sequences of design activityandfewermanagementactivitiesthanthe AC meetings.But the important point was that bothsets of meetings had the samestructure, that is, they could be describedby the same grammar, See Olson et al. (199~)for technical details. On the surThese were all very informal, intensely interactive meetings. in or~ani~ation. But our analyses tellus a face they seem to be quite chaotic different story: Design discussions at the level we described them are quite why this may beso. First, structured and orderly. There are several reasons it may be that design as a task is relatively structured, and thatin order to do design at all one must proceed with some degree of sequential orderliness. Second, our designers were quite experienced. The AC designers had

FIG. 18.4. Significantconditional probabi~itiesfortransitionsamongcoding categories,

OLSON AND OLSON

beentrainedonhow to dodesignandconductmeetings,andthe MCC designers had considerable experiencein industry prior to coming toMCC. Third, the orderliness may have been due to the high level of our analyses, without concern for who said what or how the issues interplay. The heart of the design process is the generation of design ideas, discussing their pros and cons, and selecting among these ideas those arethat to be potentia~lyincluded in the final design.We looked at the structureof theargumentsthatwereconsidered in thedesign ~eetings,usingthe Issues,Alternatives,andCriteriacategorization of designprimitivesto describe the content of the design processof our 10 meetings.We were interested in seeing when and how completely the designers stated the design why each questions explicitly, listed alternative solutions, and gave reasons alternative was good or bad. For each meeting, wedid a complete design rationale graph, a portion of of the meetwhich is shownin Fig. 18.5, Of course, the actual graph from one ings is large, sometimes taking up the better part of a wall. In Fig. 18.5, the off the issue, and the criteissue is notedin the left, the alternatives radiate ria are to the right.A criterion statedin support of an alternative is shown by a solid line,as opposed to the dashed line for a negative criterion. The median number of issues considered in a meeting was 10. There was in a meeting, rangconsiderable variationin how many issues were examined ing from one meetingin which only one design issue was discussed during the entire meeting, to another in which 44 different issues were addressed. The meeting with one issue focused on the feature of the current proposed design that was responsible for the greatest complexity, six.and alternative solutions were raised and discussed at length, In the meeting in which there were 44 issues, the participants reviewed several documents that recorded the current stateof their part of a design of a verylarge system architecture.In the process of reviewing the items recorded, new issues and alternati~esarose.

FIG. 18.5. Fragment of a design rationale.

18. T ~ ~ ~ ~ O LSUPPORT O G Y FOR ~

O

L

~

O~ O~~ eTR Ol U ~P S

The most typical issue had two alternatives discussed.Twenty-one percent had one or fewer alternatives raised (e.g., “We’ll offer only printing of the whole diagram^'), whereas 40% of the issues had three or more alternatives. These could be extensive discussions or limited ones. Clearly some issues get lots of attention In terms of an elaborate exploration of alternatives, whereas others are raised and not exploredin very much detail. Overall, only 63%of the alternatives raisedin the discussionof issues were evaluatedat all. Saidanotherway,morethanonethird of thetime,when positions are raised there is no explicit evaluation. The fact that the 10 design meetings are more similar than different is perhaps our most surprising finding. We expected much less uniformityin the groups’ behavior, due to differencesin organizations, projects, and participants. Based on ourrelatively large sampleof 10 meetings analyzedin great detail we are inclined to conclude that there is substantial regularity to this behavior in general. These initial field studies led to two other lines of fieldwork. First, in an examination ofwhatkindsof support designers might need, perhaps in a design database, we examined the kinds of ~ ~ e s t asked ~ o ~ in s design meetings (at Andersen Consulting) or questions or issues recorded in detailed minutes (atNTT in Japan) To our surprise, mostof the questions hadto dq withunderstand~ngtheuser’srequirements,notthereasonsforearlier design decisions, and the patterns were identicalin both sites (~erbsleb& Kuwana, 1993; Kuwana & Herbsleb, 1993). Second, our original field data were collected for teams using traditional software engineering methods.We have used it to help us understand the effects of another engineering method, Object Oriented Design(OOD). This work was comprised ofboth direct observations of designgroupsusing object-orientedmethods in oneorganizationandinterviewstudies of a much wider range of organizations (Herbsleb et al., 1995). The OOD groups that were examined in detail showed that unlike the traditional meetings, they were led by a chief designer, but the design discussions were very much like the traditional ones.

~ualitativeanalysis of our field groups’ behavior showed that they had difficulty in coordinating the contentof their discussions. For example, there were cases where: ~ o r ~ on n paper g copiesof a dia~ramof the system architecture the group members had agreed on to date, the group discussed some significant changes to that architecture and subsequently marked off ~ i ~ easpects ~ e ~oft the diagram.

OLSON AND OLSON

~ o r ~ on n gthe whiteboard with a olist f open issues and people who would be responsible for follow up,a group stopped their meeting when the whiteboard was full, and furthermore had greatd i f f i c u l ~discussing potential interactions among the items because the items couldn’t be moved,

Based on these and other observations (Olson, Olson, Carter, & Storrrasten, 1992)andouranalysisofavarietyofextantgroupwaresystems(Olson, k, & Wellner, 1990)) we designed and built a group editor called Guffin & Olson, 1992). The goal wasto provide a simple tool that allowedpeople to view and edit a shared object (list, diagram, etc.). We began this development effort withthe goal of merely exploring architectural issues about speed, concurrency control, and reliability, butend we in the built an easy-to-use, useful tool for the highly interactive phase of early system design. ~hrEditis a simple text editor that allows all participants to view and edit the same documents. Everyone can type simultaneously within one character position of each other. The individuals’ views of the document are independent, but can be coordinated if the discussion requires common focus. Weviewit as a simple shared el ronic workspace, an example of a “permissive” techno lo^ (Galegher & ut, 1990). This contrastswith a large set of group~aresystemsthat“prebe”anorder ofactivities,regimesfor turn-taking,access control, and so on. We have opted to build and explore because we are interested in how groups adapt to

use it successfullsin a vari t to more than 80 sites wor

1993).

18. TECHNOLOGY SUPPORT FOR C O L L ~ O ~ ~T O~ ~EG R O U P S

or

The field data we have collected and analyzed are rich and interesting. They help us understand the details of the kind of problem solving that constitutes design. Our findings of similarity among meetings and overall structuredness in the activity suggests that some aspectsof the process of collaborative designmaybesomewhatinvulnerable to the socialand organizational settingsin which this work takes place.But we continue also to be interested in whether aspects of the problem solving that constitute collaborative design can be studied under more experimental settings. There are two key reasons for wantingto bring this task to the laboratory, both related to the traditional virtuesof experimental methods. First,in the field it is very difficult to say much about what activity leads high qualto ity products. It is tricky to evaluate the outcomeof group projects as each project has a number of unique features. In contrast, in the laboratoryit is possible to give a large number of groups standard problems to work on where the qualityof the outcome can be objectively measured. A second reason for pursuing these issues in the laboratoryis that these more con~rolled settings make it possibleto understand what happens when new work techno~ogies are introduced. The technologies themselves can be varied in ways that allow a deeper ~nderstandingof what features of the techno lo^ contribute to what features of the changes in process and outto put new technolo~iesinto the work setting before their In turning to our laboratory research, we first describe what our lab ex iments are likediscussthequestionofthehomologiesbetweenthe lab and the fie1

OLSON AND OLSON

We have also developed a task that is a realistic design task, but does n requireanyspecializedknowledge.Teams of designersfromcompanies have said this task realistically mimics aspects of their work.We embed our task in a scenario about the size of the team available for implementing the design, an overall timetable, and the target audience. We also give the group a specific goal for the9~minutedesign session in which they work,so each group knows whatkind of output is requiredat theend of the session. They are to begin the design, assess the features the design will offer to the user, how it will work, equipment needed, andso forth and to make notes on this discussion thatare readable by a fictitious group member who was not present at the meeting but would be at the next. The biggest difference between the field and the lab is the organizational context. In the field our groupsare doing work on a project that matters for their company, real company resourcesare at stake, and the work playsa role in the evolving careersof the group members.We could not study the in the laboratory.On effects of organizational factors on collaborative work the other hand, if there are features of their work thatare not much affected by organi%ational factors, such as the orchestration of their problem solving as revealedin the fieldwork we described earlier, then these features might be amenable to laboratory investigation. ~ ~ We s .look

at the problem-solv-

gh the kind of coding we described

in Figs. 18.3 and 18.5. Our claim is that we can demonstrate sufficient homolo-

gies between lab and field with respect to these behaviors (which we have ady shown in the field are relatively insensitive to organi%ational set,and thus study the effectof various laboratory manipulationson the nature of these behaviors, Furthermore, because we can score the ofquality the outcomes in the laboratory, we can assess the effectof various factors on the qualityof group work. In all the conditions we report, we have groupsof three perform a standard design task that lasts for 90 minutes. Prior to this they have performed several otherwarm-up tasks so they can adjust to whatever working conditions we give them. Croupswho are workingunderconditionsdo so in a meeting room called the Collaboration Tec te (CTS; Olson, Olson, ~cCufack, Cornell,& Luchetti, 1992) in which computer displaysare in units in the table top that make up a conference table, with the displays recessed so as not to interfere with normal eye contact, shown in Fig. 18.6. All sessions are videotaped so we can analyze themin much the same way that we have analyzed our field meetings. We have run 19 control groupsof 3 who work ~ithoutany new technoloThey work in the CTS just like other groups, except we turn off the com-

18. TEC~NOLOGYSUPPORTFOR C O L L ~ O WO~GROUPS ~ T ~

FIG. 18.6. Collaboration Technology Suite atthe University of Michigan.

puters and have them use paper and pencil, and the whiteboards, ~e transcribed their videotapes and coded these. Figure 18.7 shows a side-by-side are a few plot of the behaviorof these lab groups and the field groups. There new categories here that are associatedwith the specifics of the lab task. The most important is talk about the writing (plan and write) and the actual writing timein which no one is speaking (pause). The diagrams show that there are some differences. First, as mentioned, the lab groups spenda lot of time writing; we have deleted these pauses and the associated conversation directly about the writing from the diagrams. Another m~agementtime. There is essentialway in which the two differ is on project ly none of this with the lab groups, because there is no real ongoing project at stake.If we exclude project management and writing time, the time in the spent remaining categories correlates .70 between the lab and field groups, suggesting they are using their time similarly. The story is similar for transition times. The only noticeable differenceZs in the amount of clarification. This is due to the fact that the membersof the lab group all were familiar with the problem domain (post offices), while the field groups are often put together becauseof their complementary expertise. Thus, at the level of behavioral description provided by Fig. 18.7, which focuses on the wayin which the groups use their time and structure their activity, the two situations are quite similar. A comparison of the overall issue-alternative-criteria dia~ramsshows that the lab groups raised about the same number of issues per hourin the twosituationsandmadesimilarkinds of evaluativearguments.Thelab 7 alternatives raised per groups explored the issues more broadly, with issue in the lab and only2.5 in the field. ecause the laboratory task had people generate lists of features, benefits, and costs, it is reasonable that they would consider more such alternatives.

18. TEC~NOLOGYSUPPORT FOR C O L L ~ O ~ T I V E ~ O ~ G R O U P S

It is important to be clear about what claim are we making. Despite all of the obvious differences between the field and the lab that we described earin the field and the lab organize lier, we have just seen that groups working their activityvery similarly. This,of course, focuses on the problem solving they are doing as partof design, It is at this level, and perhaps at none other, that the two kindsof groups are homologous, This encourages us to look at factors that might affect their problem solving, such as new tech~ologies.

run another 19 groups in a condition in which they use ShrEdit to do their design task. These groups were trainedin the useof this tool, used itin several warm-up tasks, andthen did the same standard design task as the unsupported groups. There are many results to lookat in comparing these two sets of 19 groups, but we focus briefly on only four. (Details are in Olson, Olson, Storrersten, & Carter, 1993.) First, the groups working with the new electronic workspace were slightly less satisfied with their process than those working in the more traditional mode. Because this was a new way of working, this is perhaps not surprising. Second, when we evaluated the final designs of the 38 groups for quality, the groups working with ShrEdit produced designs that were significantly better in quality. Quality was assessed via an objective rating scale that had very high interrater reliability. Third, an analysis of the flow of activities indicated that the supported groups used their timein ways that were similar to unsupported groups, but with some small differences.An analysis of communication processes in the group shows that they spent more time writing and less time talking. Furthermore, they summarized their work less, likely using the material they had written (the ShrEdit documents) serve to the same purpose. Also, combining the timing data with the content analysis, we found that the unsuppo~ed groups repeated their ideas more; for the ShrEdit group, these ideas were Fig. 18.8). already down in the document and even often elaborated on (see Fourth, we foundthat the groups using ShrEdit, contrary to our expectations,exploredsubstantiallyfewerdesignideasthantheunsupported groups. They considered fewer alternatives and evaluated these alternatives less. Closer examination showed that the larger setof ideas explored by the unsupported groups was ‘‘off target”; they explored ideas that were not centralto the design task given. When confinedto the coredesign task, both groups were the same. ShrEdit played a big role in the group process. Ten percent of the time, two or more of the people in the groups with ShrEdit did type simultaneously, with typically sharp burstsof this early (for brainstorming) and late

18. TECHNOLOGY SUPPORT FOR C O L ~ B O ~ ~WORKGROUPS V E

(fin~shingup the written text). The middle of the meeting was dominatedby discussion and one-at-a-time recording of the ideas discussed. In contrast, the whiteboard groups usually were by leda person at the board, who wrote the ideasas they arose.Then, late in the session, someone wrote a summary of these noteson paper. Whereasin both groups, someof the ideasin the final document were never spoken about, in the ShrEdit groups, all had a chance to see or read them, whereasin the whiteboard group, they were solely authoredby the scribe on paper(Storr~sten,1993). This study revealed that a very simple collaborative technology can have a very big set of effects on group work. Although groups using it the for first time feela bit uncomfortable about it, the technology helped them produce better work. And it did so in surprising ways, by apparently encouraging them to focus in on g0o.d ideas, and not waste their time exploringa wide range of less effective ideas.

Traditionally, all forms ofwhatwethinkof as meetings took place face-toface, in meetingrooms,commons areas, andlunchrooms.But as weail know, groups no longer need to meet in the same location; new technologies are allowing us to relax the constraintof co-location. Modern telecommunicationsmakesaninteresting array of options available: teleconferences, video conferences, ands ~ c h r o ~ ointeractions us over computer networks. These alternatives to face-to-face interact~ons have distinctive properties, and have not,in any real sense, replaced what it is possible toindoface-toface interactions. These new technologies offer the abilityof enterprises to organize in new ways and to capitalize on new flexible possibilities. Yet we need to understand better what the op~o~unities and constraints are that are offered by each mode of synchronous interaction. Modern networking has brought about the possibility for these small groups to be locatedin different places while working.A fundamental consideration in ShrEdit's design is that we assumed that the members of a group would always have other communication channels available to them. In a face-to-face setting,of course, the groups can talk and gesturein their usual interactive ways. indeed, the groupsin our studies engaged in extensive discussion while using ShrEdit as a workspace to capture and revise ing ideas. So our next step wasto investigate the useof ShrEdit by distributed groups, where, in asthe face-to-face setting, ShrEdit would be used to share the group's emerging design ideas but we would provide for them other communication channels for talking and interacting.

OLSON AND OLSON

We decided to provide communication for our groups thataswas ideal as we could make it given their distributed set-up. We did this because we wanta baseline for later studies that looked at other kinds of communication, and in particular digital desktop video.In the present study we focused on how groups of three performed when they have a shared workspace tool and nearly ideal communication. A number of sources indicate that high quality audio is important to syn& Mackay, 1993; Tang & Isaacs, 1993). So half our chronous work (e.g., Pagani groups worked withhigh quality audioin addition to the shared workspace. of far Our audio was full duplex, directional for both input and output, and in teleconferencingor most commercial video conbetter quality than found ferencingsystems.Morecontroversial is whethervideoaddssignificant value for groups doing distributed problem solving.Alth~ughthe research record is quite mixed (Egido, 1990; Tang& Isaacs, 1993), many theories (e.g., Daft & Lengel, 1986; Rutter, 1984; Short, Williams, & Christie, 1976; Weick & Meader, 1993) and most people’s intuitions are that video should add substantial value to such work. Thus, the other half of our groups had our good quality audio plushigh og video connections to each of their colleagues.The video was an optimal fashionto create the feeling of sitting arounda table with one’s colleagues, with the shared workspace in the center as shownin Fig. 18.9. We took more care than usual to create what we felt would be the best possible video conferencing set-up. We were interestedin how these video-audio groups using ShrEdit would compare to face-to-face groups using ShrEdit from our earlier study (Olson, Olson, Storr~sten, & Carter, 1993) on the same range of measures: quality of the work product, satisfaction, and characteristics of the group process.We also compared these audio-video groups toaudiwnly the groups to assess the added effect of the video. What distinguishes our study from previous investigations is the use of an established workspace tool of known value for sharing the work, the fact that the people in the study comprised intact groups, and the fact that we atook great dealof care to ensure that the audio and video were of the highest quality we could get with present communication technology. Another functionof this study was to establish a baseline from which we could conduct later studiesof a varietyof less-than~ptimal communication technologies. ~uitimediadesktop conferencing systems that run over the Internet or ISDN lines (e.g., Intel’s Proshare, SunSolutionst ShowMe, and AT&T’s Vistium) are generally quite constrainedin the quality of the audio and video they can provide. With our baseline data we can assess these situations in future investigations. The subjects were36 existing groupsof three professionals, drawn from the same populationas thatin our lab studyof face-to-face work. As shown

18. TECHNOLOGYSUPPORTFOR COLLAl3OMTm WORKGROUPS

FIG. 18.9. Sketch of the lab set-up for the videosupported condition.

in Fig. 18.9, they workedin three separate rooms, each of which had a workstation with a large screen centeredon a desk, with two 13" video monitors on each side of the screen. A camera and microphone were mounted on each video monitor, with the camera placedat the centerof the top of the monitor so that when the participants faced each other, they appeared to each other to be making eye contact.2 Furthermore, when the other two remote participants were facing each other, their images projected to the receiving participant made them look as if they were looking at each other. The microphones and speakers were similarly situated to either sideof the central screen, corresponding to the person shown on the video screen, They were open fuil-duplex channels that additionally projecteda senseof spatiallocation.Indeed, in theaudio-onlycondition,groupparticipants often moved their heads to face the speaker boxesof the people they were addressing. The audio condition used the identical microphones and speakers of the video condition; the only difference was that the video monitors were turned off. As in the earlier study, the groups all used ShrEdit. We trained all groups in ShrEdit in the CTS then took them to the distributed rooms where they worked. The groups performed the same three tasks as in the earlier study, 20 minutes each and the target design task for 90 minutes. In the first two for total, we ran 36 groups. Eighteen of the groups used the full video and audio As in the earlier study, technologies to communicate; 18 had only the audio. we assessed three things: the quality of the product, the participants' satisfaction with the process, and the process of design and coordination. (More details are available in Olson, Olson,& Meader, 1995.) __

"

zEye contact was not perfect. Participants reported that the other person appearedto be looking at their throat when they looked into their eyes.

OLSON AND OLSON

The results showed that when remote groups work with a flexible, simple shared editor to support their design work and communicate high via quality video (eye contact, spatial relations preserved), the quality of their work is not significantly different from similar groupswor~ngwithout the video, but withhigh quality, stereofull duplex audio. Furthermore, the video groups are not significantly different from similar groups working face-to-face with thesamesharededitor.Face-to-facegroups aresuperiortotheremote groups without video, however, suggesting therea small is detrimental effect overall when groups work without seeing each other. Figure 18.10 summarizes the quality results for both the face-to-face and remote conditions. More striking, however, is that the groups working at a distance without video do not like it as much as those who have the video. They reported beinglessabletotellhowtheirothergroupmemberswerereacting to things said. They also reported that the communication system got in the wayof their being able to persuade others about their ideas or to resolve disagreements. Tang and Isaacs (1993) found that groups in a field setting who were offered video in addition to shared workspace and audio used the system more heavily than those who had audio and workspacetools, suggesting that the preference we saw in our study could be a harbinger of usage patterns when these capabilities were discretionary. These meetings were then coded for their time use as before.Their patterns were very similar, with the only difference being that the time spent talking about the issue at hand was significantly longer in the audiwnly condition, perhapsa measure of grounding. However,in comparing remote work in remote work required more time to with that face-to-face, people engaged manage their work and to clarify what they meant. This suggests that the video we hadset up, although intendedto duplicate theimpo~antaspects of face-to-face,did not succeedin replicating that information. Perhaps thereis more senseof what others aredoing and what they mean when weare faceto-face than can be presented via even very good video channels. FTF

Remote Audio

Unsup. 54.7

I I

" " "

=

=

56.7

I " " " " 1 ) " -

61.5

" 1 ) " " " " " -

p cii, for any (i, j) E T, and j

iVi

,

where 6i,i+1 is thebe~intime to process e er at ion (i, j+l),and c,,the completion time of operation (i, j). The completion time ciican be calculated from in time bii,i.e., cij= 6, +- “ P,,where fr is the time required to process the operation.(Notet, is thesameforalloperations.)Theprecedencecon§traints(3) thus reflect the inter~ependencies among DMs’ decision§*

WANC, ~ E I N M A NLUH ,

~ o u n d co~s~aints: a~ To obtain the valueof task i, the begin timeb, of operation (i, j ) must be withina range specifiedby the earliest allowable time e,] and the latest possible time liltthat is, eij2bii2I,, (i, j)

E

T.

This set of constraints is actually implicit in the formulation (1)-(3). Itis made explicit here for later analysis of the ~ u ~ decision-making a n process. In the inequality (4), the eifand lijcan be determined from the task arrival time ai (or deadlinedi),the operation numberj, and the last operation4of task i: a

eij =a,+(j-l)*t,, l,=d,-(N,-j+l)*t,

(i, j) ET.

The team problem can now be formulated as selecting the begin times {b,) to maximize the team reward J,subject to the resource constraints (2), precedence constraints(3), and boundary constraints (4). Once the problem is solved, tasks can be processed according to the task processing schedule, which is optimal for the current snapshot. Once there is a new task arrival, the task processing scheduleis obsolete. A new snapshot is taken, and the team problem is resolved to generate updated schedules. The entire dynamic task processing problem is thus solved on a snapshot-by-snapshot basis (i.e., in a rolling-moving window fashion). For each particular snapshot, the statistics of future task arrivals are not considered.A typical snapshot commonly involves four to six tasks for each DM. The problem solution thus involves a planning horizop of four to six decision steps. 'Whether the future task arrivals are considered or not has little impact on the resulting decisions (as shownin solutions in Section 5), due in large part to the fact that the task processing times are required short relative to theplanning window, and futurearrivals can only be estimated statistically. roac

In the earlier formulation, decisionsof different DMs are coupled through the precedence constraints (3). These precedence constraints can be relaxed by usingLagrangemultipliers todecoupletheinterdependencies among DMs' decisions. The problem can then be solved iteratively by generating decision options for different DMs independently, checking the constraintviolations,evaluatingtheconsequences of theseviolations,and (refining)thedecisions. * ~pecifically,we introduce Lagrange multiplier A, to relax the constraint between operation (i, j) and operation (i, j+l). The resulting Lagrangian to be optimized is

23. ~ O D E L I N T ~ W COO~INATIO~

The first summationin (6) is the total task value that can be obtained by processingtaskswithintheirtimewindow,Thesecondsummationreflects penalties caused by constraint violations.If a precedence constraint is violated, e.g.,bi,j+l-cij 0, the resulting multiplier will be positive, A, >0. The violation of a precedence constraintconse~uentlydecreases theLagrangian L. Intuitively, the Lagrangian L is a composite criterion reflecting the gains obtained by timely processingof tasks minus the losses resulting from con. violations. straint With L as a criterion, decisions bijof different DMs can be madeseparately as the precedence constraintsare no longer considered as hard restrictions. Mile it is easy tofind bijfor eachDM, the precedence constraintsare usually not 'maintained, Indirectly, the Lagrange multipliers {Aii} penalize the violationsandreinforcetheconstraints. Now theprobleminvolvestwo aspects: selecting the begin times {bij}to maximizegain and determining multipliers {Aij} to minimize constraint violations. Mathematically, this leads to the following problem {Ai;,4;}=arg min max L. Ai)

~ubstituting cif= b,,

f

fr into

bi)

(7)

the expression for L in (6) results in

Let Tp denote the subset of task operations under the responsibility of DMp, p = 1,2,3. The total task operationset T is the union of the subsets,T = T,uT,uT3.Through simple manipulation, the decisionvariables {bidin the Lagrangian L can be grouped according to the subsets, and decisions of different DMs become "separated"or decoupled, that is,

where

The expression L, involves only the decisions of DMp. Because all decisions {bidare grouped intosubsets, the m~imizationof L with respect to {l+,} can be performed separately on each of the L,. Consequently, via (g), the problem (7) becomes

The structure of Eq. (11) suggests an iterative solution procedure.Given can be determined a set of beginningtimes {bid,the Lagrange multipliers{Aij} L.Then giventhe {A,,}, newlrevised decisions{l+} of each DM can to minimi~e be made to maximize the decomposed objective L,, p = 1,2, 3. Mathematically? this procedure is expressed by the following bilevel scheme: high level:

{A!;} =arg

min L, [nIJ)

low level: {b;j=argmax L,, p = 1,2,3, subject to (2) and (4). lbIJ)

vel problem can be solved by using the subgradient method the multipliers according to the degree of constraint violalow-level problem is subject to the resource availability conboundary constraints (4), and can be solved by using the ing the individual problems at the low level. S are optimalwith respect to the decornnse~uencehas to be evaluated by conside high level. There is an iterative process ggests-albeit normati~ely-that the team internaliterations(optiongeneration,

1 information of the environment (status of e s f f ~team e problem in the same way

h the same team solution. Team mem-

23. ~ O ~ E L I NTEAh4 G COO~INATION

information nor sufficient capability to reach a unique optimal solution, they

will not be able to predict each other’s decisions precisely, and the modelproduced decisions for differentDMs will differ in detail. As described in Subsection 3.2, DM1 can observe the iconsof team tasks ( T T s ) and his own individual tasks(ITS), but only knows the aggregated value of the otherDMs’ indi~dualtasks. DM1 thus hasuncert~nty about the precise state of others’ ITS, rendering DMl’s submodel different from those of and/orDM3. The distiaction among DMs’ submodels is further elaborat Section 5, where descriptive factors are introduced into the normative model. Connectingthesubmodelsforindividual DMs withcoordinationlinks leads to a~ i s ~ i 6model ~ t eof~ the team (also calleda distributed model, for short). Figure 23.4 illustrates the distributed model structure. As shown in Fig. 23.4, the team asa whole receives information about the

task environment as its input, and makes decisions to process the tasks as its output.The team members, including the coordinator, are modeled as different submodels with different local information and different responsibilities in decisionmaking.Thesubmodels interrelate via c o o r ~ i ~ f The f~o~. coordination 1inks.among the submodels involve both implicit and explicit mechanisms. Because each DM’S submodel predicts other DMs’ decisions via solutionof the team problem, the distributed model of the team contains a certain degreeof implicit coordination (no inter-DM communication). The degree of implicit coordination depends on the amount of information available to the individual submodels to solve the team problem. The uncertainty a submodelhasaboutother S’ individualtasksaffectsitsredictions of others’ decisions. Consequently, implicit explicit coordination is needed. The distrib sages to each other to announc isions {h& thus he1 dinate the decisions of the team challenge is to brin communication above the into frame~ork,includin determine communication eeds and to represent formatted

”

FIG. 23.4. Distributed model structurefor team decision making.

o ~ m u n i c ~ t i oinnt

For a given snapshot, the submodel for eachDM solves theteam problem based on its locally available information. As discussed in [31], the value of a Lagrange multiplier reflects the tightness of the corresponding constraint of the task in question. A large multiplier value impliesa high sensitivity of team performance to the constraint, and a large performance degradation will incur if the constraint is violated. If communication can be used to avoid the violation, a large Lagrange multiplier value implies large potential benefit of communication on that particular task. It is thus reasonable to assume that if a limited amount of coordination effort is available to reduce constraint violations, the effort should be allocated to handle constraints with high multiplier values. Therefore, the multiplier values can be used to provide arank~rderingas towhich task operations have the tightest time windows, andare likely to benefit most from communications. ~ommunicationcan be treated as the exchangeof decision information among DMs on a task-by-task basis. For example, the submodel ofDM1 will solve the team problem based on the decentralized information DM1 has, and it can “inform” the submodels of other DMs about its decisions{l+*].The communicated b,’ for operationj of Task i may bea processing decisionDM1 just implemented(bii10), which corresponds to a “Plan” message. When a task operation cannot be processed (not enough time available), the task will be ignored by the corresponding model (setting 6, > di)and an ignore^ message will be sent. The “Request” message is not incorporated in our modelingas it was rarely used by human teamsin the e~~eriment, (welltrained DMs in a team will anticipate needsof other DMs, and hence unsolicited information transfers dominate [111). By interpreting the communication messagesin the ways given earlier, the model is able to incorporate the use of different message types. A second modeling issue is the determinationof to whom the messages should be sent. The experiment showed that communication messages were sent mostly fromDMs responsible for preceding operations of team tasks to DMs responsible for subsequento~erations(forward flow), as ne was not feasible.We thus assume that communication messages are sentin a forward flow, that is, the submodelof DM1 sends the message onb,‘ to the submodel of the DM responsible for the subsequent operation (i, j+l). The received messageswill be usedby a submodel to adjust its decisions. A “processing” message(bii= 0 or b,,zs 0) indicates that operation (i, j) is being processed or will be processed immediately, and cannot be resched-

23. ~ O D E ~ NTEAM G COO~I~ATION

uled. The submodel will then set the earliest allowable time to process the subsequentoperationtobeoneoperationprocessingtimelater, = f r . If a "plan" message (bij 0) is received, the submodel will use the communicated in place of itsestimated b, forprocessingoperation (i, j). If an "ignore" message(bij >d,) is received, the submodelwill consequently discard task i (negotiation is not considered). By virtue of the communicated information from other DMs, a submodel is able to adjust someof its decision variables which had been previously estimatedin isolation. Therefore, communication should lead to more highly coordinated decisions among the submodels when they have diverse solutions (e.g., causedby different local information at theDMs). The model does not resolve the problem withina snapshot after receiving communication messages. In other words, the model does not iterate internally on communication. This is because (a)as time passes, communication affects future decisions rather than current ones (whichare implemented right away), (b) any decisions to be implementedin the futurewill be re-evaluated (with the communication messages considered) in the next snapshot, because the model employs a movingwindowtechnique,and updates its solutions at each new snapshot. E

The decomposition-coordination approach to solve the task coordination CODE framework to establish a correproblem can be mapped onto the spondencebetweenthequantitativeandconceptualrepresentations of human coordination and decision activities. UsingDM1 as an example, Fig. 23.5 shows theCODE framework with mathematical variables mapped onto related functional activities. Figure 23.5 shows that DM1 receives information about the environment (e.g., numbersof different tasks, task status) via the situation channel, and receives communication messages (about b,O's) from other DMs through the communication channel. The information is processed to generatea ~ ~ r ~ i mate operation completion timeci: and begin time ijo of other DMs' oper tions of the team tasks from the messages, and to estimate the earliest tim eg and latest time to process an operation. With such information, DM solves theteam problem iteratively using his mental model to determine the ion (based on the multipliers{Ai$,and makes his own S assumed to have limited communication capability, sage on an operation when the associated multi~lieris positive. Then DM1 takes actions communicate to with and to execute his task processing decisions. This decision-mak continues from one snapshot to the next.

WILUVG, ~ E I N M A NLUH , F “ “ ” ” ” ” ” ” ” ” ” ” ” ” ” ” “ ” ” ” ” ” ” ” -

1 ’

FIG.23.5. DMl’s submodel as mapped onto the CODE framework.

With respect to the HITEC experiment, the distributed model captures the followingvariations in coordinator’srole, communication structure and time pressure. ~ 1 The ~ . distribute^ modelcapturestwoof the four coordinator’s rolesexamined in the experiment (i.e., the Observer and the Advisor cases 1321). Whenacting as an observer, the coordinator is not involved in online decision making,so that the distributed team model only includes submodels for subordinates. In the Advisor case, the coordinator can send (formatted) messages to subordinates as regar~stheir processing of team tasks. The distributed model thus contains a submodel forthe coordinator-similartothoseofthesubordinates-thatalsosolvestheteam problem based on its local information, as noted in Subsection 3.2. However, by the coordinator. no tasks are actually processed The benefit of communicati~g with the coordinator liesin the unification ofteammembers’decisions.Communicationamong the§ubordinates is based on individua~ submodel solutions. Communication from the coordinator is based on the solutions of the single coordinatorsu~model.Therefore, co~municationamong thesubordinatenodestendstobemore diverse than that from the coordinator. Better coord~nation may ensueif the submodels adjust their decisions according to communications from the coordinator as opposed to lateral communications from subordi~ates.

23.

ODEL LING TEAM C O O ~ I N ~ T I O N

re. Subordinates’ communication to the coordinator affects the coordinator’s model solutions, and in turn thoseof the subordinates themselves.Two communication structures wereexamined in the experiment: the Coordinator-Centered (CC) and the Fully-Connected (FC). Theseparticularcommunicationsstructures are often referred to in the organizational literature as “star versus completely connected,” “centralized versus decentralized~ and so on. In CC, subordinates can only communicate with the coordinator so that the distributed model does not allow information exchange among subordinate DMs. Messages regarding the subordinates’ decisions are only sent to the submodelof the coordinator, which in turn can send its recommended decisions{bi:] to the subordinates.In FC, all team members can communicate among themselves. In this case, the submodelofeachsubordinate alsocommun~cateswithothersubordinate nodes, in addition to the coordinator’s node. Empirically, it was foundinthat the FC structure about one halfof the subordinates’ messages were sent to the coordinator.In the model, a particular communication message from a subordinate is sent randomly to either the submodel of (1) the coordinator, or (2) the appropriate subordinate, with equal probability. Clearly, a more refined modeling of interlevel communications isa topic for further study. s ~ ~ Two ~ e levels . of time pressure were createdin the experiment by changing the task tempo. Under low time pressure, tasks arrived Q, = 4.4 tasksfmin), had long initial time available less frequently (arrival rate (Ta= 160 sec), and tooka longer time to process (f, = 35 sec). Underhigh time pressure, thee~perimenthas a faster tempo(Q, = 6 tasksfmin,Ta= 115 sec, and f, = 25 sec). The distributed model is implemented for the actual experimental scenarios across the three independent variables comprising 2 x 2 x 2 = 8 conditions. The model uses exactly the same data with the same task arrival sequence as in the experiment. For each experimenta~ condition, there are two replications corresponding to slightly different task scenarios (generated via different random seeds). For both model and data the two replications are averaged to get results for each condition.

Thedevelopmentof a normative-descriptive model involves an iterative process of model-datacomparison,identi€icationand ~uantificationof human limitations and biases as descriptive factors, and their incorporation within the model. The process is carried out in two phases. Thefi considers basic relevant human cognitive limitations, which are kn previous research.The second phase reveals additional teamCO

WANG, ~

E

ILUH~

~

,

and decisionmaking trends or biases, and develops a normative-descriptive model of the team.

The distributed team model isfirst implemented under the assumption that all DMs have perfect mental models, with global information and sufficient capability to solve the optimization problem (Eqs. 1-4). Under these conditions, all submodels of individual DMs generate the same schedule for processing a task, or they ignore the same tasks when necessary. There is no need for communication as no constraintwill be violated, even though some may be tight with positive multipliers. In general,thismodelshowedsignificantlybetterperformancethan humanteams,withhigherteamreward (92% forthemodelvs.79%for human teams), fewer coordination failures(1.1% vs. 8.8%),and longer slack time (63 sec vs. 49.8 sec). This is primarily attributed to the unrealistic assumption of perfect mental models. According to our experiment observations and research on human cognition [lo, 12, 321, the following factors need to be considered:

1. In the experiment, the team had a decentralized information structure. Each DM had access to detailed information on all arrived team tasks and hi own indi~dualtasks, but only saw the total value (Le., aggregated information) of the otherDM'S individual tasks. Consequently,there was uncertainty about the arrival times{al)and deadlines {dl)of those ITS. This uncertainty can be accommodated by assuming that the submodel of a DM has "noisy" data for the arrival times a,' = a,+ W, and deadlinesdi' = di + W, of the ITS of other DMs, where W, and W, are random variables. 2. The earliest and latest times to process an operation of a team task could be calculated from the task arrival time, operation number, and processing time required as in Eq. (5). To do so, a human DM would needto memorize and manipulate the numerical values for each task. Given the dynamic multitask environment, one might not be able to do that precisely. This ina~ilitycan be representedby an uncertaintyin estimating eiiand Zii, viz. the submodel of each DM uses noisy estimates eii' = etj+ m3,and Zij' = $ w4,where w3 and u4are random variables. 3. A formatted communication message, as described in ~ubsection3.2, only conveys information about a D 'S task processing order (now, next, etc.) rather than the exact task processing6 , .time The com~unicatedinformation thus carries with it uncertainty about the actualb,. Thus, we model information exchange among submodels as conveying noisy begin times b,' f

23, MODELING TEAM COO~INATION

where w5 is a random noise. As a DM'S communication commonly occurs some short time after a decision is made, W, is assumed to be positive. 4. Typically, a DM would not start an operation of a team task until he received verification that the preceding operation was completed, asindicated by a change in the displayed task status. In the experiment, the status update was subject to random delays as a human DM could not monitor each task processing closely, nor update the task's status immediately after an operation was Completed. Therefore, each submodel is only informedof the completion timeof a team task operation after some random delay (i.e., cii'= cii+ w6,where is a positive random variable). Random delays in status update cause random delaysin processing subsequent operations.

= b,,+ W,,

For simplicity, the random variables {wJare assumed to be indepe~~ent and uniformly distributed. The noise levelsare selected based on extensive shows the model testing after incorporating these random factors. 23.1 Table selected noise levels, which are held constant over all experiment conditions. Table23.2 shows the averaged team reward, slack time and interoperation slack timeof the experimental data and the model solutionsover the eight experimental conditions. Table 23.2 shows that the model solutions now more closely match experimental results in team reward and timing measures on average. TABM 23.l Selected ~ ~ d Parameters o m (S~onds)

TABLB 23.2 M o d e l - ~ ~omp~sons a~ on Averaged Measures Team

~eward ~ x ~ r i ~ n t

ode1

Slack

Time

78.8

49.8 sec

10.9 sec

81.7

55.5 sec

9.1 sec

WANG, ~ E I N ~ A LUH N, ea

When model results are compared to experimentally derived measuresinvolving communication and task processing strategy across both levelsof time pressure, a numberof consistent discrepancies emerge: 1. Human teams sent more communication messages than did the model, especially “Processing” messages (related to current decisions). Theaverage ~ommunication rate for the team (excluding observer under CC communication structure) was about 1.1 msgs per minute for each DM. In the model, a message is sent on an operation only if the operation has a tight window (associated with a positive multiplier). The model solution shows about 0.3 msgs per minute, about 70% less than that of human teams! The model solution also shows a message composition with much lower percentage of ‘~rocessing” messages than thatof the human teams, especially under low time pressure. This discrepancy suggests an ‘~vercommunication” trendof human teams,as also discussedin [S, 14,271. peration tasks (30Ts) but

23. ~ O D ~ L TEAM I N ~ COO~INATION

Task Type

FIG. 23.6. ~ode1-datacomparison: (a) distribution of task types processed, (b)effect of time pressureon team reward.

WANC, ~ E I NLUH ~ ,

interpreted as a “processing” message; when b,,>0 the associated message is one of “plan to process.” Note that as a consequence ofusing the dataderived communication rate in our model, we can only examine model-data comparison with respect to thereZ~~iue (%)use of different messagetypes, as opposed to total message traffic. 2. ~vercooperation-Themodelismodified to overvalue each %operation task by 5 points and undervalue each large individual taskby 5 points. hen the model has multiple solutions aforsnapshot, it always chooses the one with team tasks scheduled earliest. 3. Planning horizon reduction-The model’s planning horizon is fixed at3 stages for low time pressure conditions and 2 stages forhigh time pressure conditions. is Afterincorporatingtheforegoingdescriptivefactors,theteammodel again applied to solve the eight experimental conditions (across the three independent variables). The model solutions now agree closely with the empirical data across all the measures. Table 23.3 shows the model-data comparison in terms of five measures averaged over the eight experimental conditions. The model solutions also match empirical data in various patterns, such as the compositionof task types processed and the composition of communicationmessagesused,asshown in Fig. 23.7.Thecompositionoftasks processed is a measure of team task processing strategy, whereas the eomposition of messages used reflects strategy of explicit team coordination. The model-data matching in pattern as perFig. 23.7 implies that themodel not only reflects what the team did (performance measures), but how it was done (process measures). The model results replicate the empirical data on a case-by-case basis, and thus exhibit the same variations toindepen~ent the variables asdid the human teams. Figure 23.8 shows the model-data comparison across four TAB= 23.3 on Avera~ed ~ o d e l - Comp~son ~~a

~

x

Model

~

~

78.8% ~ e 79.1k

~

78.5% t

8.8%

43.5 sec

10.9 sec

78.6

10.1%

44.9 see

10.2 sec

23,

ODEL LING T W COO~INATION

a EXP MDL

100 h

Task Type EXP 80

MDL

Message Type FIG.23.7. Model-datacomparison of process measures: (a) composition of tasks processed,(b) composition of messages used.

major measures: team reward, percentage of tasks processed, coordination failure, and slack time. In the figure, the cases are coded by a 3character string X Y Z with X = L or H, standing forLow or High time pressure;Y = 0 or A, for observe^ or Advisor of coordinator’s role;and Z = C or F,for CC or FC communication structure. The earlier results can be averaged across different independent variables to analyze the effects of coordinator’s role, communication structure, and time pressure in the same way as in the analysis of the e~perimental data [32]. Take team reward as an example. The model solution shows that team reward in the Advisor case (81%)is higher than that in the Observer case (77%). This indicates the benefit of having a coordinator tocommu~icate with subordinates. The change of communication structure show much effect in team reward (about 1% change). This is a t t r i ~ u t to e~ the exercise of mutual mental models which enables each DM to antici~ate

WANG,~ E I N MLUH ~ , a

ao 75

0

70

Case Index

Case Index

t / I

80

70

75

70

20

Case Index

Case Index

FIG. 23.8. Case-by-casemodel-datacomparison:(a)comparison of team reward, (b) comparison of task processing, (c) comparison of coordination failure, (d) comparisonof slack time.

others'decisions, andthus to coordinateimplicitly. As time pressure increases fromlow to high, team reward reduces from about 82% to 76%. This is mainly attributed to a reduction in planning horizon as discussed earlier. These results are consistent with the experimental data as analyzed in [32]. Similar analyses can be made on other measures.

Thecharacteristics ofimplicitand explicitcoordinationhavebeendiscussed earlier in this chapter,as well as in other research [2,3,8, lo]. It is, however, difficult to observe empirically the mec~anismsof implicit coordination as no effective methods exist with which to measure or to evaluate mental mo~elsupon which implicit coordination is based [3]. Nor is it easy in an experiment to control the mixof i~plicitand explicit coo~dination.

23. ~ O D ~ L I N TEAM G COO~INATIO~

This, instead, can be achievedby using a model-based approach, Using our normative-descriptive model as a predictive computational tool, this section examines the two questions intrinsic to team coordination. Can a team achievea reasonable performanceby using only implicit coordination or only explicit coordination, or is it necessary to have both? * What is the incremental effect of explicit coordination on team performancewithrespecttotheexistence ofanimplicitcoordination mechanism?

4

Two simulation studies are conducted by changing the implicit versus explicit coordination conditions in the normativ~descriptivemodel. The model implementation is otherwise kept the same (e.g., the noise levels). For each coordination condition, the model is applied to generate predictions for all eight experimental cases that were subsets of the HITEC experiment. The results for each coordination condition are then obtainedby averaging the resultsof the eight cases.

The first simulation study exercises the model at four coordination levels: ‘‘indi~dual~ (INDV), ‘~ommunication” (COMM), “mentalmodel” ( and “full combination”(FULL). At the INDV level, the submodelof each solves only his individual/local problem by assuming the most conditions for its own decisions. For example, consider a DM W sible for the second operation of a three-operation task. ThisD priori that the first operation will be processed at the earliest third operation will be processed right before the deadline so that he has the maximumfreedom toschedulethesecondoperation.Decisions are made by scheduling task operations within resource constraints (Eq. 2) and boundary constraints (Eq.4). Each submodel does not communicate with any other.If a submodel schedules an operation to be processed before its precedent is finished, the processingwill be carried out immediately after the precedent is completed.Thisisequivalenttotheuse of task status update to keep all submodels informed as to when an operation of an team task is finished. TheINDV level model achievesa team rewardof about 62%, as shown in Fig. 23.9. The model for the second level of coordination ( ~ O ~ aM ) INDV level with communication capability. The submodel ea of decisions asin the INDV level; but can~om~unicate with others based on the local solutions. Because the submodel does not predict others’ decisions, com~unicationis conducted with priorities on “ignore” and “process-

WANG. ~ E I N ~ A LUH N,

Coordination Level FIG. 23.9. Effectofcoordinationlevel.

ing” messages over “plan” messages, similar to the communication pattern ofhumanteams [32]. Also the communication is limited at the empirical communication rate (about 4.2 msg./min). For example, if three decisions,

one scheduled at Time zero and two at some time later, need to be communicated but only one message can be sent, the one scheduled at Time zero will be communicated (correspondingto a processing message). The team reward obtainedin this case reaches 73%. The MMDL level is the sameas the Observer casein the CC communication structure, where the submodels, however, do not communicate with each other. Here the submodel of each DM predicts teammates’ decisions and makes its own decisionsby solving the team problem based on its local information.Thiscaseemulates a coordinationconditionwithimplicit mechanisms only. Team reward improves significantly to about 75%. The FULL level is the sameas the Advisor casein the FC communication structure. Compared to the MMDL level, theFULL level adds explicit coordination capability. Each submodel solves theteam problem to predict others’ decisions (implicit coordination), and also uses communication with other submodels (explicit coordination). The advantageof communication is demonstrated by a small increasein team reward, reaching about 79%. These simulation results show that team performance improves significantly (from about 62% to 75%) by adding either team communication or mental model predictions to an INDV model. comb in in^ them generated a lesssignificantimprovement(from75% to 79%).Therefore, a teammay reach a reasonable performanceby using either implicitor explicit coordination alone in the task coordination problem, It is, however, their combination that brings the team reward up to the experimental level(79%)).

23. ~ O ~ ~ L T I N WG ~OO~INATION

Thesecondsimulationstudyexamines in moredetailtherelationship between implicit and explicit coordination mechanisms. In the first study, the MMDL and the INDV levels had no communication, whereas the FULL and the COMM levelshadcommunicationsattheexperimentalrate (1.1 msg./min per DM). This study examines two more cases, namely ~ O M M * and FULL", obtained by limiting the communication rate to 0.4 msg./min in the COMM and FULL cases, respectively. Figure 23.10 shows the results for the sixsimulation cases. The MMbL, FULL*+and FULL levels show the result of progressively increasing communication rate when the model has mutual prediction capability (solving the team problem with implicit coordination: W/ IC). At the INDV, COMM*, and COMM levels, submodelsdo not predict the decisions of each other(and there is no implicit coordination: w/oIC). With implicit coordination, the team performance does not degrade when communication rate is reduced from1.1 ms ./min to 0.4 msg./mi~. With no implieit coordination, limiting of communications causesa significant decrease in team reward. This study shows that implicit coordination largely reduces the need for communications to maintain performance. The marginal benefit of having more communication depends on how much communication capability the team already has. When the team members do not utilize implicit coordination, due to lack of training or information, more communication capability would lead to better team performance (as long as the DMs are not overloaded),

90

80

70 60 50

Communi~ation Rate Limitation FIG. 23.10. Interaction of communication rate.

WANG, ~ E I NLUH ~ ,

Coordination has long been recognized as a key issue when studying distributed systemsin general, and team decision makingin particular [15,181. We believe that team decision making cannot be thoroughly understood without a knowledgeof team coordination mechanisms,as it is coordination that makes a team different from a collection of individuals. With this in mind, we addressed the team coordination problem more directly than in previous research, and conducted a normative-descriptive study using a distributed task coordination problem within a hierarchical team. This chapter presented the anal~ical-normativeand normative-descriptive parts of this study, and made the following four contributions. 1. A conceptual framework was developed to provide a structural format for team coordination and decision making. The CODE framework integrated the SHOR model 12,331, the (mutual) mental model construct13, 71, and the coordination mechanisms inherent in an individual’s decision-making process in a team context. Team coordination was depicted as a mixof implicit and explicit mechanisms, where the balance between them shifts according to variationsin the external and internal conditions. The strength of the CODE framework lies in its capability to organize existing results and draw new insights and inferences about team behavior. With its direct treatment of coordination, we believe that the CODE framework can help integrate empirical resultsin a systematic way to obtain consistent and insightful findings about team decision making, and provideas an axiom-like basis to generate hypotheses amenable for future testing. 2. The CODE framework was used as a foundation for our mathematical modeling of the HITEC experiment. TheCODE framework guided the model development,provided a “human-like”viewofthemathematicalprocedures, and gave an intuitive feel to the~uantitativevariables. The resulting distribut~dmodel captured implicit team coordination as mutualpredictions of decisions, and explicit team coordination as decision information exchanges, as well as the actual distributed decision making of the team in its task processing. Thus, CODE provides a structure-rootedin human engineering-for computational models of team decision processes employing d~stributedintelligent“agents.”Submodelsforindividualagentsmaybe based on a dynamic programming formulation, such as ours, or use various decision rules such as in the workof Levitt et al.[35] or Carley and Prietula [36],or use a Petri net formalismas in Levis [37]. 3. Two team coordination biases (overcommunication and overcooperation),andanadaptationmechanism to time pressure (reduced planning horizon) were identifiedas descriptive factors via model-data comparison,

23, MO~ELINGTEAM ~ O O ~ I N A T I O N

The identificationof the coordination biases and adaptation mechanism was a result of comparing empirical results (what human teams actually did) with normative model solutions (what should be done). These factors were quantifiedandincorporatedintothenormativemodeltoyield a normative-descriptive model, that generated consistent data-matching results on a case-by-case basis. 4. Team implicit and explicit coordination were quantitatively analyzed by using the resulting normative-descriptive model as a simulation-computational tool.It was shown that having mutual mental models was crucial for a team to reachcoordinateddecisionsandmaintainperformancewhen communicationwaslimited,andthat satisfactory teamperformance depended on the appropriately integrated useof implicit and explicit coorin the litdination mechanisms. Althoughthese results have been discussed erature as inference, i~plication,or postulate[2,3,8,10,2?],they are quantitatively demonstrated here for the first time, Such results find application to models that ex~licitlytreat or simulate interagent coordination mechanisms, suchas in the Virtual Design Team (VDT) simulation modelof Levitt et al.[35] where communication messagesare exchanged among intelligent (and hence “implicitly coordinating”) agents via a choice of media. There are a quite a number of areas in which the results of this chapter can be expanded(or strengthened) as topics for future work. One is to consider theimpact of team versus individual goals on team performance and process measures. The current effort assumes that individualD self-centered interests, and that there are no conflicts between ~ndividual and team payoffs. A normati~e-descriptive study of individual versus team by Shi, Luh, and ~einman I251using a variantof the rewards was conducted HITEC environment.Basically,theyfoundthatunderindividualrewards there was an increase in explicit coordination, which was explainedby the fact that a DM could not as accurately predict the Iikely decisions of his t e ~ m a t e(through s the mutual mental model; i.e., implicit coordination was less effective). An issue that plagues most laboratory research involving human is teams one of scale. It becomes extremely difficult, if not prohibitive, to conduct controlled laboratory experiments (on complex problems that stress coordination) with teamsof size much greater than six. In order to make the leap to larger size teams (and organizations) we build wouldmodels thatare Validated on small teams and postulate that the predictivecapabili~of these models extends to larger teams. Alternatively, for organizations that are inherently hierarchical (suchas in the military), a model fora generic fourof person hierarchywill often find itself repeated at higher and lower levels aggregation throu~houtthe organization. Thus, the issue is how to aggre-

WANG, ~ E I N MLUH ~ ,

gate these generic modeling "units." Some workin this direction appears in Pete, Pattipati, and Meinman 1201 for distributed detection problems.

This work was supported in part by the National Science Foundation under CTCT grant 1 ~ ~ 9 0 2 ~and 5 5by , the Office of Naval Research under contract NOOOl~9~J-1~53. TheauthorswouldliketothankMr.DanielSerfatyfromAPTIMA, Inc. ~ o b u r nMA , and Professor S h i ~ h u Chang n~ from National Taiwan University, Taiwan for valuable discussions on the modeling and interpretation of implicit coordination.

[l] L. Adelman, D. A. Zirk, P. E. Lehner, R J. Moffett,and R. Hal, "Distributed Tactical DecisionMaking: Conceptual Frameworkand Empirical Results,"IEEE Transactions on Systems, Man, and ~ ~ e r n e ~ c s , (S), 1986,pp. 794-805. SMC-16 121 M. Athans, "The Expert Teamof Experts Approachto C o m m a n d - ~ d ~ o n t r(C2) o l Organizations," Control Systems M a ~ ~ i nSept. e , 1982, pp. 30-38. [3] J. A. Cannon-Bowers, and E.Salas, "Cognitive psychologyand Team Training:Shared Mental Models in Complex Systems," Symposium Address presented at theAnnual Meeting of the SocietyforIndus~ialand O~an~ational Psych (41 R. Conant, and W. R Ashby, "Every GoodRegulator of A System Must Be A Model of That System," International J ~ ~ a l ~Science, S ~ t Vol. e m1,1970, s pp. 89-91. [5] J, R. Hackman, and C. G. Morris, "Group Tasks, Group Interaction Processes, and Group Performance Effectiveness: A Review and Proposed Integration," in L. Berkowitz (ed.), Advances in ~ ~ e r i m ~ nSocial t a l ~ y c h o lVol. ~ , 8, pp. 45-99. New York Academic Press, 1975. [6] D. J. Hoitomt, P.B. Luh, E. Max, and K R Pattipati (WO), "Scheduling Jobs with Simple Precedence Constraints on Parallel Machines,"C o n ~Systems l Magazine, Vol. 10, No. 2, Feb. 1990, pp. 639-645. [7 ' 1 P. Johnson-Laird,Mental Models,Cambridge, M Harvard UniversityPress, 1983. Commande~'Information Needs, [S] J. P. k h a n , D. R. Worley, and C. Stasz, ~nde~tanding ~D/R-376l-A,RAND Arroyo Center,Santa Monica, CA, 1989. [9] G. A. Klein, "Recognition-primed Decisions," in W.B. Rouse (ed.), Advances in Man-Machine Systems ~ e s e a ~Vol. h , 5, pp. 47-92, Greenwich,CT JM Press, 1989. [101 D, L, ~ e i n m a nP., B.Luh, K R Pattipati, and D. Serfaty, "Mathematical Modelsof Team Distributed Decisionmaking," inR W, Swezey and E.Salas (eds.), Teams: Their Training and Performance,Norwood, NJ: Ablex, 1992,pp. 177-218. 1111 D. L. Kleinman, and D. Serfaty, "Team Performance Assessment in Distributed Decisionmaking," Proceedings of Interactive Ne~orkedSimulation for Traini~Conference,Orlando, FL. April, 1989. [l21 P. H. Lindsay, and D. A. Norman, ~ u m a nInformation Processing,London: Academic.Press, 1977.

23. M~DELINGTEAM C O O ~ I N A ~ O N

[131 D. G, Luenberger, Linear and~ o n f j n e a ~ ~ ~ a Reading, m m i n gM, A Addison-~Iey,1989. [141 K. R MacCrimmon, "Descriptive Aspects of Team Theory: Observation, Communication and Decision Heuristics in Information Systems," Man~ementScience,Vol. 20, No.10, June, 1974, pp. 1323-1334.

[151 T. W. Malone, and K. Crowston, "Towardan Interdisciplina~Theory of Coordination,"MI" Technical Report,CCS TR# 120, April 1991. [161J. Marschak, and R. Radner, Economic Theory of Teams, New Haven, CT Yale University Press, 1972. [171 J. E. McGrath, Groups: Interaction and Performance, Englewood Cliffs, NJ, Prentice-Hall, 1984. [l81 X. Y. Miao, P, B. Luh, D. L. Kleinman, and D. A. Castanon, "A Normativ~Descriptive Approach to Hierarchical Team Resource Allocation," IEEE Trans. on Systems, Man, and Qbernetics,Vol. 22, No. 3, MayIJune 1992, pp. 482-497. [191 B. B, Morgan, A. S. Glickman, E. A. Woodard, A. S. Blaiwes, and E. Salas, Measurement of 4), F L Navy Training SysTeam Behavio~in a Navy Environment ( ~ S C - T R ~ ~ lOrlando, tems Center, 1986. [20] A. Pete, K. R, Pattipati, and D. L. Kleinman, "Distributed Detection in Teams with Partial Information:A Normative-Descriptive Model," IEEE Trans. on Systems, Man and Cybernetics, Vol. 23, No. 6,1993, pp. 1626-1648. 1211 R. W. Pew, and S. Baron, "Perspectives on Human Performancemodel ling^ Automatica,Vol. 19, No. 6, 1983,pp. 663-676. [22] J. Rasmussen,Informatjon Processing and~uman-Machine Interaction: An Approach to Cognitive Engineering,New York North-Holland, 1986. [23] W. B. Rouse, and N. M. Morris, "On Lookinginto the Black Box: Prospects and Limits inthe Search for MentalModels," P s y c h ~ ~ i cBulletin, al Vol. 100,1986, pp. 349-363. [24] E. Salas, T. L. Dickinson, S. A. Converse, and S. I. Tannenbaum, "Towardan Understanding of Team Performanceand Training," inR. W. Swezey and E. Salas (ed.), Teams: Their Training and Performance,Norwood, NJ: Ablex, 1992, pp. 3-30. [25] P. Shi, P. B. Luh, and D. L. Kleinman, "TeamCoordination Under Individual and Team Goals," in K M. Carley and M. J. Prietula (eds.), Computat~onalO ~ a n ~ a t i oTheory, n Hillsdale, NJ: Lawrence Erlbaum Associates,1994, pp. 263-288. [261 H. A. Simon, ~ ~ eofBounded l s ~ationali~: ~ehavioralEconomics and BusinessO ~ a n ~ a t i o n , Cambridge, W The MIT Press, 1982. 1271 M. E,Shaw, Group Dynamics: The~ s y c h of ~Small o ~ Group Behavior, New York McGrawHill, 1976. [28] H. G. Stassen, G. Johannsen, and N. Moray, "Internal Representation, Internal Model, Human Performance Model and Mental Workload," Automatica, Vol.26, No. 4, 1990, pp. 811-820. [29] I. D. Steiner, Group Processes andP r ~ u c t i v iNew ~ . York Academic Press, 1972. [30] S. Streufert, and G. Nogami, "Cognitive Complexity and Team Decision Maklng," in R. W. Swezey and E. Salas (eds.), Teams: Their Training and Performance,Norwood, NJ Ablex, 1992, pp. 127-152. [31] W. P. Wang, and P. B. Luh, D. L. Kleinman, and S. C. Chang, "Coordination Mechanisms: A View from Dual Variables," Proceedings ofthe 30th IEEE Conference on Decision and Control, Brighton, UK, Dec. 1991. [32] W. P. Wang, P.B. Luh, D. L. Kleinman, D. Serfaty, "Hierarchical Team Coordination under Time Pressure: Coordinator's Role and Commun~cationStructure," submitted to IEEE Transac~ons on Systems, Man, and Cybernetics. [33] J. G. Wohl, "Force Management Decision Requirements for Air Force Tactical Command and Control,"IEEE trans. on Systems, Man, and Qbernetics, Vol. SMC-11, No. 9, September, 1981, pp. 618-639.

WANG, ~ E I N ~ LUH N ,

[M]D. Zakay, and S. Wooler, "TimePressure, Training, and Decision Effectiveness," Ergonomics,

Vol. 27, No. 3, March 1984, pp. 273-284. 1351R.E.Levitt, G.P. Cohen, J.C. Kunz, C. I. Nass, T.Christiansen, andY. Lin, "The Virtual Design Team: Simulating How Organization Structure and Information Processing Tools Affect Organization Team Performance," in K. M. Carley and M. J. Prietula (eds.), Compufat~ona~ Theory, Hillsdale, NJ: Lawrence Erlbaum Associates,1994, pp. 1-18. I361 K. M. Carley, and M. J. Prietula, "ACTS Theory: Extendingthe Model of Bounded Rationality," in K.M. Carley and M. J. Prietula (eds.), Compurationai Organization Theory, Hillsdale, NJ: Lawrence Erlbaum Associates,1994, pp. 55-88. €371A. H. Levis, "Colored PetriNet Model of Command and Control Nodes," in C. Jones (ed.), Toward Q Science of Command, Control and Communicafions,Washington DC: AiAA Press, 1993, pp. 181-192.

P A R T

.

This Page Intentionally Left Blank

C H A P T E R

i

Geoffrey C. Bowker Susan Leigh Star

Unive~ity of Californiaat San Diego

The same factors ~ h i c hhave thus coalesced into the exactness and minute prec~ion the of formoflife have coalesced intos a~ c ~ofrthe e h~hestimpersonali~;on the other hand, they have ~romo~ed a h~hly personal subjectivi~ -Simmel(1960, p. 413)

I

LE

F

I live on the net. The Internet isa new nation. * The laborato~ of the 2lStc e n has ~ no ~ walls and no boundaries, but ais v i r ~ a communi^ l * We are all ~ e ~ e now. n s * ~lec~onic comm~nica~on has revolu~on~ed the way scienceis done.

*

*

Those of us studying the use of electronic mediaare often faced with statements such as these. The popular media often confound daily life and routine work and practice, electronic communication suchas email, and that which transpires “over the net,” with the concept of “community.” There isa substantial elisionof experience, material conditions, structural positions in

B O W R AND STAR

particular social networks, communication, and location in discussions of "the net," and "the web." As social scientists, this worries and intrigues We have us. been involvedin the development and evaluation of several software systems for by usescientists. Whatever is happening with the scientists for whom the software was designed, it is clear that neither custom software efforts, nor the larger electronic environment are literally places where people live full time, nor even work full time. It does not provide either face-to-face local communi^ nor full working infrastruc~ure, as those concepts might plausibly be applied to closeknit collegial networks or to occupational communities and labor pools. Early sociologists struggled along similar lines with the concept of community as itwasaffected by anotherlarge-scaletechnologicalshift,the in~ustrialrevolution. They used the term ge~e~nschaft to refer to tribaland village life, thesort one is born into, where everyone knows who you are by virtue of family and station. At first glance, it seemed thatsort thisof life was being threatened or even destroyedby cities and factory work.~esellschaft referred to its opposite: the world of the industrialized city, markedby commerce, where who youare reflects your placein a more distanced, formally commercial and secular order. There were many arguments about whether a f t , among social scientists and urbanization had destroyedg e ~ e ~ ~ s c hboth in the general public sphere. Since their use during this time, the terms have been generalized more loosely to refer to intimate versus distanced relationships. (As well, the early sense of the disappearanceof community has been modified and complexified.)' During the infrastructural shift that is occurring with the building and integration of global electronic networks, some of the past lessons from sociologica~ studies of communities and networksare useful. Withthe interest in terms such as electronic c o ~ ~today, ~ nthei intimate ~ versus distanced distinctionsare again important.We seek here to clarify some senses in which ~ 0 t happlytothedeve~opment of collaboratories. We are learning, as did early sociologists, that online community is not a choice between familiar locale and alienated metropolis, but that elements of both are important for design and analysis (Bishop ai,, et in press). Theworkofdoingsciencestilltakesplace substantialls off line, even though electronic communication and data sharing aspects are becoming increasingly more important. Even the parts of the work that may be isolated as "information work"are not necessarily conducted via electronic media: 'This section owesa great deal to discussions with Alaina Kanfer, and draws extensively on an earlier papercoauthoredbyStarandKanfer,"Virtual Ge~~inschaft or Electronic Gesellschaft?: Analyzingan Electronic Community System for Scientists," presented at the Society for the Social Study of Science (4S), Purdue University, November,1994. Her consent to use this analysis is gratefully acknowledged.

24. SOCIAL ISSUES IN D S I G N

People talk to each other, run down the hallway, writebythings hand in notebooks and on labels, and alsoFAX: and telephone each other constantly.Further, one is not “born” into the scientific ‘~ommunity~ but one’s sponsorship and apprenticeship occur over a long period of time “offline” for the most part, in graduate school and via long-standing collegial and friendship networks. At the same time, electronic tools are of growing importance, and elements of nature are being tested, taught, and modified in virtual space. Bowker (1994a, 1994b) has used the concept ‘‘infrastructural inversion’’ to describe a conceptual shiftin the social studiesof technology, especiallyin history of techno lo^. He impliesthat a figure~roundgestaltshifthas occurred: decentering individuals, single artifacts, or even social movements as causal factorsin large-scale scientific change. Instead, when infrastructura1changeistreated as the primary phenomenon, collective processes of transformation are morerichlyexplained.Changes in infrastructural networkssuch astransportation,information,anddomestictechnologies explain a great deal about other formsof social change and social relationships-they are not simply substrate, but substance (Bowker & Star, 1999; Star, 1999). If this is true, then the substant~ve changes effected by (among many others) theNational Information Infrastructure Initiative,the Collaboratory efforts, and the Digital Library initiatives have significance both for science and far beyond it.Wenow have a chance to observe this phenomenon as itunfolds.Itis a momentthat will not recur, and which requires extraordinary effect from social, information and computer scientists. One democratic concern here is that the social inequities and distributions of information resources will become somehow frozen or reified in the infrastructural changes. This begins with the simple notion of uinformation rich” and “information poor,” but extends more complexlyas more of life’s business is conducted electronically. The digital divide has become omnipresent.

Infrastructural shifts on the scaleof global electronic networkingare not all that common in human history. They tendto generate extensive discussion of “basics”: values, concepts, moral directions. During the great shifts of the 18th and 19thce~turiesfrom rural to urban, agricultural industria^, to and to large-scale organized capitalism, some basic questions were raised about community that echo those found today in Science magazine’s articles on the global electronic laboratory and very large databases. Distinctionsbetweentypes ofhumanbonding-andthetroublesome notion of community-indeed informed the founding of the discipline of sociology. Nineteenth-century sociologists worried a great deal about the forces,

B O W R AND STAR

or nature of the links that hold people together (e.g.,Dur~eim,19W, Marx, 1970; Simmel, 1950; Tonnies, 1957), which seemed to be undergoing massive transformation throughout the Industrial Revolution. Did movement from the village, farm, and tribe to city, factory, and contracts change us, body and soul? Did what held us togetherin small, rural groupings differ radically from that which held us togetherin large urban conglomerates?If so, how? Volumes, if notlibraries,havebeenwrittenonthecontrastbetween ~emeinsc~ and a ~geseZZsc~aft,including the authors who find no essential change in human organization and bonding between country and city, farm and factory, The concept of c ~ m ~and n all i it~might entail is one of the most contestedin the historyof social sciences. Some social scientists, such as Simmel(1950), held with confidence that the metropolis has spawneda new "mental life."In his classic essay on the topic, he claims that: "There is perhapsnopsychic phenomenon, whichhasbeen so unconditionally reserved to themetropolis as has the blask attitude" (p. 413). Others went on to investigate the notion in rural societies in less-developed countries, suchas Redfield (1947). As time went on, their claims were modified and challengedby other social scientists such as Lewis (1949) who found plenty of evidence for conflict and arms'-length relationships in the village, and~illmottand Young (19603 who found plenty of g e m ~ i n s inc ~ ~ ~ the close-knit working class neighborhoodsof urban London's East End. It seems that community is not such an easy notion to place, nor such an easy one to dispense with. Social scientists need concepts to describe how it is that people "stick together" within groups. But as with other fundamental concepts, such as "species"in biology or "culture"in anthropology (Clifford 8~Marcus, 1986), ironically it becomes both the glue that binds and the thin upon. (Perhaps it is just that, as Simmel noted long ago, conflict itself is "socializing," and the agreement to disagree is what binds these disciplines together.) Stacey (1969), in a magisterial reviewof the decadesof arguments about the term c o m m ~ n inotes ~ , that "It is doubtful whether the concept ~ommunity' refers to a useful abstraction. Certainly, confusion continues to reign over the usesof the term"(p. 134). Like "power"or "profession,"th6 term community clearly points to a phenomenon of key interest for understanding large-scale change-but exactly what, or why, remains elusive. Stacey goes on to note that while one author deplores the romanticismof the notion, another claims itas the framework in which humans are introduced to civilization itself: This is so vague as to be nonsense: there is no such thing as community which does this. Various agencies are involved in this process of introduction, perhaps neighbors; almost certainly parents' kind andfriends(whichmaylive next door or miles away). These institutions may, or may not, be locally based.

24. SOCIAL ISSUES IN DESIGN

They may,or may not, be inter-related. If they are locality-based and inter-related thenthere may well bea local social system worth studying, but one would hesitate to call this a community. Nor there is any less lack of confusion in earlier usagesof the term.(p. 135)

The conflation and generalization of these terms continue to this day.

C o ~ ~ ~ n i seem t i e s to be everywhere in the media: the diplomatic communi-

ty, the international community, the women’s communication, the AfricanAmerican community, the high-tech community, the bird-watching community.Perhapsourfavorite of theseasonhasbeenthe“heterosexual community.” What is the concept that joins such disparate collections of allegiances and heterogeneities? In the world of science, the notion of “scientific community~ (and its cousin, “invisible college”) has been equaled only by the term ~ a r a ind ~ ~ the amount of disputed terrain contained within it. ~ethodologicaldivergences in how one would measurea community, where its boundariesare, whether it’s a meaningful unit of analysis,and so on, have enriched and confused the worldof social scientific work about science (STS, or science, technology, and society). Alternative notions have been proposed to emphasize one dimension or anotherof the social relations under scrutiny: communities of practice (Lave & Wenger, 1991); social worlds (Clarke, 1991; Clarke & ~ontini,1993; Strauss, 1978); relevant social groups (Bijker, 1995), and actor networks (Callon, Law,& Rip, 1986), to name just a few. The development of 1argGscale electronic infrastructures have given a new life to the battered notionof community, and have direct relevanceto collaboratory researchand its descendents, suchas studiesof cooperative work on the web and related CSCW (computer-suppo~ed cooperative work) projects. Early on, these events had two basic foci: c ~ s t o ~ s o~ro~ects ~~ffre (such as the NSF collaboratories)whichremotely link research already joined by common interests or heritage; and tools for ~ r o ~thes extant i ~ internet resourcesin order to discover useful information or possible colleagues, The collaboratory concept emergedin the late 1980s froma top-down initiative fromtheNationalScienceFoundation in Washington.Dr.Wil~iam Wulf, then directorof the NSF Directorate for Computer and Information ScienceandEngineeringwrote a foundational white paper: “The proposal, then, is to undertake a major, coordinated program of research and development leading to an electronic ‘collaboratory’ a ‘center without walls,’in which the nation’s researchers can perform their research without regard to geographical location, interacting with colleagues, accessing instrumentation, sharing data and computational resources, accessing information in digital libraries’’ (Lederberg & Uncapher, 1989, p. 19). Wulf‘s paper recalls Vannevar Bush’s canonical Science-The Endless ~ r o n ~ ewhich r , led to the

B O W R AND STAR

foundation of the NSF.Bushhad also called for improvementsto scientific information flow. Bush’s logic had been government through big science will give citizens health, wealth, and security-his clarion call opened with the words: “The 6overnment should accept new responsibilities for promoting the flowof new scientific knowled~eand the developmentof scientific talent in our youth. These responsibilities are the proper concernof 6overnment, for they vitally affect our health, our and jobs,our national security” (Bush, 1 9 ~ / 1 9 4p.~8).,

Wulf‘s logic, along similar lines, was that government through distributed science will give the nation technology, which leadsto economic and military stature and competitiveness-in the opening words of his paper: “The health of the UnitedStates, economically and militarily, depends on its technology base. The technology base depends on the number, quality, and productivity of the nation’s research scientists and engineers” (Lederberg & Uncapher, 1989, p. 19). Severalimportanthistoricaldevelopmentshave occurred between the time of Bush andthat of Wulf, of course. Most notably, has been insertedas amajor mediator between science and nature (through simulations and advanced instrumentation). A new kind of information technologyhasbecomeanessentialsupportforscientificwork,whichhas important implications for the practice of science as well as the notion of community. Wulf‘s paper led to a workshop held at Rockefeller Universityin March 1989 (Lederberg & Uncapher, 1989, p. i). The workshopreport found several needs fora collaboratory. ~nstruments, it was noted, were often to be found in environments hostile or inaccessible to people. Further, people themso that remote interaction with colselves were inconveniently distributed, leagues was necessary “whenever the appropriate mix of talents to address an interdisciplinary problem is not collocated anywhere.” Finally, not only people and instruments, but also data were d~stributed-remote interaction was needed when: “the dataare too vast to be replicated and managed at a single location,” for example with the global seismic database. Accordingly, the goal wasto build: “no less thana distributed intelligence,fully and seamlessly networked, with fully supported computational assistance designed to accelerate the pace and quality of discourse, and a broadening of the awareness of discovery: in a word, aC~llaboratory.~’ Here then is one defining reading of the word: collaboratories are about distribution in every sense of the word~istributionof things, people (thus 6allo formed a distributed college without walls-a “dream team” to fight AIDS) and information. This givesa great dealof scientific and theoretical power to computer andinformationscientists-whohavedevelopedtheoriesofdistributed databases and long-range communication networks; it allows a translation between government imperatives and new tools being developed within computer and information science: “The Collaboratory is much more than

24. SOCIAL ISSUES IN DESIGN

just a setof tools. It is a national computer-based infrastructure for scientific research." The phenomenal growth of the web scarcely needs any review in this venue; the numberof users grows exponentially. At the same time,there is considerable concern about the lackof good indexing and sorting tools, as well as a need for better means of ascertaining data quality. Thus the distinctions between custom projects and the development of the web are not absolute, bur rather foci that may blur into each other over time. Even custom projects interface with the larger networks through email and links with databases (and increasingly this is the case € omult~media r as in local sites come well); browsing tools when adopted and extensively used to form a de facto custom package over time. With the advent of the Web, the repertoire of social science tools for its analysis has also vastly expanded, to include social of studies structural and nascent communities; web-based social networks (Kiesler & Sproul~,1996); online/offline "ecologies"; digital libraries (Bishop etinal., press; Weedman, 1992), and the useof advanced modeling techniques incIuding visualization and VR.

In spite of all the controversy associated with exact definitions of comm~nity, thereis general agreement that the sense of community rests on nontrivof sharedknowledge, ial, on go in^ relationsamongpeople;somedegree understandings, material objects, or conventional practices; and the idea that these twoare not independent. Initial research on computer-mediated coIlaboration, or electronic communication systems showed some interesting effects in terms of the social relations, or role differentiation among members (Sproull& Kiesler, 1996). Thisin turn affected sharedund~rstandings, or consensus €ormation in decision making. A common fin~ingin this research wasthattechnologicallymediatedcommunicationcreatesless role differentiation among group members that did face-to-face communication. This was attribute to the fact of less visible ~ifferentiationbetween group members when communication occurs electronically (you can't see someone's race or gender, for example; you couldn't hear their voice; you by looking thei at can't impute social cue or position posture). ons sequent these researchers, there te more uninhibite~corn (vertical) ore communication between members of ~ i f f e r ~status, n t and more equal partici ation in ia ele~tronicmail iegel, ~ u b r o v s ~ '

BOWIER AND STAR

Researchers argued that the nature of decisions and agreements in such computer mediated groups may be affected, in part, by the decreased role differentiationthatoccurs in comecomputermediatedgroups.Some researchers have found thatless role differentiation allows group members to generate more ideas in problem-sol~ngsituations (Connolly, Jessup, & Valacich, 1990; Jessup & Tansik, 1991). When there aremore ideas to choose from, group members may have a harder time deciding ona solution creating a longer time to reach consensus, or point a of action. This is obviously a relative notion, depending on small, identifiable groups withc~generated histories. Whether these ideas scale up to larger electronic groups is highly debatedbothaboutwhetherthedistinctionistoosimplistic,orabout whether other forms of social cues emerge over time. The larger and more distributed the groups, the more difficult status-role differentiation to anais lyze.Baym (1995, 2 ~ 0 argued ) that emergent groups on the net, such as Usenet groups, develop strong cues within the email messages themselves, including differentiation, leadership, and novel semantic conventions based on practice. A linked finding wasthatgroupswithinorganizationscommunicating with computer mediation seemed to take longer to reach consensus than groups communicating face-to-face. Without traditional social cues to inform decision making, the process become more open, less reified or controlled by traditional lines of authority. Siege1 et al. (1986) reached the same conclusion after observing greater choice shift and inefficient communication amonggroupmemberscommunicating electronicallythanamonggroup members tryingto reach consensusin face-to-€ace settings. On the other hand, Galegher and Kraut (1990) suggested that decreased role differentiation has the opposite effect on consensus formation. They focuson the diffuse responsibility and joint ownership that results from electronicallymediated joint production. With less indi~dualidentification with the product, they argue, therewill be a greater tendencytoward conformity thus speeding up the consensus formation process. In science,technological mediation-or electronic community systems-are becoming routine parts of scientific teamwork, both locally and aatdistance. “big”science consists of scientists addressing problems so large they e solved by a lone scientist. Such large problems require the combinedeffortsofscientificteammembers.Historically,oneoftheprimary determinants of scientific collaboration has been physicalpro~mity.However as the scope of scientific endeavors grows and members of the scientific ~omm~nity are more mobile, scientists ~ o r ~ i on n gthe same problem often are not co-located. Physicists have a particularly ture to support international collaboration of this S fore, the benefits of productive col~aborations afforded by pro~mity must be achieved by other means. It was claimed and hoped that electronic omm mu-

24. SOCIAL, ISSUES IN DESIGN

nity systemswill provide the means for frequent interactions, and joint access totoolsandinformationforscientistscollaboratinglong-distance(Kraut, ido, & ~alegher,1990; Lederberg & Uncapher, 198%Schatz, 1992). Since the mid-l990s, mostof these tools operate via or alongside the Web. One feature often integrated with electronic community systems is, of course, emaii. Early studies, partly in dialogue with the organization decision support studies of electronic “rooms” and local electronic messaging, tried to show how email would affect role differentiation and consensus formation in scientific co~laborations.The findings were somewhat inconclusive. Tombaugh (1984) argued that greater role differentiation would have facilitatedtheinternationalcommunication in aninternational asynchr~ nous messaging,or conference system. Scientistsin his study felt the need for more leadership. Hiltz(1983) studied 103 scientists communicating asynchronouslyoverthenet.Thistechnologicallymediatedcommunication resulted in greater perceived understandingof other scientists’ interests or theories, thereby affecting problem solving and decision making. Although this early researchdid not take place in an~hing like the Intern ment, it is usefulin identifying problemsin the designof electronic community systems for large, geographically dispersed scientific communities, and for illuminating some of the conceptual problems addressed earlier in this chapter. Electronic infrastructural developments have made it possible to include in electronic community systemsa variety of functions in addition to electronic mail (e.g., information sharing, document editing and collaborative writing tools, and data visualization techniques and collaborative tools). Although the emergent functionalitiesof a €ommunity system will depend on how community members use it, there are two fundamental functionalities for which such systems are designed: 1, An electronic systemis designed as aquicker means for existing communications and social relations that compose a community. For instance,in a scientific community, journal articles can be viewed as a mode of communication of ideas in science-and electronic journals and preprints can provide the same function, more quickly. 2. The el~ctronicmedium can be usedto support new activities, ~iffer~nt things are possible on electronic systems and the notion of scientific community can be changed. For instance, traditionally by keeping all informationaboutan areaavailableonline,withannotations by anddialo among experienced professionals, the training of newcomers can become an activity supportedby the electronic system.

The degree to which these capacities of electronic community systems may lead to changes in the scientific community itselfare only beginnin

BOWKER AND STAR

be analyzed. The first capability may lead to quicker spread of information or diffusion among a relatively well-defined group of scientists. The second approach may make it easier for new people to come into a specialty,and more difficult to maintain boundaries.In other words, electronic community systems may have the effect of reducingthe gate keep in^ roles of interpersonal relations (this is one hope of public policy advocates of such systems). However, if large resource differences continue to stratify electronic infrastructures (e.g., with respect to advanced multimedia) they may reinforce extant disparities. These are some of the things that might be considered by designers of electronic community systems for scientists. An electronic system can be developed fora specific scientific community by community membersor by be outsiders. In addition, a“generic”electroniccommunitysystemcan developed to support all types of general activity. When a scientificcommunity adopts a generic electronic community system, its members can adapt some of the systems features for their specific use. Schatz, for example, proS next) collaboratory system to ating thecelegans ~ C (discussed es, and this has beenimplemente~for drosophila

der (1994) and Kohler (1995) for the lives of rats and dro~ophilain scientific netparticular, discusses the survival strategies the drosophila ernployto protect this highly rarefied niche. 3A d~scriptionand representation of ~~S canbefound at http:/ canis.uiuc.edu,along with a description of the Illinois Digital Library Project and the on~oingInterspace Project.

24. SOCIALISSUES IN DESIGN

linked with a directoryof those interestedin the particular subtopic, anda quarterlynewsletter-the Worm Breeder’sGazette. It alsoincorporatesa database independently developedin Europe designed for the community, acedb. Many parts of the systemare hypertext-linked with each other. ~ C was S developedin 1990-1993 by Bruce Schatz and Sam Ward at theUniversity of Arizona (after which it moved to the University of Illinois, and parts of its core programming were adapted in the Illinois Digital Library Project, ~ C was S designed fora par1998, andin the ongoing Interspace Project). ticular group with the idea of eventually migrating the structures of the software to other groups. Like the Sequoia 2000 project which deals with global change, this project was seen as being interesting both for its domain specific support and the nature of the computer science involved-meeting the information needs of scientists as well as providing a challenge for basic computer science research. (Some of the complexities and difficulties of this relationship are explored in Star & Ruhleder, 1996; and Weedman, 1992.) Star and Ruhleder (1996) worked as ethnographers on the project, considering potential s~ciolo~ica~ effects and dynamics within the system as a whole. They traveled to worm labs across the United States arid Canada, observed the use of computing and other featuresof ‘ ~ o r m aspects of routine work and communication. As well,they about other featuresof the work, such as scientific career§ etition, routine information-sharing tasks, and how ing infrastructure is managed.Theyvisitedmorethan 30 labs a

BOWXER AND STAR

ticular characteristics, such as mutation, may be mailedafrom central Stock Center to labs requesting specimens, Tracking the location and characteristics of organisms thus is an important part of record-keeping and information retrieval. Two points emerge immediately here. First, both the worm and the scientist become partof a single distributed community. The worm biologists travel a virtual network and their community only exists at their of nodes; set their subjects travel amail network, and live only at the nodes. Second, and most significantly? the worms and the scientists can only travel (really and virtually in turn) if there is a common set of standard^ shared across the network. Protocols must bein place to ensure thata given set of electronic signals means a contribution toa distributed database, andwill be understood as such at each point along the way. Equally, there must be protocols to ensure that a given package ofworm tissue means a standard mutation in each lab in the node. Berg(1997) calls this the need to discipline local practice, and he notes that his aisparadoxical feature of the attempt to create a transparent, flexible infrastructure. Usage patternsin WCS reveal the sometimes competing nature of custom versusglobalemphases in informationsystemsdesign.Manypotential users of VVCS moved to simpler, less functional web tools suchas Gopher, Usenet, or simple email (and after the Web, migrated there). Star andRuhleder (1995) analyze this in part as an unfamiliarity with some of the infrastructural tools, such as the Unix operating system,as wellas other aspects of local infrastructure and support systems.

Ruhleder (1995) has written about the waysin which classical Greek schole .~ arship chan~ed with the introductionof the ~ ~ e ~ ~ a ~ ~r~ r~a es~c ~This complete canon of classical Greek literature on database was made available online, during the1970s with updates since. Tracking down the occurrence of a single word throughout that canon, with a view to uncovering its modalities, used to be a lobar involving sensuality (the feel and smell of the book so ~eautifullyrendered by Charles Lamb), prodigious memory? and a odly setof notes. It was not something onedid at the startof one’s career. nelearned,thoughapprenticeshipoverthe years, howandwhatand where. Now this work can be done with the touch ofa button by a graduate student embarking ona PhD. Is it the same workin those two instances? Or, es’ two Don Quixotes, one writtenby Cervantes and oneby a later academic, does the work itself change with its context, in this case its infrastructura~support? We have so farlookedattheories of workpractice informing the development of collaborative infrastructure;in this section we

~

24. SOCIAL ISSUES IN DESIGN

explore how what it is to be a scientist is being affected by the development of high speed networked information infrastructures. Steven Hawking speculated that by the turn of the century theoretical physics would be the province of the computer; the role of the human being would be to attempt to understand and appreciate discoveries made elsewhere. Although this seems unlikely, thereare two basic waysin which the new information infrastructure is radically changing the natureof scientific work the natureof representation and the natureof the scientific product. is changingin the sense that theoretical work The nature o~r~~resentfftion is increasingly being delegated to the intelligent instrument-which works through terabits of data streaming in and decides (by one of a number of algorithms) which data is interesting and which is not; and then represents the interesting data graphically according to another ofsetalgorithms. The interpretivework is deliberately partially delegated to the machine,in order to cope with information retrieval.Itisof course true that the appropriate algorithm can be changed, but one suspects that once thisofact dele~ation is made and ramified (as one infrastructure submerges into another) then the attached algorithmswill be naturalized (in the anthropological sense). (Arelevant robustfinding from library science is that patrons will use a convenient electronic source they know to be incomplete in preference over a card catalog that they know to be complete.) There will literally be no other way to see the world. The nature of the scientific product is therefore changing in several ways. The scientific paper is arguably no longer the t e ~ ~ i n ad u s ~ u eItis ~an. archivaldocument of usetopeople in otherdisciplines,arguedone researchers (~cience,3/24/95,p. 1764). It is certainly increasingly the case the publication is proceeding online. This means more than the mere transposition of linear texts onto the screen. The new infor~ationinfrastructure is just as significant as Eisenstein (1979) has persuasively argued that the book was for the presentationof scientific data. Information is not being presented here in linear form, with the word as the centerof attention-rather the representation becomes the thing, with linear argument as secondary. At the limit, the scientific product becomes itself a library-the human genome for e~ampl~which is to be consulted as a huge interactive database created collaborativelyby an array of henceforth anonymous authors.

In cornputer and information science there is a unique relationship theory and practice, fact and artifact, word and The thing. word (in the form of a computer pro~ram)is the thing (the instrument, the tool, the co~muni-

BOWKER AND STAR

cation medium, the simulator); just as in a fully distributed, computer-mediated work environment the mapis the territory.So many things that we as social scientists have been tryingto prove for years-the inscription of theories in technical objects, thetheory~ladennessof observation andso forthare suddenly literal truths.We need to develop new theoretical sensibilities in order to move around in this brash new world. There isa strandin science and technology studies that a theory and values playa central rolein en~ineeringdesign ner, 19~0).In a classic article on the electric car, Callon (1989) argued that the ~soci~ngineers" designing the vehicle were at the same time betting ona theory of society in which the car itself became inevitable. Bowker (1994a) looking at ~ c ~ l ~ ~ 6 scientific e ~ e r ' work s argued that thegeop~ysicalcompany engineered work practices and social life around the oii inwell such a way that their own science became first possible thenine~itable.With the collaboratory~onceivedhere as the deliberate creationof a new information infrastructur~for scienc~there is a multiple interpenetrationof social the broadest level there is the network architecture of the national If. As is well-known, this architecture originatlate 1960s in the need felt by the military for edResearchProjectsAgency)projects-atthe loping the CYCLADES network. This early connection with the military meant that much scientific deve~opment in the new information infrastructure has been physics-led. Thus, physicists developed the electronic preprint service that for many scientists has replaced the ading of journals (currentlyto be found at http://~.lanl.gov); the World ide Web was first developedat theCEW laboratory. In a brilliant paper, Abbate(1994) discussed the theoretical understanding by the main players of the nature of information networking-and how this played outin a seriesof relatively irreversible technical choices. There has beenover the past few decades a conflict between two major protocols ~ , 2 and 5 TCP/IP) deScribing the waysin which computers can talk to each other. The ARPA model (TCPIIP) assumed that the"network itself should have low-levelstructur~thisallowed greater heterogeneity among the laboratories being interconnecte~although atthesametime it entailed a greater degree of comput~ng sophistication on the part of the labs themselves. The alternative protocol~ . 2 5 grew ) outof a CCITT (Comit~Consultatif International T~l~graphi¶ue et T~l~phoni¶ue) initiative. Drawing on the mo~el of the telephone, the federated PTT's involved chose insteadto make the network itself well-structured-so that at the other end there could be u~sophisticated dumb users, justlike clients of telephone ed ~omplexitywith the advantageof control (Abbate, 25 usa~ilitybutlittle fle~ibility.A mode~edcommunicationasbeing

24. SOCIAL ISSUES IN DESIGN

between well~quippedlaboratories with their being a premium on speed and redundancyof communication-traditional military values;CCIT put the premium on volume and usability. These differences have hadvery practical consequences for the practice of scientificworkandforscientificcommunication. It wastheflexibility offered by A W ~ E that T allowed the Internet to developin such a distributed, anarchistic fashion-the efflorescence and rapid propagation ofnew programs and standards froma huge array of sources, since each laboratory in the military web was assumed to have an interest in maximal cooperation. The success of this model is mostclearly demonstratedby the development of NCSA Mosaic in the early1990s.Mosaic revolutionized the World Wide Web (developed at CERN); its HTML standard createda new kind of usability and transparency. Within this distributed model, the interactive sharing of information has been difficult technically (so many possible standards and protocols need to be taken into account simultaneously in order to create a transparent system). Recognizing this,ARPA concentrated workin the early1980s on ways of translating email messages between incompatible systems: “The result has been that electronic mail is the only formof connection that has been fully achievedbetweendiversenetworks.Thusthecommunication aspect of networking has been emphasized, encouraging new usesof computers that stress interaction rather than calculation” (Abbate, 1994, p. 20’7). Typical of the complexitiesof the infor~ationinfrastructure environment has been that this battle between the systems of X.25 and TCPJP has not resulted in the deathof one in favor of the other. Rather, X.25 is generally used, but only that subsetof it that supportsTCP/IP (Abbate, 1994,p. 206). In computer-mediated work, these forms of level shiftare frequent. (Fora fuller account of these transformations, see Abbate, 1999,) The AWA theory of a computer networkas a setof connections between “centers of calculation” (Latour, 1987) has, then, led to a degree of local autonomy unusualin the history of large technical systems (e.g., compare by Thomas Hughes [1983]); and has the story of electricity networks as told led to an email-driven version of scientific community. There is no necessary paradox between military centralism and the democratization of Internet development. The key vision is that of the center of calculation being the node; and the question then is only who controls the nodes (anda for long period in post-World War I1 America, the military effectively controlled most significant high-tech laboratories). However, it is not only at the level of network architecture that the theory of community with which one is working comes into play; oncontrary, the in general one can say thatwith information infrastructure development of the orderof scientific collaboratories and now digital libraries that itis theory all the way down. Another levelof theory comes when one is trying to designaninfrastructuraltoolthat will actuallygetused by scientific

B O ~ AND R STAR

researchers, To do this, one needs a modelof what it is that scientists do, and thena suiteof programs thatwill permit themto do the same better and faster. Picture as a thought experiment an information systems designer assuming the truthof Merton’s norms for the scientific community (cf. McClure et of what might happento scientific norms with al., 1991, p. 93) for a discussion the introduction of electronic research networks).In the past, these norms may well have beena very useful part of the discourseof scientists: A positive self-image that demarcates science from business and projects an image of the scientist as the fair-minded arbiter of truth. Butwhenitcomes to developing an information infrastructure sucha as collaboratorywith which scientists will work to share and shape their experiments, then whether or not the norms represent work reality will make a difference, Thus, when Starassociologist on thedeteamlookedatthecommunity ofworm biologists, she found that these latter certainly did not want to share all information with rival labs on an ongoing basis, contra Merton’s p1orm of communism: indeed, particularly early in their careers (as ambitious postdoctoral students carving out their own part of the genome), these biolosts placed a great~remiumon secrecy. Scientists would not usea system at permitted no room for privacy: Working practice had to be made explicit in order to design a usable product, and that practice had to be informed caltheorStarandRuhleder (1996) noted that the WCS was htlongdistancecollaborationsamongthemany the project; co-ordinate this scientific work; and allow the ra id involvement of new scientists by way of online recruitment meet these goals, the infrastructural collaboratory had to reasonabl reflect workingpractice in order to gain initial.acceptanceassumes, of course, that a seriesof workarounds developed theory of practice over time will increasingly allow for affordances between inscribed in the infrastructure and practice itself.

In the Internet, properties are constantly being swapped between humans andnonhumans-suchthatwhatconstitutes mind, memory, intelli~ence,

and scientific work are being redefined by the new infrastructure. It also impacts efficiency. Hiltz and Turoff‘s classic work The ~ e ~ ~ t~ ifirst o ~~, in 19’78, makes the claim that: “Although a crucial endea~orto the maintenance of our society as we know it, research is a highly inefficient process when compared to other institutional functions’’(p. 212). This passage reminds us that efficiency has long been a problem and a goal in elec-

24. SOCIAL ISSUE3 IN DESIGN

tronic communication. This goal has been there, indeed, since the origin of computing: Babbages’s original 1832 design for his calculating engine was based in partonhisadmirationforProny’smethod of cal~ulatinglogarithms-itself an application of the factory principleof the division of labor (Bowker, 1994b). Hiltz and Turoff (1993) described a working computer conferencing system (CCS) that operated within an electronic information exchange system (EIES) as: Another capability being incorporated into EIES indicates the role a GC§ can play in the area of resource sharing. A fairly sophisticated microprocessor with its own computer~ontrolledtelephone dialer has been programmed to engage in the conference systemas afull-fledged member, with the same powis referred to). ers of interaction as any human member (Hal Zilog, as it/she/he This entity may perform any of the following tasks: 1. It may enter EIES and receive orsend messages orretrieve and enter items

into the other componentsof the system. 2. It may exercise certain analysis routines or generate display graphics from data providedby other EIES members, and returnthe resultsto them. 3, It may phone other computers and select data from existing data bases or obtain the resultsof a model to send back to any designated group ofEIES users. 4. Itmay drop off and pick up communic~tion items from other conference and message systems.@p. 25-26)

This new member of the network wasa full equal of any human members. Further, human members of the CCS were themselves transfor~edby the process in the sense that their disembodied intelligence would be free to act to its fullest capabilities: “computer based teleconferencing acts as a filter, filtering out irrelevant andir~ationalinterpersonal ’noise’ and enhances the communication of highly-informed ’pure reason’-a quest of philosophers since ancient times”(p. 28, citing Johansen, Vallee, & Collins, 1977): Sociological studiesof computer conferencing systems have tended to these claims of enhanced filtering (Baym, 1995; ‘Yates &.Orlik and indeed work in organi~ationtheory-notabl~inspired by Janis’ classic article on ~roupthink-hasopened the~uestionof the valueof c o ~ ~ e r a t i ~ e work (Kraut et al.,1990, p. 152). e dream remains alive, asores shadowed in the report of the on co~~a~oratories cited earlier, which aimed at creating “a distributed jnt~~~~gence, fully and seamlessly networked.” A scientific instru*It is only fair to note here that Hiltz and Turoff also point out that the authors themselves question the operation of this filtering.

7

B O W E R AND STAR

ment in this view no longera passive intermediary between the mind of the scientistandnature:"Incorporation of intelligenceintotheinstruments allows the possibility of 'self-directed' data gathering, with the instrument itself deciding when data is significant and should be transmitted, setting parameters based on local feedback, and doing preliminary data reduction. This can lead to both reduction in communications and archiving requirements and better scientific data" (pp.13-14). A formal model of information flow here can no longer have the structure subject+verb+object = scientist+instrumental action+nature; just as likely is instrument-instrumental action-nature, with the information systems designer and not the scientist as thedeus ex~ a c h i ~ a . So we have on the one hand the pure humanmind and on the other the artificial mind of the intelligent information retrieval system: each made possible by the development of a seamless infrastructure.The human and the intelligent agent meet in the infrastructural database-thus the famous Los Alamos e-print archive referred to earlier (http://~.lanl.gov) in 1997 had a welcome to humans but a warning to nonhumans to be on good behavior: "ROBOTS BEWARE: indiscriminate automated downloads from this site are notpermitted."Furtherexplorationrecoversthefollowingmessage, topped by a picture of a no entry sign superimposed on a photo~rap~ of Commander Data from Star Trek the Next Generation: This www server hasbeenunder all-toefrequentattack from"intelligent agents" (ak.a. "robots") that mindlessly download every link encountered, ultimately trying to access the entire database through the listings links.In most cases, these processes are run by well-intentioned but thoughtless neophytes, ignorant of common sense guidelines. (Very few of these same robotrunners would ever dream of downloading entire databases via anonymous ftp, but for some reason conceptualizewww sites as somehow associated only to small and limited databases. This mentality must change-large databases such as this one [which has over l ~ , distinct UWs that leadto gigabytes of data]are likely to grow ever more commonly exported viawww.) Following a proposed standard for robot exclusion, this site maintainsa file /robots.txt that specifies thoseUWs that are off-limits to robots (note that the t~tle/author listings for all archives remain available for remote indexing).Continued rapid-fire requests from any site afteraccess has been deniedwill be interwill bedecidedly pretedas a networkattack,andtheautomatedresponse unfriendly. (Clickhere to initiate automated "seek-anddestroy" against your site.)

The point hereis that as far as thec o ~ ~ u t~ata~ase er is concerned, there is very little difference between humans and nonhumans; just as for thescientific instru~entthere is very little difference between its own directive intel~igence and thatof the scientist.

~

24. SOCIAL ISSUES IN DESIGN

The collaborative infrastructure is being designed precisely in the context of collapsingthehuman-nonhumandivide.Theworkofcreating a seamless intelligent network isat present posited squarely on the possibility of rendering humans and nonhumans interchangeable. Impelling the collaboratory effortis a central focus on information as product and commodity. Notoriously, in any information~enteredvision of the world, it matters little formally where the information is being or held processed(its container): The only thing that is important is that it circulate flawlessly and be analyzedthoroughly.Earlyworkonthenature of information science after World War 11 instantiate this claim. One example is Wiener’s early, outrageous conjecture that people might be sent down a variety of telephone wires as code,and restored physically at the other end: after all, we were just information (Wiener, 1957, p. 23). Another is Turing’s famous demonstration that his abstract logic machine (an infinite roll of tape, and a mechanism for moving and marking) was equivalent to any real logic machine” whether this be lodgedin the human brainor in a scientific instrument (see Bowker, 1994b for a further explorationof these issues). Sociological and philosophicaltheory is getting literally encoded into the infrastructure of collaborative scientific work on the net. It becomesan active resource drawnon by engineers and designers as they create anew system. Thus any given theory-if encoded into successful software-can be truer now than it was before. Winograd and Flores’ Coordinator, for example, was oneof the earlyand most influential groupwork programs; it is one of several such programs that utilizes Searle’s speech-act theory (Rodden, 1992, p. 5; Winograd & Flores, 1987). The program structures office communication into assertives, directives, commissives, expressives, and declarations; and any future researcher analyzing collaborative work using sys- this tem will find thesenaturalizedcategoriesstructuringallcomputer communication and having ramifications for noncomputer mediated work, One could say that the infrastructure can only work if the theory that everything is information is true; a preferable statement would be that it can only work if it can make the theory that everythingis information true-change the worldin such a way that this isa fair descriptionof the natureof things (see Suchman,1994 for a critiqueof Winograd and Flores on these grounds, for example). This is a somewhat pessimistic viewof the collaboratory as architectof humans as vessels of pure reason, computers as same, and computer networks as reason dist~ibution engines. There is of course another way of reading the collapseof the human/nonhuman divide-as representedby the works of Haraway (1985,1997), Star(1991,1999) and Latour(1993). According to this reading, categories that were assumed to inhere in the head (suchas ‘ ~ e m o r“cognition,” y~ and “learning” can themselves be understood as spatially and temporally distributed), Cicourel(199Q), for example, pointed to a

unity between social analysis and infrastructural development: “The of idea socially distributed cognition refers to the fact that participantsin collaborative work relationships are likely tovary in the knowledge they possess, and must therefore engage each otherin dialogues that allow them to pool resources and negotiate their differences to accomplish their tasks. The notion of socially distributed cognition is analogous to the idea of distributed computing” (p.223). Or again, following historically on the lead of Maurice Ha~bwachs, an influential group of workers within the worldof computersupportedcooperativework ( C S C ~havearguedthatmemoryis invariably a social phenomenon, which is both spatially and temporally disiddleton lk Edwards, 1990). Work in distributed artificial intelligence also speaksof “composite systems”of humans and machines.

Ratherthanindulge in an intellectually vacuous exercise of det~rmining “where is the boundary” between online work and offline or work, the exact distinction between “electronic communities” and “scientific communities,’’ we offer some findings and cautions about the visions of collaboratories, The community metaphor is a powerful one, but one whose heritage is so fraught that it is almost useless to try to retrieve it intact (seeJones, 1998for excellent discussions of the matter in cyberspace). flet, the sheer level of use of the term demonstrates that we are using sit ofor ~ ~ t ~impor~ant ing to us.) Rather, we would like to make the following observations. Collaboratories contain inherently competing goals (as do all large systems). This iswhy they are communicative systems. With the growth of big science,forinstancewith cetegans a designatedmodelforthehuman genome initiative, an electronic community system is designed with the followinggoals in mind: tosupportlong-distancecollaborationsamongthe many researchers in the large scientific effort; to help coordinate the large scientific effort; to allow the rapid involvementofnew scientists by wayof onlinerecruitmentandtraining.Theworm researchers,however,were explicit in that theydid not want to lose that informal, close-knit community feeling.A more structured scientific effort may impose formals ties and weaken the informal ties upon which the community has been built. The availability of online training, and reduced need for mentorship MAY indeed decrease the boundary definition and intimacyof the community, although this is by no means clear as yet. Therefore, the goalsof the system are antithetical to the feelingsof the existing scientific community. Therefore, the goalsof the system are antithetical to the feelings of the existing scientific community. Indeed the goals of the systemare in conflict with each other in precisely this way. Simultaneous attempts to build g ~ ~ e i n s and c ~ geseZZsc~a~t; a~ to blue

24. SOCIAL ISSUES IN DESIGN

locale and global reach; to preserve intimacy and extend community infinitely will not work without careful attention to the distribution between handsg e ~ e ~ n sand c ~gae~s e f f s c ~ a ~ ~ . on and automated work, between In scientific communities like celegans, the scientists valued the closeknit collaborative workingenvironmentandwereopento.anelectronic community system that supported this. An electronic system without a formal structure of information and communication links might best support this closeness. However, usersofan electronic system get frustrated without some structured access to information, and the desire to use the system for training suggestsa system with a more arms’ lengthset of relationships and structures. These twore~uirementsof an electronic sci~ntificcommunity system create an inherent design conflict. The solution to this conflict in response to the may come from a more “organized” system, that evolves community evolution, or perhaps from newly evolving forms and conventions that we cannot yet imagine. The notion of collaboratory itself aismoving target. However, key elements are an orientation to information flow-between instruments, people, and documents~mbeddedin an integrated information infrastructure.It is assumed that within this infrastructure the map will become theterritory. Just like Huysmans (1981) in A ~ e deciding ~ nototo visit ~London ~ because he had already in a quayside restaurantin Calais experienced all the sensations that the visit itself would produce;so do the new scientists not need to see each other, their instruments, or the world in order to do their valuable work of theory production. At the same time, theoriesof situated action and workplace studies (I-Iutchins1995; Lave, 1988; Star, 1995; Suchman, 1987) show us that without attention to the local contingencies and differences in hands-on, craft aspects of even formal work, systems and knowledge risk irrelevance and rigidity. We see the importantof vase, dense electronic networks that scientists use as an opportunity not to engage in boundary disputes, but rather to use the conceptual tools from different parts of the social sciences to understand the phenomenon empirically and theoretically. One thing is clear: Media hype is not helpful in this enterprise, nor is extending the inherited clutter and entropy associated with the conceptof community. However, we hope that perhapsa real opportunity to combine the empirical work from ethnography and situated studiesin analyzing new forms of communication, media for work practices, and affiliations, if not communities, may help shed some light on this old problem. There is nothing new about scientific collaboration being distrib~tedin time and place-the correspondence circles of 17th~enturyscientists; Darwin’s enormous rangeof correspondents; the large-scale collaborations during World War I1 leading to the development of the atomic bomb. What is new and interesting is the embedding of specific fopms of collabor~tioninto an infor~ationinfrastructure that impacts the very nature of scientific work.We

BOWKER AND STAR

havearguedthatastheinformationinfrastructurebecomesevermore willitbe natudeeply engrained, then the successful theories inscribed into ralized. Theywill come to seem true and unproblematic. However, it should be noted that this is not equivalent to saying that the best theory will win. The history of standards and infrastructures makesvery it plain thata series will triof contingencies mean that often the second best (or even the worst) umph. The Lotus123 spreadsheet, theDOS operating system, and VHS format are (in)famous examples.Nor does it mean that the theory which becomes naturalized will give the best possible description of scientific work and practice. Afterall, Merton’s norms were long naturalized in the scientific community; but they certainlydid not describe the waysin which scientists acted. But this natura~ization does count for something. The new truth becomes the problematic in terms of which theory-critical and otherwise-is defined. Social theorists have takenon a very active role in the development of collaboratories. Muchgoodsociologicalanalysishasbeenwritten by authors involvedin the design of systems. Where the field of science studies has in the past called attention to the sophistication of the actors’own perspective and served as spokespeople for the actors, the shoe is now firmly on the other foot.The designers of the new infrastructureare calling attention to the sophisticationof sociology’s viewsof information and work, and are serving to represent these views within their programs (Bowker, Gasser, Star, &C Turner, 1997). A fearful, but quite pleasing, symmetry.

WCSwas partially funded by NSF under grants 1 ~ - ~ 1 51~-92-57252, ~7, and B I R - 9 ~ 1 9The ~ . Interspace project is currently housed within the Community Systems Laboratory(CSL), headed by Bruce Schatz, affiliatedwith the Graduate School of Library and Information Science and the National Center for Supercomputing Applications at the University of Illinois,Urbana~hampaign. Additional support was providedby the Universityof Arizona and the University of Illinois. CO-PIS Bruce Schatz and Sam Ward, and developers Terry Friedman and Ed Grossman were extremely generous with their time, comments, and access to data and meetings; we also thank our anonymous respondents for their time and insight. Additional thanks to Karen Ruhleder Bill Turner. and

Abbate, Janet. 1994. “The Internet Challenge: Conflict and Compromise in Computer Networking.” In J. Summerton (ed.), C h a ~Large i ~ ~echnicaiSystems (pp. 193-210). Boulder, CO: Westview Press.

24.SOCIALISSUES

IN DfESIGN

Abbate, Janet. 1999,Inuenting the Internet.Cambridge, NIA: MIT Press. Ackoff, Russell L., Thomas A. Cowan, Peter Davis, Martin C. J. Elton,James C. Emergy, Maybeth L. Meditz, and Wladimer M. Sachs. 1976. Des~ninga National S c i e n ~ and ~ c Technol~icalCommunica~onsSystem. Philadelphia: University of Pennsylvania Press. Baym, Nancy K 1995. "From Practice to Culture on Usenet." In S. L. Star (ed.), The Cultures of Computj~(pp. 29-52). Oxford, UK: Blackwell. B a p , Nancy K. 2000. Tune In, Log On: Soaps, Fandom, and Online Community.Thousand Oaks, CA SAGE. Berg, Marc. 1997.R a t i o n a l ~Medical i~ Work-Decision Support Techniques and Medical Problems. Cambridge, MA: MIT Press. Bijker, WiebeE. 1995. OfBicycles, Bakelite, and Bulbs: Toward a Theory of Sociotechnical Change. Cambridge, MA: MIT Press. i oTech~ Bijker, WiebeE.,Thomas P. Hughes, and Trevor Pinch (eds.).1987. The SocialC o ~ s ~ u c tof nological Systems: New Directions in the S o c i o l ~and Hktory o f T e c h n o lCambridge, ~. MA: MIT Press. Bishop, Ann, Laura Neumann, SusanLeigh Star, Cecelia Merkel,Emily Ignacio, and Robert Sandusky. In press. "Digital Libraries: Situating Use In Changing Information Infrastructure." Journal of the American Society for information Science, 51. Bowker, GeoffreyC. 1993. 'Howto be Universal: Some Cybernetic Strategies.' Social Studies ofScience, 23, 107-127. Bowker, GeoffreyC. 1994a.Science on the Run: Info~ation Management andIndus~ialGeophysics at Schlumbe~e~ 1920-1940.Cambridge,M A MIT Press. Bowker, Geoffrey, 1994b. "Information Mythology and Infrastructure." In L. Bud-Frierman, ed. information Acumen: The Unde~tandingandUse of ~ n o w l e ~ineModernBusiness (pp. 231-247). London: Routledge. Bawker, Geoffreyand Susan Leigh Star. 1%. ma owl edge and ~ n f r ~ t r u c t uinr e InternationalInformation Management:Problems of Classification and Coding." In L. Bud, ed. informa~on Acumen: the Unde~tandingand Use le^ in Modem Business@p. 187-213). London: Routledge. Bowker, GeoffreyC., Les Gasser, Susan Leigh Star, and William Turner (eds.). 1997.Socialscience, Technical Systems and C~peratiueWork:Beyond the Great Divide. Mahwah, NJ: Lawrence ErIbaum Associates. Bowker, Geoffrey and Star, Susan Leigh. 1999. Sorting Things Out: Classi~cationand its Consequences. Cambridge, MA: MIT Press. Bush, Vannevar.19~/1945.Science-The Endless Frontier:A Report To The President on a ~~~m forPostwar Scienti~cResearch. Washington, DC: National Science Foundation. Callon, Michel. 1989."Societyin the Making: The Study of Technology as a Tool for Sociological Analysis." In W, E.Bijker, T. P. Hughes, and T. Pinch (eds.), The SocialCons~ction Technoof logical System(pp. 83-105). Cambridge, MA:MIT Press. Callon, Michel, John Law, and Arie Rip (eds.). 1986. map pi^ the Dynamics of Science and Technology. London: Macmillan. Cicourel, Aaron V. 1990. "The Integration of Distributed Knowledge in Collaborative Medical Diagnosis." In J. Galegher, RobertE. Kraut, and Carmen Egido (eds.), Intellectual Teamwork: SocialAnd Technol~icalFoundations Of C ~ p e r a ~ uWork e (pp. 221-242). Hillsdale, NJ: Lawrence Er~baumAssociates. Clarke, AdeleE. 1991. Social Worlds/Arenas Theory as OrganizationalTheory, In D. Maines (ed.), Social O ~ a n ~ a t i and o n Social Process:Essays in Honor of Anselm Strauss(pp. 119-lSs>. New York Aldinede Gruyter. Clarke, AdeleE.and Theresa Montini. 1993. "The Many Faces of RU486 Tales of Situated Knowland Human Values,18,42-78. edges and Technological Contestations"Science, T e c h n o l ~ Clifford, James and George E. Marcus. 1986. Writing Culture: The Poetics and Politicso f E t h n ~ a phy. Berkeley: Universityof California Press.

'

B O W R AND STAR of Anonymity and Evaluative Tone on Idea Conolly, T.,L. M. Jessup, & J. S. Valacich. 1990. "Effects Science, 36,689-703. Generation in Computer-Mediated Groups.Man~ement Durkheim, Emile. 1984.The Division &Labor in Society. Translated by W. D. Halls. New YorkFree Press. Eisenstein, ElizabethL. 1979, The Printing Press as an Agent of Change: Communications and Cub tural Transformations in Early Modern Europe. Cambridge, England: Cambridge University Press. Galegher, Jolene, Robert E. Kraut, and Carmen Egido (eds.). 1990. Intellectual Teamwork: Social and ~ e c h n d ~ i c a l F o u n d a t i ~ s & CWork. ~ p e ~Hillsdale, t i v e NJ Lawrence Erlbaum Associates, Groupe G6ode. 1995, Planifier l'kmergence: la formation d'une s~atkgieduns une kconomie des comp~tences des et savoirs,Paris: C E ~ I / C N May. ~, Haraway, Donna. 1985."A Manifesto for Cyborgs: Science, Technology, and Socialist Feminism in the 1980s."Socialist Review,80,65-107. FemaleMan-Meets-OncoMouse: FemiHaraway, Donna. 1997. M~est-Witnes~Second-Millennium. nism and Technoscience.New York Routledge. Hiltz, Starr Roxanne and Murray Turoff. 1993/1978. The Network Nation: Human Com~unication Via Computer(rev.ed.). Cambridge, MA:MIT Press. Hutchins, E. 1995. Cognition inthe Wild.Cambridge, M A MIT Press. Huysmans, J. K. 1981.A rebours. Paris: Imprimerie Nationale. Jessup, L. M, and D. A. Tansik. 1991. "Decision Making in an Automated Environment: The effects of Anonymity and Proximity with a Group DecisionSupport System." Decision Sciences, 22, 226-279. Jones, Steven G. (ed,). 1998. ~bersociety2.0: Revisiting Computer-Mediated Communication and Community.Thousand Oaks, C A SAGE. Kiesler, S., J. Siege1 and T. McGuire. 1984. Social Psychological Aspects of Computer-Mediated Communication. American Psychol~ist,39, 1123-1134. Kiesler, Sara (ed.). 1998. Cultures of the Internet.Mahwah, NJ: Lawrence ErlbaumAssociates. Kohler, Robert E. 1994, Lords of the Fly: Drosophila Genetics andthe ~perimentalLife. Chicago: University of Chicago Press. Kraut, Robert, Carmen Egido, and Jolene Galegher. 1990. "Patterns of Contact and Communication in Scientific Collaborations."In J. Galegher, Robert E. Kraut, and Carmen Egido (eds.), IntellectualTeamwork:SocialAndTechnologicalFoundations of C~perativeWork (pp. 149-172). Hillsdale, NJ: Lawrence ErlbaumAssociates. Latour, Bruno.1987.Science in Action: Howto Fdlow Scientists and Engineers T h r o ~Society. h Cambridge, MA: Harvard University Press. Latour, Bruno. 1993. We Have NeverBeen Modern. Translated by Catherine Porter. Cambridge, MA: Harvard University Press. Lave, Jean and Etienne Wenger. 1991. Situated Learning: LegitimatePeri~heralParticipation.Cambridge, England: Cambridge University Press. Lave, Jean. 1988. C~nition in Practice: Mind, Mathematics, and Culture in EverydayLife.New York Cambridge UniversityPress. Lederberg, Joshua and Keith Uncapher. 1989.Towards a National Collaboratory: Report of an Znvitational Workshopat the Rockefeller~niversity,March 17-18. : Restudied,Urbana, University of Illinois Lewis, Oscar, 1951. Life in a Mexican V i l l ~ eTepoztldn Press. Marx, Karl. 1970. Capital, Vd.I. London: Lawrence and Wishart. N e ~ o r k ~ N R Research E N ~ : and McClure, Charles, et al. 1991.The National Research and Education Policy Perspectives.Norwood, NJ: Ablex. Middleton, D. S. and D. Edwards (eds.). 1990,CollectiveRemem~ering:Memory in Society. London: SAGE.

24.SOCIALISSUES

IN DESIGN

Olson, GaryM. and Daniel E. Atkins. 1990. "Supporting Collaboration with Advanced Multimedia ElectronicMail: the NSF EXPRESS Project." In J, Galegher, Robert E. Kraut, and Carmen Egido (eds.), Intellectual Teamwork: Social andTechnol~ical~oundationsOFC ~ p e r a t Work i ~ (pp. 429-451). Hillsdale, NJ Lawrence Erlbaum Associates. Theory andColOlson, Gary, Ray Levitt, and Laurence Rosenberg.1993.Intr~uction] C~rdination la~ora~on Techn~ogy Workshop,NSF, July 8-10. Rader, Karen. 1994. "The Production, Distribution and Uses of Genetically-StandardizedMice for Znnouation. Pasadena, C A CalResearch, 1909-1962. In D. Barkin (ed.), Ideas, Instruments and Tech Workshop Conference. A Mexican ~llage.Chicago: Universityof Chicago Press. Redfield, Robert. 1947. Tepoztl~n: Rice, R. E. and Rogers,E. M. 1984. "New Methods and Data for the Study of New Media." In R. E. Rise & Associates (eds)., The New Media: Communication, Research, and techno lo^. Beverly Hills: Sage. Rodden, Tom. 1993. "Technological Support for Cooperation." In D. Diaper and Colston Sanger (eds.), CSCW in Practice:An Introduction And Case Studies(pp. 1-22). Berlin: Springer-Verlag. Ruhleder, Karen.1995. "Reconstructing Artifacts, Reconstructing Work From Textual Editionto On-line Databank"Science, T e c h R o l ~ and H u ~ a nValues,2#,39-64. Schatz, B. 1991. Building an electronic community system.Journal OFManagement InFormation Systems, 8,87-107. Science Policy Study, Background Report No. 5.1986. The Impact oFI~Formation T e c h n o l on ~ Science. Report prepared by the Congressional Research Service, Library of Congress, transmitted to the Task Force on Science Policy; Committee on Science and Technology, U.S. House of Representatives, 9 9 t h Congress, Second Session, Serial T, September. Siegel, J.,V. Dubrovsky, S, Kiesler, and T. W. McGuire. 1986. Group Processes in Computer-Medi37, 157-187. ated Communication. Organ~ationalBehauior and Human Decision-Processes, Simmel, Georg.1950/1~8. "The Stranger."In Kurt Wolff (ed.).The Sociologyof George Simmel (pp. 402-408). Glencoe, I L Free Press. Kiesler, Sara and Lee Sproull. 1996. Connec~ons:New Ways of Working inthe etw worked Organization. Cambridge, hhk MIT Press. Stacey, Margaret.1969. "The Mythof Community Studies."British Journal oFSociology, 29,134-147, Star, SusanLeigh. 1989.Regions of the Mind: Brain Research andthe QuestFor Scientific Certainty. Stanford, C A Stanford University Press. Star, Susan Leigh. 1991. "Power, Technologies and the Phenomenology of Standards: On Being Allergic to Onions." In J. Law (ed.), A Sociolqgy of ~onsters? Power, T e c h n o l and ~ the Modern World:S o c i ~ ~ i cReview al Monograph 38. Oxford: Basil Blackwell, 27-57. Star, Susan Leigh (ed.). 1995. The Culturesof Computing. Sociological Review Mon Oxford: Basil Blackwell. Susan Leigh Star. 1999. "The Ethno~raphyof Infrastructure," American Behauioral Scientist,43, 377-391. Star, SusanLeigh and Karen Ruhleder.1996. "Steps towardan Ecology of Infrastructure:Design and Access for Large Information Spaces." InFormatioRSystems Research,7,111-134. Star, SusanLeigh and Alaina Kanfer. 1993. ~ ~Gemeinscha~ a l or Electronic Gesellscha~?: Analyzing an Electronic CommunitySystem For Scientists. Paper presented at theAnnual Meetings of the Society forthe Social Studyof Science (4S), Purdue University, West Lafayette, Indiana. Strauss, Anselm. 1978. "A Social World Perspective."Studies in Symbolic Interaction, I, 119-128. Suchman, Lucy.1987, Pfans andSituatedAction. Cambridge, England: Cambridge University Press. Tombaugh, J. W.1984. "Evaluation of an International Scientific Computer-Based Conference. Journal oFSocialIssues, 40,129-144.

BOWKEX AND STAR

und Gesellscha~~. Translated and Tannies,Ferdinand. 1957. Communivand Sociev (Gemeinscha~ edited by Charles P. Loomis. East Lansing,MI: Mich~ganState University Press. Traweek, Sharon.1988. ~eamtimesand Lifetimes: The WorldofHighEnergy Physicists.Cambridge, M A Harvard University Press. Scientific Collaboration: A Report on Research in Weedman, Judy, 1992. Origins of ~u~ti-sector Progress. Paper presented at the Annual Conference of the International Network for Social Network Analysis, Dallas,TX, June. Wiener, Norbert.1948. ~bernetjcs;or, Control and Communicationin the Animal and the Machine. New York Wiley. Young, Michael Duniop and Peter Willmott. 1957. Family and ~ i n s in h East ~ London. London: Routledge andKegan Paul. Winner, Langdon. 1980. Do Artifacts Have Politics?~aedalus,19\09,121-136. i~ and C~nition.Norwood, Winograd, Terry and Fernando Flores. 1987. U n d e ~ t a n d Computem NJ: Ablex, Yates, JoAnne and Wanda J. Ortikowski. 1992. “Genres of Organizational ~ommunication:A StructurationalApproach to Studying Communicationand Media.” Academy ofMan~ement Review, 17,299-326.

C H A P T E R

Hsinchun Chen

University of Arizona Bruce R.Schatz

University of Illinois

Despite the usefulness of database technologies, users of online information retrieval systems are often overwhelmed by the amount of current information, the subject and system knowledge required toaccess this information, and the constant influx of new information 171. The result is termed infor~ation o v e r l [3]. ~ ~A second difficulty associated with information retrieval and v ~ a b ~ roble^, l a ~ whichis a conseinformationsharingistheclassical quence of diversity of expertise and backgrounds of system users [6,2l, 221. Previous research in information science and in human-computerinteractions has shown that people tend to use different terms (vocabularies) to describe a similar concept-the chance of two people using the same term to describe an object or concept is less than 20% [22].The ‘‘fluidity” of concepts in the scientific and engineering domains, further and vocabularies, especially c~mplicatesthe retrieval issue16, 14,201. A scientific or engineering concept may be perceived differently by different researchers and it may also convey different meanings at different times. To address the information overload and in a large informationspace that is used by searchers the vocabulary problem of varying backgroundsa more proactive search aidis needed. The problems of information overload and vocabulary difference have become more pressing with the emergence of the increasingly popular internet resource discovery services and digital libraries [16,19,25]. ~etrievaldif-

CHEN AND SCHATZ

ficulties,webelieve, will worsen asthe amountofonlineinformation increases at an accelerating pace under the National Information Infrastructure (NII), Although Internet protocols and World Wide Web software support significantly easier importation of online information sources, their use is accompaniedby the adverse problemof users not being ableto explore andfind what they want in an enormous information space [2,4,30]. The main information retrieval mechanisms provided by the prevailing resource discovery software and other information retrieval systems are either based on“keywordsearch”(invertedindex or full text) or “user r e c i s i (the o ~ proportionof the browsing.’’ Keyword search often causes ~low retrieved documents thatare judged relevant) and poor recall (the proportion of the relevant documents thatare retrieved) due to the limitations of controlledlanguage-basedinterfaces(thevocabularyproblem)andthe inability of searchers themselves to fully articulate their needs. Furthermore, browsing allows usersto explore only avery small portionof a large and unfamiliar information space, which,in the first place, was constructed based on the system designer’s view of the world.A large information space organized based on hypertext-like browsing can also potentially confuse and disorient its user, giving rise to the “embedded digression problem”; and can cause the user to spend a great dealof time while learning nothing specific, the “art museum phenomenon” [S, 181. This research aimsto provide a semantic, concept-based retrieval option that could supplement existing information retrieval options. Our proposed approach is based on textual analysisof a large corpusof subject vocabdomain-specific documentsin order to generate a largeof set ularies. By adopting cluster analysis techniques to analyze the cwccurrenceprobabilities of thesubjectvocabularies,asimilarity(relevance) matrix of vocabularies can be built to represent the important concepts and their weighted “relevance” relationships in the subject domain[g, 1’7,291.To create a networkof concepts, which we refer to as the ‘koncept space” for the subject domain (to distinguish it from its underlying “information/object space”), we propose to develop general AI-based graph traversal algorithms .,serial, optimal branch-and-bound search algorithms or parallel, Hopfield net like algorithms) and graph matching algorithms (for intersecting concept spaces in related domains) to automatically translatea searcher’s in preferred vocabularies into a set of the most semantica~ly relevant terms the database’s underlying subject domain. By providing a more understandable, system-generated, semantics-rich concept space (as an abstraction of the enormously complex information space) plus algorithms to assist in concept/information spacestraversal, we believe we can greatly alleviate both information overload and the vocabulary problem. In this chapter,we first review our concept space approach an algorithms in Section 2. In Section 3, we describe our experience in ~

25. CONCEPT- ED I N F O R ~ A ~ OACCESS N

using such an approach. In Section 4, we summarize our research findings and our plan for building a semantics-rich~ n t e ~ p afor c e the Illinois Digital Library project.

To alleviate information overload and the vocabulary problem in information retrieval, researchersin human-computer interactions and information science have suggested expanding the vocabularies for objects andlinking vocabularies of similar meanings. For example, Furnas et al. [21, 223 proposed “unlimited aliasing,” which creates multiple identities for the same underlyingobject. In informationscience,Bates [l] proposedusing a dom~n-specificdictionary to expand user vocabularies in order to allow users to ‘kiock” onto the system more easily. The general idea of creating rich voca~ulariesand linking similar ones together is sound and its usefulness has been verifiedin previous research as well in asmany real-life information retrievalenvironments(e.g.,referencelibrariansoftenconsulta domain-spec~fic thesaurus to help usersin online subject search). However, the bottleneck for such techniques is often the manual process of creating vocabulari~s(aliases) and Iinking similar or s y n o n ~ o u ones s (e.g., the effort involved in creating an up-to-date, complete, and subject-specific thesaurus is oftenoverwhelmingandtheresultingthesaurusmay ~uickly become obsolete for lack of consistent maintenance). Based on our experiences in dealing with several business, intelligence, and scientific textual database applications, we have developed an algorithmic and automatic approach to creating a vocabulary-rich dictionary/th~ saurus, which we call the concept space. In our design, we concept spaceby first automatically extracting concepts (terms) from texts in domain-specific databases. Similar conceptsare then linked together by using several elaborate versions of cwccurrence analysis of concepts in texts. Finally, through generating concept spaces of different (but somewhat related) domains, intersecting common concepts, andp r o ~ d i ngraph ~ traversalalgorithmstoleadconceptsfrom asearcher’s domain (que~ies expressed in his or her own vocabulary) to the target database domain, the concept space approach allows a searcher to explore a large information space more effectively. We present a blueprintof this approach next. Selected algorithms and examplesare given for illustration. The first task for concept spaceg~neration used in the textual documents. P)techni~ue§ have been used f

CHEN AND S W T Z

detailed, unambiguous internal representation ofEnglish statements [28, 321. However, such techniques are either too computationally intensive or

are domain~ependentand therefore inappropriate for identif~ngcontent descriptors (terms, vocabularies) from texts in diverse domains.An alternative method for concept identification that is simple and domain-independent is the automatic indexing method, often used in information science for indexing literature [291. Automatic indexing typically includes dictionary look-up, stop-wording (removing function words), and term-phrase formation. Another technique (often called "object filtering") which could supplement the automatic indexing technique involves using existing domain-specific keyword lists (e.g.,a list of company names, gene names, researchers' names, etc.) to help identify specific vocabularies used in texts. In our previous Worm Community System (WCS) project [131, we incorporated several thousand worm biology abstracts to generate a worm concept space.A sample entry is shownin Fig. 25.1. The indexin automatic indexing and object filtering for thisentry are shown in Fig. 25.2. The parentheses indicate the number of occurrences of the specific term. (For algorithmic details, please see [131.) ace ~ e f f e ~ a ~ i o Although f f ~ automatic indexingand in texts, the relative importance of object filtering identify vocabularies used each term for representing concepts in a document mayvary. That is, some of the vocabularies used may be moreim~ortantthan others in conveying meanings. The vector space model in information retrieval [29] associates with each term a weight to represent its descriptive power (a measure of importance). Based on cluster analysis techniques, the vector space model

J ~ r n A~ bl ~ t r a ~

-

Li t type : Journal " Reference : "3 Author: "Abdulkader N;Bruin JL" Title: "Induction, detection and isolationof temperature-sensitive lethal and/or sterile mutants in nematodes. 1. The free-living n e ~ t o d eCaenorhabditis elegans" Date: "1978" Journal: "Rev. N m t o l " stract: "Applying a series of techniques intended to induce, detect lethal and/or sterile temperature-s the self-fertilizing h e ~ p h r o d i t e Bergerac strain (Abdulkader et Bruin, 19761, 25 mutants were found. Optimal conditions for the applicationo f mutagenic treatment and the detecti of such mutations are discussed," Sourc~: "Rev. N to1 1978 l(1): 27-38" 'I

'I

FIG.25.1. A sample journal abstract used in concept space generation.

FIG. 25.2. Indexedtermsgenerated for a sample journalabstract.

CHEN AND SCHATZ

could be extended for concept space generation, where the main objective is to convert raw data (i.e., terms and weights) into a matrix of “similarity” measures between any pair of terms [17,271. The similarity measure computation is mainly based on the probabilities of terms cwccurring in the texts. After terms were identified in each documentin the WC3 project, we first computed the term frequency and the document frequency for each in term a document. Term frequency, tf4,represents the number of occurrences of Term j in Document i. Document frequency, df;.represents the number of documents in a collectionof n documents in which Termj occurs. j in Document i, dg, We thencomputedthecombinedweightofTerm based on the product of term ~equencyand inverse d ~ u m e n t ~ e q uase foln~y lows:

where Nrepresents the total number of documents in WCS and wjrepresents the number of words in Descriptor j. Multipleword terms were assigned heavier weights than singleword terms because multipl~wordterms usually conveyed more precise semantic meaning than singleword terms. We then performed termcwccurrence analysis basedon the asymmetby Chen and Lynch [g]. ric “Cluster Function” developed

q k indicates the similarity weights from Term j to Term k. d# was calculated based on the equation in the previous step. dilkrepresents the combined weight of both Descriptors j and k in Document i. dgkis defined similarly as follows:

k thenumber of occurrences of both Term j and Term k where ~ g represents in Document i (the smaller number of occurrences between the terms was chosen). d{k represents thenumber of documents (in a collectionof N document§)in which Termsj and k occur together.wj represents thenumber of

In order to~ e ~general ~ l terms ~ e(terms that appearedin many places) in the c~oc~urrence analysis, we developed the following weighting scheme which is similar to thei n ~ e r ds ~~ ~ ~~ ee n~ t~ function: e ~ c y

25.

CONCEPT-È” ED INFO~ATIONACCESS

Terms with a higher dfkvalue (more general terms) had a smaller weight-

ing factorvalue,whichcaused the,urrenceprobabilitytobecome smaller. In effect,generalterms W $l down in the c~ccurrence table (terms in the co~ccurrencetable were presented in reverse proba-

bilistic order, with more relevant terms appearing first). Sample results from the cluster analysis procedure are shown in Fig. 25.3. * I

tal problem in information retrieval is to link the vocabularies used by a searcher (thosehe or she considers most comfortable and natural to use to

FIG. 25.3. Sample results from worm concept space generation.

CHEN AND SCHATZ

express his or her own information need) withthe vocabularies used bya system (i.e., indexesof the underlying database). By creating a target concept space using the texts of the underlying database (e.g., a C elegans worm database) and another concept space from texts representative of the searcher’s reference discipline(e.g., human genome, fly genome)and intersecting and traversing the two concept spaces a~gorithmically,we believe we will be able to create a owle edge able online search aide that is capable of bridging vocabulary differences between a searcher and a target database, thereby helping alleviatethe information overload problem ina large information space. In previous research we have tested a serial branc~-andbound search algorithm and a parallel Hopfield-like neural network algorithm for multipl~thesauriconsultation [lo,111. The Hopfield algorithm, in particular, has been shown to be suitable for concept-based information retrieval. The Hopfield net [23]was initially introd~cedas a neural network that can be usedas a content-addressable memory. Knowledge and information can bestored in singl~layered, interconnected neurons (nodes) and weighted synapses (links) and can be retrieved based on the Hopfield networ~s ~arallei~ e l ~ a ~and i o conver~enc~ n methods. The opfield nethas been used s~cces~fully in such pplications as image classification,character recognition, and robotics [2 ,311. Each term in the network-like thesaurus can be treated as a n e ~ r o nand the asymmetric weight bet wee^ any two terms is tion between ne~rons.Using field algorithmactivates their

n-l

i=O

25. CONCEPT-B~EDINFO~ATIONACCESS

essing, and by creating concept spaces for the target databases and other related subject disciplines (i.e., preprocessing selected source textual documents), it is possible for a system to help searchers articulate their needs and to retrieve semantically (conceptually) relevant information. *

We have tested the proposed techniques in several domains, most recently in the contextof the Worm Community System. ~ ~ s s i Q ~ Inc19, o lo], ~ ~we~generated t i ~ ~ a: Russian computin~concept space basedon an asymmetric similarity function we had developed. Using the indexes extracted from about 40,000 documents (200 several weeksof CPU time ona V M VMS minicomputer, we were erate adomain-specific Russian computing thesaurus that containeda 20,000 concepts (countries, institutions, researchers' names, a areas) and 280,000 ~eightedrelationships. We performed a memory-association experiment, comparin~the (concept) recall and precision levels of the concept space and those of four oc computing experts in associating concepts (for50 ssian computing terms).~ o ~ crecQ11 e ~ is t defined as t 4

CHEN AND SCHATZ

Community System ~ C § [13). ) It took about4 hours of CPU time on a DEC Alpha 3000/600wor~stationto analyze5 , ~ O worm abstracts and the resulting worm thesaurus contained 845 gene names, 2,095 researchers’ names, and 4,691 subject descriptors. ~e tested theworm thesaurusin an experiment with six worm biologists of varying degrees of expertise and background. The biologists were asked to suggest terms relevant to a list of 16 previously selected worm-specific concepts (one term at a time). The worm thesaurus~suggested terms were by thebiologists.Theexperiment thencomparedwiththosegenerated showed that the worm thesaurus was able to help suggest more terms for retrieval (on an average, from6.2 terms to 8.5 terms).The worm thesaurus was significantly better than the subjects in concept recall, but inferior to the subjects in concept precision. ~e also found that the worm thesaurus was an excellent memory-jogging tool and that it supported learning and serend~pity browsing. Hy co~ceptspace g e ~ e r a ~As o ~an: extension of the worm thesaurus project and in an attempt to examine the vocabulary problem across differentbiologydomains,werecentlygenerated a fly thesaurususing ~,OOO abstracts extracted from Medline and Biosis and literature from FlyBase, a databasecurrently in use by molecularbiologists in the ~ r o s o p ~ i l ~ ~elanog~ster-related research community [121. The resultin includedabout 18,000 terms(researchers’names,genenames,function names, and subject descriptors) and their weighted re~ationships. In a similar fly thesaurus evaluation experiment (using 10 pre-selected flyspecific concepts) involvingsix fly researchers at the University of Arizona, we confirmed the findings of the worm experiment. The fly thesaurus was found to be a useful tool to suggest relevant concepts, improve concept recall, and to help articulate searchers’ queries. However, fly subjects were better in concept precision than thefly concept space. * ~ y - ~ concept o r ~ space aver sal: Our structural comparison of the fly and worm thesauri revealed a significant overlap of common vocabularies acrossthetwodomains.However,eachthesaurusmaintainsitsunique organism-specificfunctions, strtures,proteins,and so on.Table25.1 shows the numbersof terms in t orm and fly thesauri,These include the numberof authorterms, numb geneterms,numberofsubject terms, number of function terms(fly thesaurus only), and umber of method terms (worm thesaurus only). The lastthree columns report the number of terms appearing in both thesauri and the respective proportion of each thesaurus thlappingtermsrepresent. It is notsurprisingthatnooverlapexists in ames:thenamingconventionsforthetwodomains are e~remely different, Furthermore, it is noteworthy that 252 author names appear in both thesauri, The format for author names is last name and first initial, 0

25. CONCEPT-BASED I N F O ~ ~ ACCESS ON

TABLE 25.l

Structural Comp~sonof the f l y and Worm Thesauri

or^ Terms

2095

03

Authors

Genes

25.8Subjects 32.0 Functions Methods Total 18101

3.4

845 4691503 l n/a

12

Fly

Terms 722 12.0 1 4875 5821 182 n/a

~verlapping Term 252 0 n/a n/a

Percentage of Worm Term

Per~~n~age O f W

Terms

I _

-

-

_ .

7657

which could present some ambiguity. Still it is likely that some authors have published in both domains. The extent of overlap for the subject descriptors was 25.8%for the fly thesaurus and 32% for the worm thesaurus. With this much overlap,thelikelihood of finding intermediatetermsforconcept space traversalis promising. We believe that by intersecting concepts derived from the two domainspecific concept spaces and by providing AI search methods wewill be able to bridge the vocabulary differences betweena searcher's (e.g., fly a biologist's) domain and the target database's (e.g., the worm database's) subject area. We are in the process of testing and fine-tuning several search algorithms [111and wealso plan to expand our subject coverage to other model organisms including e. coli, yeast, rat, and humanin the near future. We have incorporated a worm thesaurus aand fly thesaurus into the WCS Release 2. An windows thesaurus-browsing interface which can accept multiple terms, identify other relevant terms by means of the thesaurus, combine the weights associated with terms, and rank terms in order, has been developed, as shown in Fig. 25.4 (the thesaurus component is'shown on the right side of the screen). Searchers can use the Thesaurus Window to elicit suggested terms from the thesaurus. In a recentC, eZe~ff~s rneetin held at the Universityof Wisconsin at son (with about 650 participa S), the WCS and the worm thesaurus were used extensively and perceived favorablyby the worm biologists during the %day live demo sessions. They were able to consult the worm thesaurus whenevertheywishedtoexploreothertopicsrelevant to their initi~l ate students and postdoctoral researchers,in particuinterest in the ~ C and S the thesaurus, especially for their potential valueto help them quickly enter a new and dynamicresearc

25. CONCEPT-B~ED I N F O ~ ~ ACCESS ON

area. Many viewed the WCS as a comprehensjve, tightly integrated digital library with an online reference librarian(i.e.,the thesaurus and its browsing interface).

The concept space approach to supporting semantic retrieval has recently been extended from the National Collaboratory domain to Internet and Digital Library applications.We highlight the statusof our research and future directions next.

TheWCS was created using proprietary GUI software and suffers from its poor portability. Recently, we have migrated part of the WCS design to a more portable and popular interface, theNCSA (National Center for Supercomputing Applications) Mos internet resource discovery softt, exhibits full-textsearchingcapaware [30]. The prototype,calle bilitythrough WAISindexingandsearchingandMosaic’s form~apable a thebrowser. In addition, it provides semantic retrieval capability through ~ioQuestconsists of saurus component anda thesaurus browsing interface. about 4,000 worm documents (e.g., Worm Breeders’ Gazette, conference proceedings abstracts, etc.), a 7,OO~term worm thesaurus, and a l$,OO~terrnfly thesaurus. Readers are encouraged to access the ~ioQuestURL at the following address: http://bpaosf.ba.arizona.edu:$~O/cgi-bin/Bio~uest The following sequence of screendumps illustrates the semantic retrieval In a pilot study involvinga process supported by the thesaurus camponent. in searching “worm junior worm biologist, the subject expressed an interest genes which determine sex, especially those that cause feminization.” The user first entered “sex determination” in the ~ i o Q ~ emain s t searchscreen, as shown in Fig. 25.5. The option she chose was to search “Worm Document” (as displayed in the middleof Fig. 25.5). The system searched the worm database, which was indexed using the WAIS inverted inde~ngscheme, and displayed 40 top-ranked documents. Each document was assigned a document id (e.g., ~ 3 $and . the ~ scoring scheme was based WAIS, on with 1000 bein the highestdocumentscore),Forexample,Document15 ~ 4 6 3 5 , ~ tis” ) shown in Fig. 25.6. After browsing these retrieved documents, the user did not find any abstracts of relevance to feminization and decided to invoke the system’s thesaurus component,as shown in the middle of Fig. 25.7 (search“Worm-~ly

sex d @ t @ r m ~ ~ i ~ n (I

able irdcx of inform~ion.Nore: This svvice CM OB@ be useal/rom a

FIG.25.5. System response when “sexdetermination’’ enteredin the BioQuest main search screen.

.. "

/oi/tna/~/dot./in6//m16/463Sf t x t

1I

variable, but clearly show

FIG. 25.6. hampie of document found as a result of using Usex determination" entry.

CHEN AND SCHATZ

of ~ o r m ~ i o &Utn : ?%is svvlct CM only be usedfmm a

Score: 632,lM 3 3 , Size: l Bytes,Type:text flle ize: 7 ~

~

~text file ,

T

~

:

FIG.25.7. Worm thesaurus invoked with user using%ex determination^ entry.

Thesaurus”). Several thesaurus terms related to “sex determination” were retrieved, including “SEX-DETE~~INATION-GENE.” In VVAIS indexing, hyphenation of a term was treated differently, for example, “sex-determination gene9’ was indexed as “§EX-DETE~INATION~ENE.~.txt” and “sexdeterminationgene”(withouthyphen)wasindexed as “§EX”ETE GENE.txt,”Clickingon “ § ~ - ~ E T E ~ I N A T I O N - G produced E N ~ specific

25. C O N C E P T - B ~ EI~~ F O ~ T ACCESS IO~

terms (gene names) such as “FEM-1,”“FEM-2,” and “FEM-3” (genes which cause ~e~inization) as shown in Fig. 25.8. All related terms were ranked in reverse weighted order and the sources of these terms were indicated by “f” for fly thesaurus or“W”for worm thesaurus. For example, the thesaurusindicated that “FEM-1,” “FEM-2,”and “FEM-3” are worm terms only.By using the newly obtained terms, “FEM-1,” “FEM-2,” and “FEM-3,”to search the worm database, ~ i o ~ u eretrieved st many more documents which were relevant specifically to genes that cause feminization, as shown in Fig. 25.9. The user browsed a few documents and was satisfied with the search results.

TheIllinoisDigitalLibraryInitiative (DLI) project entitled: “ B u ~ l d i nthe ~ Interspace: Digital Library Infrastructure fora University Engineering Community,”is one of six projects funded recentlyby NSF/A~A/N~A [26]. The goal of the project is to evolve the Internet into Int~rs~ace, the in particular by bringing professional and semantic search and display of structured documents to the Net.To accomplish this, weare constructing a large-scale digital library testbedof SGML journal articlesin the engineering domain while concurrently performing the underlying research to provide effective interaction across networks with such a library[30j. The semantic retrieval component of the Illinois project aims to create graphs of domain-specific concepts (terms) and their weighted cwccurrence relationships for all major engineering domains. Merging these concept spaces and providin~traversal paths across different concept spaces could potentially help alleviate theu ~ u 6 ~ l a ~ ~ ~ i f f eevident re~ce~ ~ in large-scale information retrieval. In order to address the scalab~lity issue related to large-scale information retrieval and analysis for the current Illinois DL1 project, we recently proceeded to experiment with using the concept space approach on several parallel supercomputers. Our initial test collection was24- GBs of computer science and electrical engineering abstracts extracted fromINSPEC the database and the concept space approach called for extensive textual and statistical Inianalysis based on ~uto~atic indexing and cooccurrence anulysis algorithms. tial testing results using a 512-node ~ h i n ~Machine) ng C and a 16-processor SGI Power Challenge supercomputer were promising. The user-friendly shared-memory microprocessor-basedSGI Power Challenge (a cluster of SGI workstations), in particular, is appropriate for the memory-intense digital library analysis and was later selected to generate a large-scale computer engineering concept space of about 2 7 0 , terms ~ and 4 , 0 ~links , using ~ ~ 24.5 hours of CPU time. Preliminary results have been posted at the~ ~ u e s server at:http://ai.bpa.arizona.edu~tml/csquest/

5 6 7 8

{0.28966) [0.26945) (0,254601 (0.241291 9 (0.231641

[fw]

[ W] [ W) ( W) [ W)

~ 1 2 ( 0 . 1 7 9 0 ~ ) f W] GENE ~ 13 [ 0 . 1 7 5 8 8 ] [ W) FEW2 14 to.174131 f f w ] 15 [ 0 , 1 7 3 8 9 ) f f w ] 16 [0.15544] [ W] 17 [0.151951 [ f w ] 1 8 [O.l5l351 [ W] 19 t0.149001 f W ) ~~~,M. 20 [0.14490) [ W) SQimL, T. 2 1 ( 0 . 1 4 3 0 7 ~ { W] O ~ P, ~ 2 2 (0,141331 ( f w ] 23 (0.136061 l f w ] 24 ( 0 . 1 1 9 7 5 ) [ f w ] 2 5 [0.10703) [ f w ] T 2 6 (0,103891 ( f w ] S 2 7 [O.lOlOS) ( f w ] I 2 8 (0.098891 [ W] W 29 (0,095931 [ f w ] €3 30 $0.095671 f f w ] 31 [ 0 . 0 8 9 9 7 ] [ f w ] 32 f o . 0 7 8 4 l l f f w ] 33 [0.07648) [ f w ] 34 (0.05651 J ( W] 35 (0,052161 f f w f ~ ~

U

~

R

~

,

~

P

~

FIG. 25.8. Worm-thesaurus suggested terms which are relevant to "sex determination gene."

~

T

25. CONCEPT-B~EDI N F O ~ T I O NACCESS

S e l e c t M index to ;x

:22, Size: 8439 bytes. Type: .. lext rile :S. Size: 122 byles. Type: -. lext m e

5: 6

:1ubytes. Type: text file

:22,Sfie: 936 bytes. Type: text fife ..

:6. Size: 160 bytes, Type: text fife

FIG. 25.9. Documents identified as relevant to FEM-1, FEM-2, and FEM-3.

Our recentsystemevaluationoftermassociation invol~ng12 knowledgeable subjects revealed that the automatically created computer engineering concept space generated sig~ificantly higher concept recallthan the human-generated INSPEC thesaurus (concept space :INSPEC thesaurus = 69.08% :17.71%). However, the INSPEC thesaurus was more precise than the :INSPEC thesaurus = 59.50% : automatic concept space (concept space

CHEN AND SCHATZ

Using the INSPEC thesaurus as the benchmark for comparison, we believe the computer engineering concept space has demonstrated its robustness and potential usefulness for suggesting relevant terms for search. However, we are convinced that multiple interfaces and multiple vocabulary search aids are necessary for effective concept-based search across multiple largescale repositories and domains. Our current work mainly involves: (1) creating concept spaces for other major engineering domains(in roughly the followingorder: chemical, materials, systems and industrial manufacturing, mech~ical,aerospace, automatic, mining, marine, and nuclear civil, agricultural and biosystems, geological and using the &processor SGI Power Challenge Array, 16-processor ,and Wprocessor Convex Ekemplar (all accounts already have with NCSA); and (2) develo~in~ robust graph matching and trathms for cross-domain, concept-based retrieval. Future work also will include generating individualized concept spaces for assisting in user-specific concept-based information retrieval and developing genetic algorithm-based search agents for users, based on the Java will be incorporated into an operational language. Results from our research L search interface for the IllinoisDL1 engineering testbed.We are also investigating methods by which this semantic retrieval capability might be Net or theNII). extended and scaledup to large distributed repositories (the

This research was mainly supportedby the following grants: NSF/ARPA/NASADigitalLibraryInitiative, I~-941131$,1994-1998 (B. Schatz, H. Chen, et. al, “Building the Interspace: Digital Library Infrastructure for a University Engineering Community”), * NSFCISE, 1~-9525790, 1995-1998 (H. Chen, “Concept-based Categorization and Search on Internet: A ~achineLearning, Parallel Computing Approach”), * NSF CISE ResearchInitiationAward, 1~-921141$,1992-1994 (H. Chen, “Building a Concept Space for an Electronic Community System”), * NSFCISE Special Initiative on Coordination Theory and Collaboration techno lo^, 1 ~ - 9 0 1 ~ 01990-1993 7, (B. Schatz, ‘‘~uildinga National Collaboratory Testbed”), * AT&T Foundation Special Purpose Grants in Science and Engineering, 1994-1995 (H. Chen), and * National Center for Supercomputing Applications (PICSA), High-performance Computing Resources Grants, I ~ 9 5 ~ 31994-1996 N, (H. Chen). *

25. CONCEPT-B~EDINFO~ATIONACCESS

Most of theseprojectsweremadepossiblethroughthesupportand vision of late Dr. Larry Rosenberg of the National Science Foundation.

[ l ] M. J. Bates. Subjectaccess in online catalogs: a design model. Journal oftheAmerican Society forInformation Science,37(6):357-376, November 1986. [2] T.Berners-Lee, R. Cailliau, A. Luotonen, H. F, Nielsen, andA, Secret. The World-Wide Web. Communications of the ACM,37(8):76-82, August 1994. [3] D. C. Blair and M. E.Maron. An evaluation of retrieval effectiveness for a full-text documentretrieval system. Communications of the ACM,28(3):28~299,1985. [4] C. M. Bowman, P.B. Danzig, U. Manber, andF, Schwartz. Scalableinternet resource discovery: research problems and approaches. Communications of the ACM,37(8):98-107, August 1994. [5] E. Carmel, S, Crawford, and H. Chen. Browsing inhypertext: A cognitive study. IEEE Transactions on Systems, Man and Qbernetics, 22(5):865-884, September/October 1992. [6] H. Chen.Collaborativesystems:solving the vocabulary problem. IEEE ~ M P ~ E ~ , 27(5):58~6,Special Issue on Computer-Supported Cooperative Work (CSCW), May 1994. [7] H. Chen and V. Dhar. User misconceptions of online informationretrieval systems. International JournalofMan-Machine Studies, 32(6):673-692, June 1990. [8] H. Chen, P. Hsu, R. Orwig, L. Hoopes, and J. F. Nunamaker. Automaticconcept classification of text from electronic meetings. Comm~njcationsoftheACM, 37(10):56-73, October 1994. [g] H. Chen and K. J. Lynch. Automatic construction of networks of concepts characterizing document databases. IEEE Transactions on Systems, Manand Qbernetics, 2 2 ( 5 ) : ~ 5 ~ 2 , September/October 1992. [101 H. Chen, K. J. Lynch, K. Basu, and D. T. Ng. Generating, integrating,and activating thesauri Intell~ence for concept-based document retrieval.IEEE PER^ Special Series on Artificial in Tit-based Information Systems,8(2):25-34, April 1993. [111 H. Chen and D. T. Ng. An algorithmic approach to concept exploration in a large knowledge network (automaticthesaurus consultation): symbolic branch-and-bound vs. connectionist Hopfield net activation. Journal of theAmericanSociety for InformationScience, 46(5):348-369, June 1995. [121 H.Chen, B. R Schatz, J. Martinez, and D. T.Ng. Generating a domain-specific thesaurus automatically:An experiment onFlyBase. In Center forMan~ement Information] of College of Business and PublicAdminis~ation]~niversityOfArizona,Working Paper,CZMI-WS 94-02,1994. [131 H.Chen, B,R.Schatz, T. Yim, and D. Fye. Automaticthesaurus generation for an electronic community system.Journal of the American Societyfor Information Science,46(3):175-193, April 1995. [141 J. Courteau. Genomedatabases. Science, 254.201-207, October 11,1991. [151J. Dalton and A. Deshmane. Artificial neural networks. IEEE Potentials, 10(2):33-36, April 1991. [161 0. Etzioni and D. Weld. A softbot-based interface to the Internet, Communications of the ACM, 37(7):72-79, July 1994. [171 B. Everitt. Cluster Analysis. Second Edition, Heinemann Educational Books, London,England, 1980. [181 C. L. Foss. Tools for reading and browsing hypertext. Info~mat~on Processing and Management, 25(4):407"418,1989. [191 E. A. Fox, R. M. Akscyn, R.K Furuta, and J. L. Leggett. Digitallibraries: introduction. Communicutions oftheACM, 38(4):22-18, April 1995.

CHEN AND SCHATZ [20] K. A. Frenkel. The human genome project and informatics. Communicat~onsof theACM, 34(11):41-51, November 1991. 1211 G.W. Furnas. Statistical semantics:How can a computer use what people name things to guess what things people mean when they name things. In ~ o c e e d i of ~ the s ~uman Factors in Co~puter Systems Conference, pages 251-253, Gaithersburg, MD, Association for Computing Machinery, March 1982. [22] G, W. Furnas, T. K. Landauer, L.M. Gomez, and S. T. Dumais. The vocabulary problem in human-system communication.Communicationsofthe ACM, 30(11):964-971, November 1987, [23] J. J. Hopfield. Neural network and physical systems with collective computationalabilities. Proceedings of the ~ationalAcademy of Science, USA, 79(4):2554-2558, 1982. [24] K. Knight. Connectionist ideas and algorithms. Communications of the ACM, 33(11):59-74, November 1990. [25] K.Obraczka, P. B. Danzig, and S. U.Internet resource discovery services. IEEE C O ~ P ~ E R , 26(9):8-24, September 1993. [26] R. Pool, Turning an info-glut into a library. Science, 266:20-23,7 October 1994. [27] E. Rasmussen. Clustering algorithms. In Information Retrieval: Data Structures and A k a rithms, eds., W.B. Frakes andR. Baeza-Yates, Prentice Hall, Englewood Cliffs, NJ, 1992. [28] N. Sager. ~atural Langu~e IRfo~ation Processing: A Computer Grammar of English and Its A~p~ications. Addison-Wesley, Reading, M A , 1981. [29] G.Salton. Automatic Tat Processing. Addison-Wesley Publishing Company, Inc., Reading, MA,1989. (301 B. R Schatz and J. B. Hardin. NSCA Mosaic and theWorld Wide Web: global hypermedia protocols for the internet. Science, 265:895-901,12 August 1994. I311 D. W. Tank and J.J. Hopfield. Collective computation in neuronlike circuits.Scientific American, 257(6):104-114, December 1987. [32] W.A, Woods. An experimental parsing system for transition network grammars. In ~atural Langu~eProcessing,pages 113-154, ed. R. Rustin, Algorithmics Press, New York,NY, 1972.

C H A P T E R

ary N.Olson Daniel E. Atkins Robert Glauer Terry Weymouth Thomas A. Finholt Atul Prakash Craig Rasmussen Farnam jahanian

University of Michigan

The Upper Atmospheric Research Collaboratory (UARC) is a project to provide upper atmospheric physicists with networked computer technology to support the practiceof their science. UARC is an instance of a collaboratory, the ‘‘...combination of technology, tools and infrastructure that allow scientiststo work with remote facilities and each other as if they were co-located . ,”(VVulf, 1989, p. 6). A National Research Council (1993) report defines a collaboratory as “a...center without walls,in which the nation’s researchers can perform their research without regard to graphical location-interact in^ with colleagues, accessinginstru~entation, sharing data and co~putationalresources [and] accessing informationin digital libraries”(p. 7). Put simply,a collaboratoryis the useof computing and communication technology to carry out geographically distributed scientific activity. For more background on collaboratories, seeFinholt and Olson (1997).

OLSON ET AL.

Collaboratories provide scientists with access to each other, with access to information bases (digital libraries), and with access to remote facilities (seeFig. 26.1). Throughout thehistory of UARC the focus has been on the s~nchronouscommunication for the people-to-people and peopleto-facilities componentsof the collaboratory. There have been two broad phases to the project.From its inception in 1993 until about 1996 the space in this project were all users of scientists who were the testbed users ground-bas~dinstruments at theSondrestrom UpperAtmospheric ResearchFacility in Kangerlussuaq,Greenland. We providedthemwith real-time access to data from these instruments as well as conferencing so they could interact with each other over the data they were capabilit~es observing. The collaboratory tools were all builtanddeployed in the ~ e ~ T S tprogramming ep environment. Around 1996 the project entered a second phase. The user technology base shifted to Java-applets accessed from a Web browser, which enabled a considerable expansionof the user population. The science shifted from a single ground-based site to include multiple datasources, both land-based andsatellite, coveringmuch of the northern hemisphere. Real-time access to model outputs wasalso added. The present chapter focuses on the first phaseof UARC; the second phase is still in progress.

FIG.26.1. A summary of the main re~ationshipsin the conceptof a col~aborator~

26. D I S T ~ B ~ ETEAM D SCIENCE

The UARC project has a seriesof interrelated research goals: Design methods: To discover systematic and detailed approaches to the design of effective team-science support systems; to verify these designs throughtheanalysis of realuse of prototypesbased on user-centered, object-oriented principles. Architectures for colla~oratio~: To discover the supporting architecture required by a closely cooperating groupof scientists with complementary expertise and goals who need to examine multiple instruments with specialized datadisplays; toimplement this architecture to allow operations over a wide span of time zones via networks with low bandwidth and~ossibilities for failures; to implement an architecture that allows expansion of the numbers and types of instruments and users. 3ehavi~ralchanges in science due to new techno lo^^ To understand the effects of the introductionof collaboration technology on the patterns of scientific collaboration over the Syear duration of this project. Impact on upper a ~ ~ p h e r~hysics: ic To provide support for upper atmosa pheric physicistsin performing timely and coordinated observations using variety of ground based instruments; to use the UARC system to increase understanding of solar influences on the Earth’s highly conductingupper atmosphere and the coupling of these influences to lower atmospheric layers.

These goals have required the tight synergy of a researchteam comprised of computerscientistsandengineers,behavioralscientists,andupper by rapid cyclesof softatmospheric physicists. The project has been paced ware development to support the physicists: creationof new software versions, deployment for conducting space science experiments, systematic evaluationof actual use, and subsequent modification of new versions based on users’ experiences. Since the spring of 1993, the UARC testbed has gone throughnumerouscycles of development,approximately 5 monthsper cycle. UARC 5.3 was the final versionof the initial roundof software develwhereas startingin 1996 the opment in the NeXTStep software en~ronment, UARC software migrated toa seriesof distributed servers accessed through Java applets activated via a World Wide Web browser.

Upperatmosphericphysicistsstudythe interac~onsamongthe earth’s upper atmosphere, the earth’s magnetic field, and thewind. solarA common ‘~orthernlights.” manifestation of this interactionin the aurora borealis, or

OLSON ET AL.

The physicists develop models based on observations from ground-based instruments, satellites,and rockets. The UARC project initially focused on a community of upper atmospheric physicists who used ground-based instruments in Greenland. The Sondrestrom Upper Atmospheric Research Facility has as its core instrumentanincoherent scatter radar that is supported jointly by the National Science Foundation and the Danish Meteorological Institute. SRI International in Menlo Park, California, provides overall management for the facility. In addition, there area number ofother instruments at the site that are managed by a variety of principal investigators. Figure 26.2 shows a view of the Sondrestrom facility, The radar is the most complex and expensive instrument at the site. It operates about 150 hours per month, and must be attended by the local staff. Its operation is usually scheduledin advance by making requests to SRI. There are many modes in wh7ch theradar operates, andsomeofthemrequire extensive real-timedecision-ma~ng d~pending on ionospheric conditions. Otherinstrumentsvary in theircomplexityandneedforinteraction. Some, such as the ima~ingriometer (IRIS), run 24 hours a day, 365 days a year, and have no settings that can be varied. Others, such as optical instru-

FIG.26.2. The observatory in Greenland. This picture h ~ g h l ~ gthe ~ t sradar and the buildings in which additionalobservational equipment is housed.

26. D I S T ~ ~ ~TEAM E DSCIENCE

ments like the all sky camera, are used during darkness, which at Kanger-S lussuaq is abundant during the winter months but scarce in the summer. Somehavedifferentmodes of operationoradjustableparameters. For example, various optical filters tha# canset beon the allsky camera. In the past, most data collection required the physicists to bein Greenland to operate the instruments and monitor ionospheric conditions. If multiple instruments were involved, several scientists might arrange to be in Greenland at the same time. The physicists call such coordinatedactivity a data campaign.A campaign usually hasa particular scientific focus and may involve simultaneous observations using several instruments. Campaigns are usually scheduled to take advantage of particular viewing conditions (e.g., moonless nights) or to coincide with other data collection events (e.g., a satellite passing overhead). Withina campai n period observations often take place only when relevant conditions are present. Campaigns involving tion of the radar with optical instrumentsare particularly frethe winter months. In 1990 NASA installed in a 56 Kb data link to the Sondrestrom facility. This enabled access to data on local discs over the network, and opened the door to the possibilityof remote interactions with the instruments. The initial set of UARC users came from five sites: the Danish Meteorological Institutein Copenhagen, the University of Maryland, the University of Michigan, SRI International in MenloPark,California,andLockheedPalo Alto, alsoin California. During theearly years severalnew sites were added: Corneli University, the University of Alaska, the University ofNewHampshire, Phillips Laboratory, Florida Institute of Technology, and the High Altitude Observatory in Boulder, Colorado. Thus, evenby the end of the early phase of UARC the user community had grown considerably.

Discussions about the possibility of using the network conqection to Sondrestrom to develop a collaboratory project beganattheUniversity of ~ichiganin 1991. In October of 1991 a workshop sponsoredby the National Science Foundation provided an occasion for allof the principals to gather and discuss the feasibilityof such a project. This led to the definition of the overall project goals described earlier. The participants in the workshop submitted a formal proposal for a collaboratory project toNSF, and it was funded as a cooperative agreement in September of 1992. Subcontracts were arranged forSRI,DMI,Maryland, and Lockheed. The project was formally launched with a workshop of all key participants in late 1992. Project goals were reviewed and detailed plans were made. In addition, extensive work was done on gathering information about

OLSON ET AL.

the ,practicesof the physicists and developing user scenariosto be used in designing theearly versionsof the software. ~uring the earlymonths of the project further observations were made of the upper atmospheric research community, both to develop specifications for the software and to document the practices of the community prior to the introduction ofnewtechnology.These observations included regular visits to the labsof participating scientists.In addition, an extended observation was made of a campaign team in Greenland over a 1O.day period in arch 1993. This was just prior to the introductionof the initial versionsof the UARC software. In parallel with early observations of the scientists, programming staff was assembled, and a prototype of the radar viewing software was developed. An early project decision was to develop software using the NeXTStep programming environment, with distributed objects writtenin Objective C. The rapid prototyping capabilityof this software development environment was an extremely important feature, as it facilitated the evolution of the software in response to how the scientists actually used it. The initial version of the ARC application allowed users to access radar data in real time, and provided a simplepublicchatwindowforcommunication. All messages exchan~edthrough the chat window and all user actions were time stamped and saved in log files These data (which we continue to collect) have provided important information about whether new UARC designs meet the needs of physicists as they conduct observations. Thedecision touse a homo~eneouscomputingenvironmentforthe entire project was possible because of the small number of sites. Thisof course simplified interoperability. The project equipped each site with a NeXT workstation. Subsequent expansionof the project to additional sites used ~ e X ~ trunning e p on Intel 486 platforms. As mentioned, this phase of the project endedin 1996 whenUARC software development shifted atonew architecture and a Web-based user environment. The NeXTStep version of the software was available after 1996, and a few users continued to use it. The initial versionof the software was first usedin a scientific campaign in April of 1993. A senior space physicist wasin Greenland, and his graduate student participated from Ann Arbor, ~ichigan.A member of the development groupobserved thiscampaign from Greenland. The campaign provided valuable user performance data, and extensive revisions were made in the UARC software. In June of 1993 this revised versionof the software was used in another campaign with the same scientific focus, but with the same senior physicistin Ann Arbor and the same studentin Greenland. Based on the June campaign, further extensive revisions to the software were made.In addition, two new instruments, the imaging riometer and the magnetometer, were added.By the fallof 1993 campaigns using these three instruments were supported.

26. D I S T ~ ~ TEAM ~ E DSCIENCE

In December of 1993 the second annual project workshop was held. Prototype versionsof new collaboration capabilities, an annotation feature and the ability to share windows, were demonstrated. Shortly after this workshop NSF held an external reviewof the project and approved plans for the remaining years of the project. During 1994 the UARC software developed amidst extensive experience with users, By the annual workshopin Ann Arbor in December of 1994 the basic design of UARC 5.0 was set. UARC 5.3 was the final versionof the software built using the NeXTStep environment. On the basis of experience gained during the first 3 yearsof the project, the technical goal shifted to develop a more generic toolkit to support a broader rangeof collaboratories across multiple platforms. This toolkit was called the Col~aboratoryBuilders En~ironment(CBE). The UA itself was rebuilt during 1995-1996 into a seriesof modules that capture the key functionality and interface clients that allow the UARC displays to be shown on any platform froma Web browser. This has allowed for considerable expansion of users and data sources, allowing UARC functionality to extend to other ground-based facilities, satellite data, and model outputs. Additional information aboutCBE is described in Lee, Prakash, Jaeger, and Vvu (1996).

As mentioned earlier, theUARC project has been committed to a user-centered, iterative, rapid-prototyping design strategy, employing object-oriented methods, Although various textbooks on object-oriented analysis and design describe development methods (e.g., Coad& Yourdan, 1991a, 1991b; Jacobson, Christerson, Jonsson,& Overgaard, 1992), the methods are presented at a fairly abstract level, and it can be difficult to know exactly how to proceed. Therefore, we devoted considerable attention to the development of concrete design methods. These methods, described in more detail cDaniel Olson, and Olson (1994), used ideas from object~rientedanalyn, business process reengineering, and human-computer interaction. A detai~edset of steps were worked out, ~onsistingof: *

*

collection and distillation of use cases, essentially user scenarios for how they do their work specificatio~of w~ichaspects of the work should be automat priority should be put on these functions in design of software modules that implement these specifications

OLSON ET AL.

7 ft

e ft

applicationofinterfaceanalysistechniquesfromhuman-computer interaction rapid prototyping of the software analysis of usertests of the prototypes iteration through the entire sequence again as a resultof user testing

This strategy has resulted in useful and usable software for our user community.

of Figures 26.3 and 26.4 show screen dumps of the final NeXTStep version the U M C a~plication.Figure 26.3 shows (clockwise from the upper left corner): (in the tenthe main menu, two radar data displays, the radar status window ter), theIRIS image window, anAll Sky Camera image window,a Fabry-Perot interferometer data display, a display of three tim~~arying data streams

FIG.26.3. A screen dump from UARC5.2 duringan observational campaign in

February of 1995. Several instrument data displays and status displays are shown.

26. DIST~BUTEDT W SCIENCE

FIG. 26.4. Showsa screen dumpwith several displaysdepicting substorm activity during a campaign in early November of 1993.

from a magnetometer, and theUARC message window (the “chat window”). Figure 26.4 shows anarray of IFUS displays depicting interesting features during an event which is calleda magnetospheric substorm and is usually associated withvery activity aurora near midnight local time.

In this chapter we focus on the architectureof the most mature versionof

the ~eXT§tepsystem. This system hada number of good properties, but it based versionsof UARC were intended tosolve these problems.

ber of clients (see Fig. 26.5). The path of data from the

instr~mentto the

OLSON ET AL.

FIG.26.5. Overview of the architecture.

26. D I S T ~ B ~ EW D SCIENCE

the clients. Hence, collaboration was supportedby client to client communication. The system was based on a set of interacting objects. For example, in the flowof data from instrumentto view, there were objects representing: the instrument course, the data package, the client instrument connection, theinstrument display manager, and the particular views of the data. The overview of the architecture shownin Fig. 26.5 depicts the major objectsand the connectionsof data flow among them. In actuality, there were a number of client-server relations in UARC A breakdown of the detailsof the servers is as follows: ~ ~ s ~ ~Seruer: ~ e each ~ tinstrument ~ ~ t had f aftask or task-set which gath-

instrument, packaged it for delivery on the net and supplied it to the centralinstrument~ataserver, the Routing Server. ~ o ~Sertter: t i The ~ routing server maintained the list of all user applications that were using the system. It had four functions:(1) it maintaineda list of clients who were currently running; a client could register or un-register, and this placed the client on the user a~plicationlist or took them off; clients were sent a copy of the user application lista n ~ i mite was updat (2) it maintaine~a list of all instrument data servers currently prov~di data; all clients were given the current oflistinstruments whenev~rt h ~lit changed;(3) it routed any objects sent to it to all thea user

OLSON ET AL.

Fttbry-Perot

0,001

FIG. 26.6, Data rates for the principle instru~entsin Greenland.

‘~ummary” or “quick look” data for the high data volume instruments, the incoherent scatter radar and the allsky camera. In the case of the radar the data sent was that needed to generatea rough sketch display just like the display which was onsite at the operator console in Greenland. The image from mera was compressed usingP E G , a standard for “lossy” comfull data from the radar and all sky camera were recorded locally in Greenland for later transfer via ftp or the shipment of backup media. These instruments were handled by instrument dataservers. In all cases, the ~ackagingof the data was done through NextStep objects. First the data wereparsedintoaninst~ument-specific object for the graph controller object and data view object of the User Client Application (see the following). Then it was ‘%rapped”in a generic dataobject called a “stream” so that NextStepcouldtake care ofserializingthedatausing a readandwrite method for each type of object (which is defined when the object is imple mented). NextStep streams could be “sent” to remote objects as ar~uments ributed~bjectmethods. In this way the implementation of communibetween parts ofthe system was done at the object level ~ithoutconcern as to whether the objects are running the on same machine or not. The user client consisted of a central controller dispatcher, instrument stand-inmonitorsforeachinstrument,multipleinstrumentdisplays,the messagewindowmonitor, a shared windowmonitor,andanannotations monitor. It servedas a client for the Routing Server, the Mess Annotation Server, and the Shared ~indowServer. It cont ment proxy for each instrument type that unpacked the instrument data into a data object that was used as the source for data display. ment data objects were passed to the graph controller that intermediate information and layout for the data displays. The data views object then generated the actual display, presented the user interface, and passed the results of user actions. The annotation server received annotations from user clients, created u n i ~ ide~t~fiers ~e for them, createdan object for each annotation, and sent out ap~ropriateinformation about each annotation to all active user clients. The annotationobject kept the following information for each annotation: an

26. D I S T ~ ~ TEAM ~ E D SCIENCE

er id that identified each annotation uniquely; the creator of the anno;an optional subject field; the instrument for which this annotation was made; the modein which this instrument was being operated; and the time and date of the data stream for which the annotation was attached.A storage object kept trackof the name of the user, the text entered, and the time it was entered for each annotation or amendment made by users. Qn the client side an Annotation Coordinator in the user client communicated with the AnnotationServer, kept a local copyof the annotation object list, and displayed and maintained the Annotation Browser and Annotation ~indows.The notation Coordinator was connected to one Annotation Manager for each instrument. These Managers performed the instrument specific coordinate conversions to display the new annotation in the displays. An Annotation Interface object responded to mouse downs and button clicks, changed the cursor to the annotation cursor, and drew annotation icons. The Annotation Interface objects of each instrument talked to the instrument specific Annotation Manager. Annotations were not saved beyond a session unless users specifically save them, although a plausible extension would have been to keep the annotations in a database that would allow them to persist beyond a session and would giveusers thefull range of search and retrieval capabilities. The shared window capability was based aonprototype implementation of DistView (Prakash & Shim, 1994), a toolkit that provideda shared window server,shared window manager, application object manager, export window manager,andimportwindowmanager.Thesharedwindow server of DistView played the role of the bookkeeper for exported windows. It was responsible for creating export and import window managers when requested by the applicationin which it wascreated.An application object manager supported the application object replication and state transfer operations. An exportwindowmanagerprovidedseveralprimitivesforthe membershipandinterfaceobjectreplicationandstatessnchronization services. The export window managerrecursively followed the connections between the interface objects of the exported window retrieving the type, size, location, and state information and sent the acquired information to each client import in^ thewindow, An importwindowmanager requests for the importing of windows and requested the export the state of the correspond in^ interface object and tri~gereda 1 pointi~ capability ~ was more effectively and efficiently a c ~ i ~by~ ~e odm m ~ -

OLSON ET AL.

nicating low-level events so that smooth traces of the cursor movements were produced on exported windows.In order to prevent the applications from being blocked due to the telepointing facility, we adapted a nonblocking protocol for telepointing messages in which a message was droppedif it could not be immediately sent. This scheme may temporarily have left the views of shared windows out of synchronization. However, the number of messages for the cursor movements was large enough that the synchronization was often quickly restored. In practice, we found that the overall quality of the regenerated traces of the cursor on imported windows did not sig~ificantly suffer evenif some messages were dropped. As mentioned, this system architecture worked well in the early periodof ,when the number of users and instruments was modest. Wefound that as the numberof active users grew beyond ten and the number of instruments increased beyond five performance with the system seriously degraded.High network traffic loads increased the severity of the proble~. Thisledtothedesignofthe CBE architecture and its subsequent implementation in the Web phase of UARC These issues are discussed in more detail in Hall, Mathur, Jahanian, Frakash, and Rasmussen (1996) and Lee, Prakash, Jaeger, andWu (1996).

The ~ e ~ ~ version § t e pof the UARC software was used extensively from April 1993 until the springof 1996. We captured the behaviorof our users in several ways. First, all user actions with the software were recorded and time stamped in an action log. Second, the contents of the message window were saved by the message server for later analysis. Third, we hired behavioral observers in each of the major sites, and they directly observed the physicists using the software. The observers often asked questions about the software or madenotesabouthowthe scientists used it. §ometimes they videot~pedthese sessions for later analysis. Fourth, users themselves volunteered their reactionsto the software, often via email. §tar~ingin the fall of 1994 the software itself had a feature where users could report bugs and suggestions directly within the application to make this even easier, and it was used extensively. As noted earlier, We focused our behavioral observations on campaigns. the early campaigns in 1993 focused on looking for convection boundary radar data were available over the network. By the fall of d magnetometer data were available, and in the earlywinter ering data from the Fabry-Ferot Int erometer were available. in 1993-1 the UARC software W nter c a ~ p a i gseason ~ ly. Some particularly noteworthy events were a January19

26, ~ I ~ T ~ TEAM B ~ SCIENCE E D

campaign with a number of significant events that led to the conduct of a replay campaign in March 1994 (a specific descriptionof this follows).In February 1994 two space scientists from outside theUARC sites came to Michigan to conducta campaign rather than goto Greenland. One ofthese scientistswasatheoreticianwhohadneverbeforeseendatacollection in progress in real time. Throughout this time monthly World Days, when the Greenland radar operates in a standard, predefined mode, were observed by scientists throughout theUARC network. In Mayof 1994 another convection boundary reversal experiment was done. In the summer and fallof 1994 further extensive revisionsof the software were made. A full complement of Fabry-Perot Interferometer data displays were added, the allsky camera was brought on line,a separate operator’s window to support the Greenland site crew was added, the annotation and shared window capabilities were added, and numerous small fixes and additions were made to make the client application more useful and usable and to make the server more reliable. To give an impression ofwhat these campaigns were like, we include some data obtained between April1993 and November 1995, during the use of the Ne~T§tep system for access to the Greenland facility. More recent campaigns using the Java-based implementation aand much wider range of users and data feeds have not been thoroughly analyzed Theyet. first setof data are based on analyses of the message window files.We coded the content of the messages usinga five-category coding scheme: Science-space science phenomena, data, methods §ession-scheduling, timing, planning * Techno~o~-UARC system, NeXT, network * Display~rienting to data displays * §ocializing-greeting, jokes, personal matters

* *

This scheme was developed from earlier experiences with coding conversations while people used collaborative technology (e.g., Olson, Olson, Storrgsten, & Carter, 1993). Coders were trainedso they had acceptable reliability in the coding of the messages. Figure 26.7 shows the distribution of messages by these categories for four different classes of users, summed over all campaigns between April 1993 and November 1994. It is encouraging that for the scientists themsel~es the most frequent communications were about the science itself. This suggests that the UARC software was servinga useful purposein the conductof their science, a suggestion confirmed by detailed observationsof their work and discussions with themduring and after campaigns* The huge amount of discussion of the techno lo^ by the Others is because these are the pro-

OLSON ET AL.

I

FIG. 26.7. Messages classifiedby category of Usage over four different classes of user, summarized over all the major campaigns,1993-1994.

grammers and the behavioral scientists concerned with the development of the software itself, who often used the message window to query users online about the functionality and usability of the software. Indeed, analyses of the content of these Technology discussions has played a major role in the iterative developmentof the software. Figure 26.8 shows the three convection boundary reversal experiments mentioned earlier. This shows only the data for the scientists. The distribution of message typesis quite constantover this l~monthperiod, with Science messages dominating throughout. This is with three different versions of the software. Figure 26.9 shows data fora number of other campaigns carried out during this period.We discuss the replay campaign shortly in more detail. How do these electronic conversations compare with earlier face-to-face ones? We made such a comparison by using transcripts of face-to-face conversations made at the Sondrestrom site in March of 1993 with a selectionof episodes from electronic conversations using the UARC chat facility. The episodesselectedforboththeface-to-faceandelectronicconversations

26. D I S T ~ ~ TEAM ~ E DSCIENCE

0.7

0.6 0.5 0.4

0.3 0.2

0

FIG. 26.8.Messages

by category in threeconvectionboundaryreversal exper~ments. Note: this involved three different versions of the software.

were from campaigns that involved similar science goals. The conversations focused primarily on the real-time data from the radar. In the case of the face-to-face conversations the data were displayed on a bank of monitors along a wall, whereasin the UARC case they were in windows on the individual pa~ici~ants workstations. Figure 26.10 shows the relative amounts of conversation in the five coding .Both kinds of conversation are dominated by science talk, parhen i~terestingthings were happeningin the displays.~uringthe times when upper atmospheric activity was limited, the face-to-face groups tended to socialize,whereasthe UARC U d totalkaboutthetechnology.Suchtalkwasaboutimprovements,problems,andwish 1' added functionality. Interestingly, there sociali~ingin the chat window. Most of it was greetings and goodbyes as participants came and went. Butthere were periodsof jokes, teasing, weather discussions, and even a discussion of an on oing NGAA tournament basketball

OLSON ET AL.

FIG.26.9. Messages by category for additional campaigns, 1993-1994.

26. DIST~BUTEDTEAMSCIENCE

Science

T e c h n o l ~ ~ y splay

Session

ocial

FIG. 26.10. Comparison of content for face-to-face (FTF) and computer-mediated (CMC) conversations.

are very similarforthetwokinds of conversations, though significantly there is a much longer tailin the caseof UARC,indicating that thetec~nology allowed a large number of “lurkers,” people who observed butdid not partic~pate very much. This long tailis of course quite important, as it shows how it is possible for the techno lo^ to bring the real-time practice of science toa much broader community. out to be useful betweencam~aigns. The UARC software has alsoturne~ Figure 26.12 shows the extentof use of the software between campaigns for between~ampaignusage focused,of course, on instrum ,such as the day that a caribou W

t h ~ r ewa§ a campai~nin inter~stto the §cientists. of the scienti~ts iganforotherpurposes,several ssible to replay the 2 days in ~ ~ e s t i so o nt in real time could be examined m

35 *6

* *

30

-

.- .

. " .

S

*

6

25 20 15 10

.............

5

^.. . . . .

.

-

..................

...........

.................

........... ." ......................................

0 FIG. 26.11, Percentage of time taken up by different participants, ranked by frequency.

onto UARC for 14 en^

1

FIG. 26.12. Use of the UARC software by thesix most active users,1993-1995.

26. D I S T ~ ~ ' ~I E" DSCIENCE

Arbor. The principal participants were in Copenhagen, Boston, and Menlo Park. The participants reported that it was extremely valuable to reexamine the session, being ableto fast forward over quiet periods, pause or replay interesting periods, and converse through the message window about the phenomena. This mode of operation had not been anticipated. However, the scientists have reported that this kind of replay campaign is useful both for the science itself and for the training of students and young scientists. We decided not to explicitly support this capabilityin the current NeXTStep-basedversions of the UARC software. This capability is allowed for in the new CBE architecture, but has not yet been implemented. C for Graduate Student

One feature of the UARC software that is potentially very important is its role in the trainingof graduate students. Previously, when data were collected by means of trips to Greenland graduate students would rarely be ableto participate in data collection.But with this phase of the science available online students can participatein any campaign, evenif only as an observer. This allows students to observe the interactions among a wider range of scientists than is available in their home laboratory, and indeed we have seen episodes of intense interaction between a student in one site and a senior scientist in a different site. Someof our recent extensionsof UARC software to new sites will enhance and elaborate the its use for student training. The early UARC software was used extensivelyby a growing community of scientists for real-time data acquisition from Greenland. This provided a foundationforexpandingthecommunity of UARC users, makingsimilar capabilities available for other sites where space scientists collect data, and making UARC available for the examination and analysis of archived data. Thus, the early phase of UARC represented a strong beginning for transforming the practiceof science. The technical and social visionof UARC were of course overtakenby the explosive growthof the World Wide Web in the 1994-1995 time period, leading to new phase in the project. The change to a Web-based system also enabled much wider access. Indeed, manyof the space physics sites-both ground-based and satellite-based-on their own began providing real-time data feeds over the World Wide Web, which made it easierUARC for to incorporate a wide range of inputs withina unified framework. This enhanced the functionality of UARC and made it useful to an even broader com~unityof users. Early indications are that this broader scope will bring about even

OLSONET AL.

more dramatic changes in science practice. But this more recent story is still in progress,and will require further experience and analysis before we understand thefull significance of the later periodof UARC development.

The UARC projectissupportedbya Coop~rativeAgreementfromthe National Science Foundation. Many people have been involved in the development, testing, and deployment of the UARC software and the longitudinal study of UARC users, including Susan McDaniel, Eric Heldman, Stephanie Mackie-Lewis, Michael Gallo, Torben Elgaard Jensen,Efrat Elron,Joe Magee, ~ t r i n Wade, a Galye Farbman, Bill Mott, Bill, Shim, Ramani, Bob Sitar, and Aaron Ridley.We are deeply indebted toa number of people at theNational Science Foundation for their stewardshipof the project. We owe a special debt to Larry Rosenberg, who was responsiblefunding for and managing the NSF side of the UARC Cooperative Agreement during its early years.Subsequently Su-Shing Chen and Les Gasser have served in this role. Y. T.Chien also provided advice and counsel throughout the project.

Coad, P,,& Yourdan, E. (1991a) Object~rientedanalysis. Englewood Cliffs,NJ: Yourdan Press. des@. Englewood Cliffs, NJ YourdanPress. Coad, P., & Yourdan, E. (1991b) Ob~ecf~riented Finholt, T. A., & Olson, G. M. (1997) From laboratories to collaboratories:A new organizational j c a ~8,28-36. form for scientificcollaboration^ ~ s y c h ~ ~Science, Hall, R W, Mathur, A., Jahanian,F., Prakash, A., & Rasmussen, C. (1996) Corona:A communication service for scalable, reliable group collaboration systems. In Proceedings of CSCW '96 (Nov. 16-20, Boston, MA). ACM, New York, pp. 140-149, Jacobson, I., Christerson, M., Jonsson, P., & Overgaard, G. (1992). Ob~ect~rieRted s o ~ ~ aengire neering:A use case drivena p p r ~ c hReading, . M Addison-Wesley. workspaces Lee, J. H,,Prakash,A., Jaeger, T.,& Wu, G. (1996) Supporting multi-user, multi-applet in CBE. In Proceedings OF C S W '96(Nov. 16-20, Boston,MA). AGM, New York, pp. 344-353. McDaniel, S. E., Olson, G. M,, & McGee, J. (1996) Identifying and analyzing multiplethreads in computer-mediated and face-to-face conversations. In Proceedings of CSCW '96(Nov. 16-20, Boston, MA). ACM, New York, pp. 39-47. McDaniel, S, E., Olson, G. M,, & Olson, J. S. (1994) Methods in search of methodolo~~ombining HCI and object orientation. In Proceedings OFCHI '94, ACM, New York, pp. 145-151. National Research Council. (1993) ~ationulcollaboratories:A ~ p ~ i inFormation ng t e c ~ n oFolr~scientific research.Washington, D C National AcademyPress. Olson, J. S., Olson, G, M,, Storrmten, M., & Carter, M. (1993). Groupworkclose up: A comparison of the group design process with and without a simple group editor,ACM ~~Rsactions OF InFormation Systems, 11,321-348.

26. DIST~BUTEDTEAMSCIENCE

Prakash, A., & Shim, H,S. (1994) DistView: Support for building efficient collaborative applications using replicated active objects. In Proceedings of CSCW '94, ACM, New York, pp. 153-164. Sproull, L., & Kiesler, S. (1991) Connections: New waysof working in the network~do ~ a n ~ a t i o n . Cambridge, M A MIT Press. Wulf, W. A. (1989) The national collaboratory-a white paper. Appendix AIn Towards a n a ~ o n a ~ c o ~ ~ a ~ o r aUnpublished to~. reportofaninvitationalworkshopatRockefellerUniversity, March 17-18.1989.

This Page Intentionally Left Blank

A Aalbersberg, I., 1 0 6 , 121 Abbate, J.,726,727, 734, 735 Adelman, L., 674,675,708 Agarwala-Rogers, R., 22,49 Ahiswede, R.,589,594 Ahuja, S. R., 297,306 Aiken, P., 343,366 Akscyn, R. M,,405,407 Allen, J.P., 18,43,5ll, 513,514,515, 516,517,519,520,521,522,524, 528,53 1,532 Allen, T.J., 22,43 Altenburg, K.,413,444 Altmann, E.,602,605,621 Altmann, S., 414,444 Alvesson, M., 52,64 Amabille, T. M.,253,259 Amari, S. I., 589,594 Anderson, I. H.,463,472 Anderson, J.R., 463,469 Anderson, N.,488,504 Argote, L,, 245,259 Argyris, C., 53,64 Arkin, R. C., 411,413,414,417,419, 420,421,424,425,444,445,446 Aron, S., 414,445 Arrow, K.J.,22,43, 198,259 Ashby, W. R., 679,708 Athans, M., 679,702,706,707,708 Atkins, D.E., 722,737 Attewell, P., 29,44,510,531 Aumann, R. J.,23,44 Axelrod, R., 55,64 Azulay-Schwarts, R,, 120, 121 B Bacchus, F., 87,90 Baecker, R. M.,543,556

Bahgat, A. N. F.,474,496,499, 505 Bainbridge, L., 326,337 Baker, B. S., 16,44 Baker, M.G., 400,407 Bakhtin, M. M.,54,64 Bakos, Y., 511,513,514,515,516,517, 518,519,520,521,525,526,531,532 Balasubramanian,V.,501,502 Balch, T. R., 416,419,420,431,425, 444,445,446 Baligh, H. 24,44 H., Banks, D.,32,44 Bannon, L. J.,328,334,337,339,669, 672 Barghouti, N. S., 551,554 Barnard, C. I., 13,24,44 Barnard, P,J.,562,582 Barnes, B., 53,64 Baron, S., 675,709 Barron, A.R., 509,594 Bartlett, F. C., 53,62,64 Bartunek, J., 53,64 Basar, T., 652,671 Basu, K.,746,759 Bates, M. J., 741,759 Baym, N. K.,720,729,735 Beard, D.,32,44 Beckers, R., 414,445 Begeman, M.L., 31,32,44,147,148, 159,453,470 Bell, B., 547,557 Bellotti, V.,488,504,543,555,581,583 Beni, S., 412,445 Benjamin, R. I., 28,48,51,65 Bennett, S, L., 509,531 Bennis, W.G., 559,582 Berg, M.,724,735 Berger, T.,588,589,592,593,594 Berlin, L., 467,468,469 Berners-Lee, T.2 740, 759 Be~stein,P.,40,44

7 Beuschel, W., 513,515,516,522,527, 53 1 Bier, E. A., 342,366 Bigelow, J.,343,366 Bignoll, C., 264,307 Bijker, W E., 717,735 Billard, E., 380,383,386,388,389 Binmore, K.A., 111, 121 Birman, K.P,,403,408 Bishop, A., 714,719,735 Blahut, R. E., 589,593,594 Blair, D. C., 739,759 Blaiwes, A. S., 675,709 Blake, B., 95, 114, 123 Blakeslee, A., 544,554 Bobrow, D. G., 32,49,246,308,449, 472,541,557 Boehm, B., 282,306 Bogia, D. P., 286,297,307 Boland, R. J., 51,53,57,63,64,65 Bond, A. H., 7,40,44,116,121 Borenstein, N., 297,306 Bostrom, R.P., 475,502 Bougon, M. K.,55,66 Boulding, K. E., 53,65 Bowker, G., 715,726,729,731,735 Bowman, C.M,,740,759 Boyer, M., 121, 124 Braybrooke, D., 66 1,671 Brecher, M,,119, 121, 124 Brehe, S., 541,554 Brehmer, B., 3 16,339 Brennan, S. E., 328,337,562,582 Brenner, D., 365,366 Brenner, M. J., 661,67 1 Brett, J.M,,652,656,657,672 Brewer, M. B., 619,620 Brewer, R.S., 286,307 Bridwell-Bowles, L. S., 541,554 Briggs, F. A.,17,45 Briggs, R. O., 316,337 Brobst, S. A., 1,4,27,30,32,47,69,71, 91, 127, 147,160 Brooks, F. P,, Jr., 560,582 Brooks, R. A.,412,419,445 Brothers, L.,286, 306 Brown, R., 54,65 Bruner, J.S., 56,65 Brunner, H.,569,583

AUTHOR INDEX Bruns, W.J., 29,44 Brynjolfsson, E., 28,44,518,5 19,521, 53 1 Buchanan, B. G., 452,469 Bullen, C. V., 509,531 Burdett, K.,593,594 Burton, R. M.,13, 19,24,44,634,647 Bush, P., 546,548,554 Bush, R. R.,602,603 Bush, V., 718,735 Bushman, J.B., 317,337 Bushnell, L. G., 626,647 Buss, M,, 412,445 Bussler, C.J.,669,671 C Cailliau, R., 740, 759 Callon, M,,717,726,735 Calsamiglia, X.,164, 191 Cammarata, S. D., 104,121 Campbell, B.,405,407 Campbell, D., 253,256,259 Cannon-Bowers, J.A., 675,676,678, 679,682,702,706,707,708 Caplinger, M[., 405,407 Capps, M.,365,366 Card, S. K., 132, l60 Carey, L., 541,545,555 Carley, K.M.,596,597,598,599,601, 602,603,620,621,626,640,647, 679,7 10 Camel, E., 740,759 Carraher, D. W., 463,469 Carraher, T. N., 463,469 Carriero, N., 40,44 Carroll, G, R.,596,620 Carter, D.E.,16,44 Carter, M,,264,308,487,488,505,536, 539,556,564,565,566,567,570, 579,584,775,782 Carver, N., 104,121 Castanon, D. A.,673,698,709 Catlin, T., 546,548,554 Cavalier, T.,548, 554 Ceci, S. J., 493,505 Champy, J.,626,647 Chandhok, R.,32,48,263,297,308, 539,543,548,552,554,556

AUTHOR INDEX Chandler, A.D., Jr., 21,44 Chandrasekn, B.,96, 121 Chang, S. C., 692,709 Chapman, D., 18,44 Chen, A.,463,469 Chen, H.,739,740,744,746,747,749, 759 Chid~baram,L., 475,485,495,502 Chimera, R., 548,556 Chionglo, J.,319,337 Choudhary, R,,268,271,272,279,305, 306,543,555 Christerson, M., 767,782 Christiansen,T,R.,618,621,626,627, 63 l, 634,641,642,646,647,648, 679,706,707,710 Christie, B.,578,584 Chu, R.W., 317,337,338 Ciborra, C. U.,10,25,44 Cicourel, A.V.,73 l , 735 Clancey, W.J., 625,647 H., Clark, H. 54,65,328,337,562,582 Clark, K.B., 510,5 12,532 Clark, R.J., 420,445,446 Clark, R. W., 255,259 Clarke, A.E., 717,735 Clarke, E., 36 1,366 Claypool, M.,284,288,289,305,306, 308 Clifford, J., 716,735 Goad, P.,767,782 Cohen, G. P.,618,621,626,63 1,634, 647,648,679,706,707,710 Cohen, M.D., 1,4,27,30,32,45,47, 69,7 1,91,127, 147, 160,596,620, 626,647 Cohen, P.,23,44 Cohen, S. S., 317,337,510,531 Cole, J.R.,493,502 Cole, S., 493,502 Conant, R., 679,708 Conklin, E. J., 462,472 Conklin, J., 31,32,44, 147, 148, 159, 453,470 Connolly, T., 33,48,720,736 Conroy, S. E., 72,91,96, 115, 116, 121 Consolini, P.M., 596,621 Contractor, N. S.,643,648 Converse, S. A.,674,675,709

Cook, l?, 287,306 Coombs, J. H.,404,407 Cooper, R.B.,510,531 Corkill, D. D., 79,91 Cornell, P.,572,584 Couger, J. D., 501,502 Courteau, J., 739,759 Cover, T. M,,589,594 Crawford, A.B.,Jr., 26,44 Crawford, S., 740,759 Croft, W. B., 32,44 Cross, G. A., 545,546,555 Crowley, T. R., 297,309 Crowston, K.G., 11, 12, 13,21,27,39, 44,47,132, 157, 159, 160,268,307, 319,338,370,389,474,504,508, 533,635,648,651,668,672,673, 674,681,709 Croy, M., 414,445 Cruz, G. C., 342,366 Csiszar, I., 589,594 Culler, D. E., 15,43 Curnmings,L.L.,492,502 Curtis, B., 10,39,45,501,502 Cvetanovic, Z , 1 0 4 , 121 Cytron, R., 14,45 Czech, B., 485,503 Czech, R. M.,483,503 D Daft, R.L., 501,502,562,582 Dalton, J., 746,759 Damouth, D., 79,87,90,91 Dantzig, G. B.,42,45 Danzig, P,B.,740,759 Danziger, J,N.,29,45 Darwin, C., 254,259 Davis, R.,34,45, l l , 21, 34,49,104, 115,116, 1 2 4 Dearborn, D. C., 619,620 Debreu, G., 13,25,41,45 Decker, K.,116, 121 Deitel, H.M., 14,40,45 Delbecq, A.,474,506 Delisle, N.M,,405,407 DeMarco, T.,448,470 Demers, A.,551,555 Deneubourg, J. L.,43,45,414,445

7

AUTHOR INDEX

Dennis, A. R.,22,32,45, 294, 306,473, 504 Dertouzos, M.L.,23,45,512,532 DeSanctis, G., 32,45,473,475,476, 478,492,493,495,502,503,505, 1,648 Deshmane, A., 746,759 Desimone, R.,257,259 Dewan, P., 264,266,268,270,271,272, 279,289,297,305,306,308,543,555 Dhar, V.,739,759 Dickinson, T.L.,674,675,709 Dickson, G., 495,503 Diesing, P.,119, l24 Dietterick, T.G., 463,469 Dijkstra, E. W., 40,45,79,91 DiMaggio, P.J.,513,533 diSessa, A. A., 137, 159 Dobler, D. W., 670,672 Dougherty, D., 52,54,65 Dourish, P.,543,555,581,583 Drake, J., 284,287,297,308 Drexler, K.E., 11,34,35,45,48 Drogoul, A., 413,445 Druclcer, P.,5 1,65 Dubois, M., 17,45 Dubrovsky, V., 719,720,737 Dudek, G., 412,445 Dufner, D. K.,483,484,498,503 Dumais, S. T.,739,741 ,760 Durfee, E, H.,41,45,71,73,74,77,78, 79,85,86,87,89,90,91,92,96, 115, 116,122,123,412,446,618,620 D u r ~ ~ i E., m ,716,736 Dutton, W H.,29,45 Dyer, M,,413,446 E

Easton, A. C., 492,503 Ekcles, R.G., 14,45 Edwards, D., 732,736 Edwards, M. R.,548,555 Egido, C., 2,4,578,583,720,721,729, 736,739 Eisenberg, E. M,,619,621 Eisenstein, E.L.,725,736 Elliot, M.,515,520,522,523,526,528, 53 1,532

Ellis, C,A., 15,29,32,45,264,268, 284,287,306,316,337,449,470, 473,503,669,671 Emerson, E. A., 361,366 Engelbart, D. C., 405,407 English, P.NI., 342,366 Ensor, J.R.,297,306 Ephrati, E., 79,91, 115, 117, 118, 122 Epple, D., 245,259 Erman, L.D., 30,41,45,104,115,123 Estrin, D., 10,46 Everitt, B., 740,744,759 F Fadel, F. G., 319,337 Fagan, M.,282,307 Fan, T,,365,366 Farrell, J., 23,45 Ferber, J., ;1.13,445 Ferguson, G. J., 541,556 Ferris, T.,259 Ferront, F., 485,503 Festinger, L.,54,65 Feulner, C., 284,294,297,308 Fikes, R.E., 18,45,34,47, 115, 116, 123,130,159 Finholt, T.A., 3,4,761,782 Fischer, G., 264,307,453,455,456, 464,466,467,468,470,471 Fischer, K.,104, 122 Fish, R.S., 32,45,548,555 Fish, S., 53,65 Fisher, R.,655,668,672 Fjermestad, J.,475,476,485,486,499, 503,506 Fleck, L.,53,65 Flores, F., 30,45,497,503,738 Flower, L.,537,540,541,542,545,555, 556 Ford, C., 485,503 Ford, D. L.,486,504 Foreano, D,, 413,445 Forsdick, H.C., 297,309 Forsyth, D., 560,583 Forte, G., 263,307 Foss, C.L.,740,759 Foster, G., 32,49, 449,472,541,557 Fox, B., 547,557

AUTHOR INDm Fox, M.S., 11,46,319,337,635,647 Franklin, R.F., 413,445 Frankowski, D.,284,288,289,305,308 Franks, N.R.,43,46,414,445 Frenkel, K.A.,739,760 Frieder, O., 106, 122 Frost, P.J.,492,502 Fry, C.,33,47,147,153,160,328,338 Fukuda, T., 412,445 Fukuoka, H., 284,309 Furnas, G. W., 739,741,760 Furuta, R.,341,342,343,351,359,361, 362,365,366,367,405,408

Fussell, S., 54,65 Fye, D.,742,748,759 G

, Gabarro, J. J.536,555 Gadamer, H. G., 63,65 Galbraith, J.R., 10,21,42,516,532,

596,620,624,626,628,637,640,647

Galegher, J., 2,4,539,555,570,583,

720,721,729,736 Gallanti, M.,312, 338 Gallupe, R. B.,32,45,473,475,476,502 Gantmacher, F. R.,257,259 G d h k e l , H., 52,54,65 Gargano, G., 10,49 Gasser, L., 7,40,44,46, 115, 116, 117, 121,122,513,532,618,620,734,735 Gelernter, D.,40,44 George, A.L., 119, 122 George, J.F., 473,504 Gibbons, J.,332,337 Gibbons, R., 15,45 Gibbs, S. J., 29,32,45,264,268,284, 306,316,337,449,470,473,503 Giddens, A.,54,65,643,647 Gillespie, T., 537,544,548,556 Gintell, J.,286,307 Girgensohn,A.,455,456,467,470 Glamm, B.,297,308 Glickman, A. S., 675,709 Glockner, A., 384,389 Gmytrasiewicz, P.G., 89,91 Gochman, C.,119, 123 Goetsch, W., 414,445 Goldberg, S. B.,652,656,657,672

Gomez, L. M.,739,741,760 Goodisman, A.,342,366 Goodman, J.M.,405,407 Goodman, N.,40,44 Goss, S., 414,445 Gould, J.D.,536,555 Goyle, V., 312,314,317,338 Graf, M.,287,306 Grant, K.R.,1,4,34,47,69,71,91, 115,116,123, 127, 137,147,160

Grant, R. A., 527,532 Grant, R.,512,533 Graves, M.,30,45,497,503 Gray, J.,23,46 Greenberg, S., 264,308 Greene, D.,54,65,551,555 Greif, I., 29,46,264,268,279, 307,308 Gronbaek, K.,328,337 Gross, M.D.,463,470 Gross, S., 43,45 Grosz, B. J., 115, 117, 122, 123 Grudin, J.,33,46,264,307,455,466, 467,470,509,532

Gruninger, M,,635,647 Guindon, R.,488,503 Gurbaxani, J., 28,44 Gurbaxani,V., 28,46,511,513,517, 518,519,526,531,532

Guttag, J., 18,47 H

Haake, J.M.,271,307,405,408,549, 555

Haan, B. J.,404,407 Haas, C., 537,541,555 Hackman, J.R.,448,470,560,583,674, 675,708

Hackwood, S., 412,445 Hadamard, J., 253,254,256,260 Hal, R.,674,675,708 Halasz, F. G., 405,407,548,549,555 Hall, J., 486,503 Hall, R.W., 774,782 Halpern, J.J., 22,23,46,49,373,389 Halstead, R.H., 14,46 Hammer, M.,626,647 Han, T. S., 589,594 Haraway, D.,73 1,736

Hardin, J. B., 740,751,755,760 HareI, D.,312,337 Harmon, L.A.,413,445 Harrington, J., 5 11,532 Harris, M.,95, 114, 113, 115, 121, 123, 1 24 Hart, P.,10,46 Hartfield, B., 30,45,497,503 Hartman, J. H., 400,407 Hauser, C.,551,555 Hayes, J. R.,537,540,541,542,545, 555,556 Hayes, R. H., 510,512,532 Hayes-Roth, F.,30,41,45 Heath, C.,562,583 Hedlund, J., 599,620 Helmbold, D.,43 1,445 Heminger, A. R., 486,487,505 Hendler, J., 18,43 Herbsleb, J. D.,564,566,567,569,583 Hewitt, C., 11,35,46,47,51,65,96, 122,536,555 Hicks, D.L.,405,408 Higgins, A.,527,532 Higgins, L.F.,501,502 Hill, W. C.,461,470 Hiltz, S. R., 474,475,476,483,484, 485,486,487,493,495,496,499, 500,503,504,505,506,721,729,736 Hirokawa, R. Y., 564,584 Hirsch, E.D.,23,46 Hoare, C.A.R.,40,46,79,91 Hobbs, J. D.,414,444 Hoffer, H. A.,495,504 Hogg, T.,34,36,46,50 Hollan, J. D.,137, 160,449,450,461, 470,47 1 Holldobler, B., 414,445 Hollenbeck, J. R.,599,620 Holley, K.,113, 115, 121, 122, 124 Hollingshead, A. B., 3 16,338 Hollnagel, E.,3 1 l , 337 Holt, A. W., 10, 31, 39,40,46,343,366 Hong, S., 107, 122 Hoopes, L.,747,759 Hopfield, J. J., 746,760 Hopmann, P.T.,121,122 Horn, D.N.,297,306 House, P., 54,65

House, R. J., 496,504 Howard, J. H., 400,404,407 Howard, M.T.,34,47, 115, 116, 123 Hsieh, H., 365,366 Hsu, E.Y. F?,499,504 Hsu, F?, 747,759 Huber, G. P., 51,65 Huber, M.J., 79,91, 87,90,91,412,446 Huberman, B. A.,7,11,34,36,46,47,50 Huff, A. S., 55,65 Hughes, R.,414,445 Huhns, M.N.,7,40,46,120,121,122 Humm, A.,32,44 Humphrey, W., 282,307 Hurwicz, L.,41,46,164, 192 Hutchins, E.L.,51,65, 137, 160,462, 471,471,562,583,733,736 Hutchinson, N.,28 l ,308 Hymes, C.M.,294,307 I Iacono, S., 510,512,532 Ignacio, E.,714,719,735 Ilgen, D.R., 599,620 Irish, W., 551,555 Irving, R. H., 527,532 Isaacs, E.,578,580,584 Isakowitz, T,,501,504 Iscoe, N.,501,502 Itani, J., 253,260 J Jablin, F. M.,614,619,620 Jacobson, E. S., 342,366 Jacobson, I., 767,782 Jaeger, T.,767,774,782 Jago, A. G., 496,504 Jahanian, A.,774,782 Jaikumar, R.,5 12,532 Janis, I. L.,22,46 Jasek, C.A., 318,321,322,334,337,338 Jeffay, K.,392,408 Jeffries, R.,467,468,469 Jenkin, N.,412,445 Jennings, N.R.,120, 123 Jessup, L.M,,22,32,45,720,736 Jewett, T.,525,532

AUTH~R INDEX Jin, Y., 618,621,626,636,637,641, 642,646,647,648 Joey, E G., 22,32 Johannsen, G., 311,339,678,709 Johansen, R.,7,29,46,68,91,448,449, 47 1 Johnson, K., 483,485,503 Johnson, P. M.,286,287,307,541,554 Johnson-L~rd,D.J., 675,676,678,679, 706,708 Johnson-Laird, P., 54,66 Johnson-Lenz, P., 473,497,504 Johnson-Lenz, T., 473,497,504 Jones, B., 495,502 Jones,P. M.,311,312,313,314,316, 317,322,323,329,330,334,337, 338,339 Jones, S., 732,736 Jonsson, P., 767,782 Jordan, J., 164, 192 Joseph, T, A., 403,408 Judd, T. H.,342,366 K

Kaelbling, L., 419,445 Kahan, J, P,, 675,678,682,698,702, 707,708 Kahn, K. M.,32,49,264,308,449,472, 541,557 Kahn, V. A.,404,407 Kahneman, D.,22,24,46,54,65 Kaiser, G. E., 281,308,55 1,554 Kaldor, N.,162, 192 Kambil,A., 28,44,518,519,531 Kandel, E. R.,257,260 Kaplan, R.E., 448,470 Kaplan, S., 286,297,307 Kaufer, D.S., 32,48,263,297,308,537, 542,544,548,554,556 . Kaufmann, J., 414,445 Kawauchi, Y., 412,445 Kaye, T.,542,543,556 Kazar, M.L., 400,407 Kehler, T., 130, 159 Keim, G., 537,544,556 Kelly, T. J., 593,594 Kenny, P., 412,446 Kephart, J. O., 34,50

e

Khatib, O., 420,443 Kidder, T., 25,46 Kiefer, N.M.,593,594 Kiesler, S., 1,4,22,46,719,720,737, 778,783 Kim, Y.J.,494,504 King, J.L., 22,47,263,307,511,513, 517,532 King, N.,488,504 King, R.,547,557 Kirste, K. K., 619, 621 Kistler, J. J., 400,408,55 l ,552,556 Kjaer-Hansen, M.,601,620,626,640, 647 Klein, G. A., 679,690,708 Klein, H.,569,583 Kleinman, D.L., 22,48,626,647,673, 674,675,676,678,681,682,685, 686,692,694,696,698,701,702, 704,707,708,709 Kling, R., 10,23,29,45,46,508,509, 510,512,513,514,515,516,517, 519,520,521,522,523,524,525, 526,527,528,531,532 Knight, J.C., 286,307 Knight, K., 654,672,746,760 Knights, D.,513,532 Knister, M.J., 264,307 Knoblock, C. A., 85,91 Kobayashi, K., 589,594 Koestler,A., 206,254,255,260 . Kohler, R.E., 722,736 Kohler, W., 40,46 Koopman, T. C., 250,260 Kopelrnan, E.,668,672 Korf, R.E., 85,91 Korner, J., 592,594 Kornfeld, W. A., 35,46,47,96, 122 Kowal, J.A., 320,338 Kraerner, K. L., 22,47,29,45,263,307, 511,513,517,532,560,584 Kramer, R.M.,619,620 Krasner, H.,501,502 Kraus, S,, 94,95,99, 101, 108, 109, 110, 113,114,115,117,121,122,123,124 Krauss, R.,54,65 Kraut, R.E., 2,4,32,45,539,548,555, 570,583,720,721,729,736 Kreps, D.,95, 123

Krishna, R.,512,533 Krogh, B.,420,446 Kuffner, T. A., 461,463,471 Kuhn, N.,104,122 Kullback, S., 589,593 Kumar, V.,365,366 Kunz, J.C.,618,621,625,626,641,

642,646,648,679,706,707,710

Kunz, W., 3 1,48,462,47 1,488,504 Kupfer, M. D.,400,407 Kurland, D.M.,548,555 Kurose, J. F.,34,47 Kuwabara, K., 72,91,96, 123 Kuwana, E.,264,268,308,569,570,583 Kyng, M.,328,337 L La Porte, T. R.,596,621 Ladd, B.,365,366 LaFasto, F.,312,338 Lai, K. Y.,,33,47, 127, 137, 138, 139, 147,153,154,160,328,338

Laird, J.E.,602,604,605,621 Landauer, T. K., 739,741 ,760 Lanning, S., 32,49,449,472,541,557 Lantz, K. A., 264,275,307 LaPotin, P.,596,621,626,648 Larson, C.,312,338 Larson, J.,551,555 Latour, B.,727,731,736 Laudon, K., 5 13,533 Lauwers, J. C.,275,307 Lave, J.,463,471,717,736 Law, J., 717,735 Lawrence, P., 10,42,47 Leavitt, H.J., 26,27,47 Lebow, R.N.,119,123 Lederberg, J., 717,718,721,736 Lee, A., 461,471 Lee, G., 156,160 Lee, J., 11, 12,23,31,32,39,47, 132, 138, 139, 146, 160,412,446,635, 648,767,774,782 Lee, L., 670,672 Lefkowitz, L. S., 32,44 Leggatt, J.J., 405,408 Lehmann, D.,115,117, 122 Lehner, P. E.,674,675,708

Leifer, L. J.,541,557 Leitch, R.,312,338 Leland, M. D.P.,32,45,548,555 Lemke, A. C.,264,307,455,466,467, 470,47 1

Leng, R.,119,120,123 Lengel, R. H.,501,502,562,578,582 Leplat, J., 316,339 Lesser, V. R.,41,45,72,79,85,91,96, 104,115,116,121, 122,123

Lester, R. K., 512,532 Levesque, H.J., 23,44 Levine, J.A., 548,555 Levis, A. H.,706,710 Levitt, R.E.,618,621,636,637,642, 646,647,648,679,706,707,710

Lewis, C.,547,557,669,672 Liker, J.K., 495,504 Lin, F., 27,44 Lin, J. K., 392,408 Lin, Y., 679,706,707,710 Lin, Z.,596,597,598,602,603,617, 619,620,621

Lind, M.,484,506 Lindblom, C.,66 1,67 1,672 Lindsay, P. H.,696,708 Liskov, B.,18,47 Lister, T., 448,470 Lochbaum, K. E.,115,117,123 Long, M. D.,467,472 Lorsch, J., 10,42,47 Lovasz, L., 164, 192 Lucas, R. E.,Jr., 198,250,260 Lucco, S. E.,297,306 Luce, R.D.,652,672 Luchetti, R.,572,584 Luenberger, D.G., 690.709 Luff, P.,562,583 Luh, D.L., 674,676,682,685,686,692, 694,696,698,701,702,704,709

Luh, P. B.,22,48,673,675,676,678,

68 1,686,696,698,702,707,708,709

Lumer, E.,36,47 Luotonen,A., 740,759 Lynch, K. J.,740,746,759

M MacCrimmon, K. R.,698,709

AUTHOR INDEX Mack, L. A., 570,572,584 MacKay, W. E.,132,133, 148, 160,578, 584

Mackenzi, K. D.,596,621 MacKenzie, D.,416,417,421,444,446 MacLean, A., 488,504,564,583 MacLennan, B., 413,446 Maebara, K.,284,309 Maeno, K., 284,309 Maes, P,,71,91, 126,135, 160,412, 419,445,446

Magnusson, B., 543,556 Mahoney, M. J., 493,504 Majchrzak, A., 618,620 Major, D.A., 599,620 Malone, T. W., 1,4, 11,12, 18, 19,21,23,

24,26,27,28,30,32,33,39,42,44,47, 48,51,55,69,71,91, 115, 116, 123,127, 132,135,136,137, 138,139, 146,147, 154,160,268,307,319,328,338,370, 389,474,504,508,509,518,519,531, 533,536,538,556,595,619,621,635, 648,65 1,668,672,673,674,68 1,709 Manber, U.,740,759 Mancini, G., 311,337 Mann, L., 22,46 March, J.B., 596,620 March, J,G., 21,42,48,626,647,648 Marcosson, I. F,, 664,665,666,672 Marcus, G. E.,716,735 Markus, M. L,, 33,48 Maron, M. E.,739,759 Marschak, J.,41,48,673,709 Marschak, T., 54,65, 161,164, 178, 190, 192 Marshall, C. C., 365,366,459,471,501, 504 Martin, J., 286,308 Martinez, J.,748,759 M a s h a y e ~ iV,., 284,287,288,289,294, 297,305,308 Mason, R.O., 485,504 Masuch, M,, 596,621,626,648 Mataric, M., 412,431,445,446 Mathur, A., 774,782 Mazola, J., 16,48 McArthur, D., 104,121 McCall, R.,264,307,453,455,457, 466,467,470,471,472

7 McCandless, T., 461,470 McCarl, R.,602,604,621 McClain, J., 16,48 McClelland, J.L.,7,49 McClure, C., 728,736 McCracken, D.L., 405,407 McDaniel, S. E.,767,778,782 McDowell, C., 43 1,445 McFarlan, F. W., 29,44 McGee, J., 767,778,782 McGrath, J.E.,24,48,316,338,476,

483,485,487,490,504,554,556, 559,560,562,564,583,673,674, 709 McGuffin, L. S., 264,268,308,570, 572,581,583,584 McGuire, T. W., 22,46,719,720,737 McIntyre, S. C., 501,502 McKinney, M,, 657,672 McLeod, P. L., 495,504 McMahon, M., 467,472 Mead, G. H.,54,65 Meader, D.K., 562,578,579,581,583, 584 Melnyk, S. A., 511,533 Memmi, G., 286,307 Menees, S. G., 400,404,407 Menges, J., 392,408 Mennecke, B. E.,495,504 Merkel, C., 714,719,735 Meyer, A., 96, 115, 116, 121 Meyer, J.W., 514,525,533 Meyer, R.A., 72,91 Meyrowitz, N. K., 404,407 Miao, X.Y.,22,48,673,698,709 Middleton, D.S., 732,736 Milgrom, P., 23,48,637,648 Milios, E.,41 2,45 Miller, D., 412,446 Miller, J.G., 43,48 Miller, J.R., 148, 160 Miller, K.L., 619,621 Miller, M. S., 11,34,35,45,48 Miller, M., 286, 306 Milliken, F. J., 53,66 Minor, S., 543,556 Minsky, M., 257,260 Mintzberg, H.,18,24,27, 30,42,48, 596,62 1,624,634,648

7

AUTHOR INDEX

Mitchell, C. M., 311,3 17,323,330,337, 338,339

Mitchell, W. M., 602,621 Mitroff, I. I., 485, 504 Moch, M. K.,53,64 Moffett, R. J.,674,675,708 Mogensen, P.,328,337 Mokyr, J., 255,256,257,258,259,260 Monge, P.R.,619,621,643,648 Montgomery, T.A., 79,86,87,90,91 Montini, T., 717, 735 Moon, D. A.,467,472 Moorman, K.,420,446 Moran, T. R,488,504,564,583 Moray, N., 678,679,709 Morch, A.,466,467,470 More, G., 412,445 Morgan, B. B., 675,709 Morgenstern, O., 652,654,672 Morris, C. G., 674,708 Morris, J.H., 32,48,263,297,308,537, 548,554,556

Morris, N. M., 676,679,709 Morris, P.,15,45 Morris, R. A., 342,366 Moser, S., 119,121,124 Moses, J., 18,48 Moses, X, 115, 116, 123,373,389 Mosteller, F., 602,620 Mount, K.,164,192 Moynihan, M.,414,424,446 Mundy, K.B., 342,366 Munson, J., 271,279,305,308 Murata, T., 344, 366 Murphy, R. R., 419,445 Murray, F., 513,532 Murugappan, P.,32,44 Myers, E. A.,286,307 Myerson, R.B.,41,48 N Nair, A., 32,44 Nakagawa, S., 412,445 Nakakoji, K.,453,464,467,468,470, 471

-

Nalebuff, B.,121 Namioka, A., 16,49 N a r a s i ~ a nR., , 511,533

Nardi, B. A., 148,160,452,471 Narendra, K.,382,389 Nasrallah, W., 640,648 Nass, C. I., 618,621,626,641,642,648 Nass, T.,679,706,707,710 Nelson, M.N., 400,408 Nemiroff, P.M., 486,504 Neumann, L., 714,719,735 Neu~irth,C. M.,32,48,263,297,308,

536,537,539,543,544,548,554, 556 Newell, A.,250,260,601,602,604,605, 620,621,626,640,647 Newman, J.,542,556 Newman, R.,542,556 Ng, D. T.,746,748,749,759 Nichols, D. A.,400,404,407 Nielsen, H. F.,740,759 Nii, H.P.,30,41,48,79,91,323,338 Nikhil, R.S., 16,44 Nilsson, N. J., 18,45 Nishimara, A.,253,260 Nitz, E., 421,425,444 Nogami, G., 675,709 Noreils, F. R.,413,446 Norman, D. A.,51,65,137,160,548, 556,562,583,696,708 Norman, R.,263,307 Nunamaker, J.F.,22,32,45,294,306, 316,337,473,492,503,504,747,759 Nylan, L,, 365,366 Nystrorn, P.C., 53,65

0 O’Conaill, B., 562,583 O’Day, V. L., 467,469 0 7 ~ ~ 1 T. 1 , J., 257,260 O’Hare, G. M. P.,120, 123 Obel, B., 13, 19,44,634,647 Ocker, R.,485,487,498,503,504 Olsder, G. J., 652,671 Olsen, J.P.,596,620,626,647 Olson, G. M., 3,4,264,268,294,307, 308,487,488,489,501,505,536, 556,560,562,564,565,566,567, 569,570,581,583,584,722,737, 761,767,775,778,782

AUTHOR INDEX Olson, J.R., 264,268,308,487,488, 489,501,505,536,539,556,560, 562,564,565,566,569,570,572, 574,575,583,584,775,782 Olson, M.H.,2,4,612,621 Oralkan, G. A.,637,638,643,648 Orlikowski,W. J.,33,48,513,533,729, 738 Orwig, R., 747,759 Osborne, A.F.,474,505 Osborne, M.J.,94,123 Oscar, L.,716,736 Ostwald, J., 264,307,453,455,464, 466,467,468,470,471 Ouchi, W. G., 619,621 Ousterhout, J. K.,400,404,407,408 Overgaard,G., 767,782 P Paepcke, A.,467,469 Pagani, D. S., 578,584 Papageorgiou, C. R,601,62 l Papamarcou, A., 589,594 Park, D.,596,601,620 Parnas, D. L.,285,308 Partridge, J.L.,495,503 Pasmore, W. A., 52,65,486,504 Pasquale, J., 378,380,383,384,386, 388,389 Pass, S., 512,533 Pasteels, J., 414,445 Patterson, E.S., 314,315,338 Pattipati, K.R., 673,675,676,678,681, 686,696,702,707,708,709 Pavicic, M.,413,444 Pavlin, J., 96, 123 Payton, D. W., 419,446 Pearce, M.,420,446 Pearson, M.,419,445 Pelletier, S. D.,342,366 Pentland, B., 11, 12,39,47, 157, 160, 635,648 Pepper, S. C., 53,65 Perkins, D. N.,254,256,260 Perrow, C., 513,533 Pete, A.,673,708,709 Peters, D.P., 493,505 Peters, T.J., 559,584,635,648

Peterson, D., 29,48 Peterson, J.L., 15,31,48,344,366 Pew, R. W., 675,676,678,709 Pfeffer, J., 10, 13,48 Pfeffer, J.,513,516,533,599,620 Pingali, K.K.,16,44 Pinsonneault, A.,560,584 Plato, I., 253,260 Pnueli, A., 3 12,337 Polanyl, M,,251,253,260,453,461,471 Polucci, T.A.,342,366 Pool, R., 755,760 Poole, M.S., 476,478,493,495,503, 505,564,584,641,648 Porter, L,W., 614,620 Posner, I. R., 543,556 Post, B. Q., 264,308 Powell, W. W.?513,533 Prakash, A., 264,307,767,773,774, 782,783 Prietula, M.J., 595,601,603,620,626, 640,647,679,710 Prins, J.,365,366 Pu, C., 281,308 Purser, R. E.,52,65 Putnam, L.L.,495,505,564,584,614, 620 R Rader, K.,722,737 Radner, R., 41,48, 163, 164,192,673, 709 Raiffa, H.,652,672 Ram, A.,420,445,446 Rana, A., 490,493, 505 Rao, R., 127,132, 137, 160 Rao, U.,495,500,506 Rasmussen, C., 774,782 Rasmussen, E.,744,760 Rasmussen, J., 316,339,678,679,690, 709 Rechner, P. L., 485,486,505 Reddy, D. R.,30,41,45 Reder, L.M.,461,463,471 ‘Redfield,R., 716,737 Reeves, B. N.,264,307,312,339,455, 457,462,466,467,470,471 Reichelstein, S., 164, 192

AUTHOR INDEX

79 Rein, G. L.,17,29,32,45,49,264,268, 287,306,316,337,449,470,473,503 Reisig, W., 344,366 Reisner, M., 657,660,672 Reiss, S. P.,264,308 Reiter, S., 41,48, 164, 192,256,260, Reitsma, R. F.,669,672 Repenning,A,,466,471 Rescher, N.,3 l , 48 Rhyne, J., 462,472 Rice, R. E.,474,505,7 19,737 Rich, E.,654,672 Richartz, L.,19,44 Riedl, J.,284,287,288,289,294,297, 305,306,308 Riley, J.H.,404,407 Rip, A.,717,735 Rittel, H.W. J., 3 1,48,462,471,488,504 Roberts, J., 637,648 Roberts, K.H.,10,22,49,596,614,620, 621,643,649 Robertson, G. G., 297,309 Robinson, L.,495,503 Robinson, M,, 328,339,669,672 Rockart, J.F.,10,26,28,47,49 Rodden, T.,73 l, 737 Rogers, E.M., 719,737 Rogers, E.,22,49 Rommetveit, R., 54,65 Rosechein, S., 419,445 Roseman, M., 264,308 Rosenblitt, D.,127, 137, 160 Rosenbloom, P.S., 602,604,605,621 Rosenschein, J.S., 79,89,91, 104, 115, 118, 119,122,123,124 Ross, L.,54,65 Ross, S., 41,49 Roth, J., 564,584 Roth, T., 343,366 Rothman, L.W., 619,621 Rouse, W.B., 676,679,709 Roven, B.,512,533 Rowan, B.,5 14,533 Rubin, K.S., 317,323,337,338,339 Rubin, L.,493,502 Rubinstein,A., 94, 111, 121, 123 Rueter, H.H.,564,566,567,583 Ruhleder, K.,723,724,728,737 Ruiz, J. C.,362, 367

Rule, J., 29,44 Rumelhart, D.E.,7,49 Rutter, D.R.,578,584 S

Sager, N.,742,760 Saisi, D.L.,317,338 Sakata, S., 284,309 Salancik, G. R., 13,48 Salas, E.,674,675,676,678,679,681, 682,702,706,707,708,709 Saloner, G., 23,45 Salton, G., 740,742,760 Sambamu~y,V.,492,505 Sandberg, W. R., 485,486,505 Sandell, N.R., Jr., 589, 594 Sandholm, T., 104,123 Sandusky, R., 714,719,735 Santmire, T. E., 120, 123 Sarin, S., 264,268,308 Satyanarayanan,M,,400,404,407,408, 55 1,552,556 Savage, C. M,,626,648 Scarbro, H.D.,342,366 Schaab, B.,453,471 Schaaf, R.W., 297,309 Schatz, B. R., 405,408,731,722,737, 740,742,748,751,755,759,760 Schelling, T.C.,7,25,33,49 Scheurich, C.,17,45 Schliemann, A.D.,463,469 Schmidt, K.,315,328,334,337,339 Schnase, J.L.,405,408 Schneider, A.K.,668,672 Schneider, G. M.,286,308 Schnepf, J., 284,288,289,305,308 Schoen, D.A.,461,471 Schon, D.,53,64 Schonberger, R., 16,49 Schriver, K.A.,541,545,555 Schuler, D.,16,49 Schuler, W.,453,471 Schutt, H.,405,408 Schwartz, F.,740,759 Schwartz, M. D.,405,407,549,555 Schwartz, R., 95,120,123 Schweiger, D.M.,485,486,505 Schwenk, C.R.,485,505

AUTHOR INDEX Scott, W. R., 513,525,526,533 Searle, J.R., 30,49 Secret, A.,740,759 Seeley, T,D.,43,49 Sego, D.J.,599,620 Selfridge, P,G., 467,472 Seliger, R., 279,307 Selten, R., 94, 123 Sembug~oorthy,V.,286,306 Sen, S., 71,73,74,77,78,79,87,90,91, 92 Serfaty, D.,626,647,674,675,676,678, 681,682,685,686,692,694,696, 698,701,702,704,707,708,709 Sewell, W. H,, 643,648 Shackelford, D.E., 398,408 Shalaby, H, M.H.,589,594 Shan, Y. P.,32,44 Shani, A. B.,512,533 Shannon, C.E., 22,49 Shaw, M. E., 674,698,707,709 Shen, H.,268,270,272,306,271,279, 289,305,308 Sheridan, J.W., 512,533 Sheridan, T.B., 311,339 Sherman, G. R., 256,260 Shi, E., 517,526,532 Shi, P.,673,707,709 Shim, H. S., 773,783 Shipman, F. M., 312,339,365,366,453, 455,457,459,462,464,466,467, 470,471,472,501,504 Shirriff, K.W., 400,407 Shoham, Y., 23,49,79,92, 115, 116, 123 Short, J.A.,578,584 Short, J.E., 10,49 Shortliffe, E. H., 452,469 Sidebotham, R. N.,400,404,407 Sidner, C.L,, 115, 117, 122, 123 Siegel, J., 22,46,719,720,737 Siegelmann, H. T.,106,122 Sijstermans, F., 106, 121 Silver, M.S.; 493,505 Simha, R., 34,47 Simmel, G., 713,716,737 Simon, H. A., 13,21,22,26,42,48,49, 56,66,250,260,488,505,619,620, 626,648,675,698,709 Singh, B., 17,39,49

Singh, M.,120, 121, 122 Sistla, S., 361,366 Slack, M. G., 419,446 Sloane, A. M., 669,672 Slovic, P., 54,65 Smith, A., 13,49 Smith, C.S., 255,260 Smith, D.K.,404,408 Smith, F. D.,312,328,339,392,408 Smith, J.B., 312,328,339,392,408, 541,556 Smith, R. G., 11,21,34,45,49,72,79, 92,104,115,116, 124 Smith, S. A., 18, 19,21,24,48 Smith, T.,287,306 Smoke, R., 119,122 Smolensky, P., 547,557 Snyder, G. H.,119,124 Solomon, M.,268,264,306 Solow, R. M.,512,532 Sproull, L., 1,4,719,737,778,783 Stacey, M.,716,737 Stahl, G., 453,464,468,470 Stankovic, J., 34,49 Star, S. L., 63,66,714,715,723,726, 731,733,734,735,737 Starbuck, W. H., 52,53,65,66 Starkey, B., 121, l24 Stassen, H. G., 678,679,709 Stasz, C.,675,678,682,698,702,707, 708 Steeb, R., 104, 121 54,66 Steedman, M,, Stefik, M.J., 32,49,264,308,449,472, 541,557 Stein, J.G., 119, 124 Stein, L., 413,446 Steiner, I. D.,562,584,674,681,709 Stelzner, M.C.,625,648 Stilgoe, J.R., 1 , 4 Stoddard, D.,29,49 Stogdill, R. M., 496,505 Stone, H., 106, 124 Stornetta, S., 34,50,449,471 Storrosten, M. N.,264,308,487,488, 505,536,539,556,564,565,566, 567,570,577,579,584,775,782 Stotts, P.D.,341,342,343,351,359, 361,362,365,366,367,405,408

Stout, S. K.,22,49 Stratman, J., 545,555 Strauss, A.,717,737 Streitz, N.A., 405,408 Streufert, S., 675,676,709 Suchman, L.,32,49, 135, 136, 160,328, 339,449,472,541,557,731,733,737

Sugihara, K.,640,648 Sumner, T.,453,464,466,468,470,471 Swigart, R.,316,337 Sycara, K.P.,115,117,124 T Tambe, M.,79,92 Tan, T., 95, 124 Tang, J.C.,541,557 Tang, J.T., 578,580,584 Tank, D.W., 746,760 Tannenbaum, A. S., 16,49 Tannenbaum, S. I., 674,675,681,709 Tansik, D.A.,720,736 Tanter, R.,119,124 Tatar, D.,29,49 Tate, A., 18,43 Te’eni, D.,51,57,65 Teasley, S. D.,562,584 Telem, M.,505,505 Tenkasi, R.V,, 51,52,57,65 Tennenholtz, M.,79,92, 115, 116, 123 Tenney, R. 589,594 R., Terveen,L.G., 467,472 Teversky, A.,54,65 Thathachar, M.,382,389 Thimbleby, S., 463,472 Thomas, J.A.,589,594 Thomas, L.J.,16,48 Thomas, R.H.,297,309 Thompson,C, W., 391,408 Thompson, E.,512,533 Thompson, J.D.,10,42,49,596,621, 624,649

Thompson, L.,119, 124 Thomsen, J.,642,648,649 Tjahjono, D., 286,287,307 Tjosvold, D.,485,505 Toffler, A.,9, 27,30,49,559,584 Tolone, W., 286,297,307 Tombaugh, J.W., 721,737

Tomlinson, R. S., 297,309 Tonnies, F., 716, 738 Toulmin, S., 31,49 Travers,V.M.,297,309 Traweek, S., 720,738 Tsai, W., 284,286,287,294,297,308 Tsitsiklis, J.N.,589,594 Tung, L. 486,487,505 L., Turbak, F. A., 1,4,27,30,32,47,69,71, 91, 127,147,160

Turner, W.,734,735 Turoff, M.,14,32,49,474,476,483, 484,478,485,493,495,496,499, 500,501,502,503,504,505,506, 721,729,736 Tversky, A., 22,24,46

U Ullman, D.G., 461,463,469,471 Uncapher, K.,717,718,721,736 Ury, W. L.,652,655,656,657,672

v Vakil, T. F., 492,502 Valacich, J.S., 294,306,473,504,720, 736 van Biljon, W. R.,343,367 Van de Ven, A.H.,474,506 Van Zandt, T., 163, 164, 192 Vasilik, E.,305,306 Vaughn, D.,419,445 Vazirani, U., 161, 178, 190, 192 Verfurth, S. C., 317,339 Viswanathan, H.,588,589,592,593 Vogel, D. R.,22,32,45,473,492,503, 504 Von ~eumann,J.,652,654,672

W Wagner, I., 509,533 Wainer, J., 669,671 Waldspur~er,G. A., 34,50 Walker, J. H.,467,472 Wallace, A. R,,254,26 Wan, D.,286,307 Wang, J.,413,446

AUTHOR INDEX Wang, W. P., 674,676,682,685,686, 692,694,696,698,701,702,704,709 Warner, T.N., 510,533 Watabe, K., 284,309 Waterman, R. H., Jr., 559,584 Watson, R. T., 483,484,498,506 Weaver, W., 22,49 Weedman, J .,719,723,738 Weick, K. E., 22,50,55,66,562,578, 584,643,649 Weihl, W., 279,307 Weinberger, H., 164, 192 Weinreb, D. L., 467,472 Weiss, D. M,,285,308 Weiss, G., 120, 124 Weiss, S. F., 541,556 Welch, B. B., 400,404,408 Wellman, M., 104,115, 124 Wellner, P., 570,584 Wenger, E., 717,733,736 Werlang, S., 95, 124 Werner, G., 413,446 West, M. A.,400,404,407,487,488,506 Whang, S., 28,46 Wharton, C., 467,469 Wheelwright, S. C., 510,512,532 Whisler, T. L., 26,27,47 Whittaker, S., 562,583 Wickens, C. D., 562,584 Wiener, N., 731,738 Wilbur, S., 562,583 Wilkenfeld, J. M., 94,95,99, 101, 108, 109, 110, 113, 115, 119, 120, 121, 122,123,124 Wilkes, D., 4 12,445 Williams, E., 578, 584 Williams, M. D., 625,648 Williamson, 0. E., 10, 14,21,25,28,41, 50,595,621 Wilson, B., 271,307,549,555 Wilson, E., 414,445,669,672 Wilson, R., 95, 123 Wilson, V.,669,672 Winner, L., 726,738 Winograd, T.,7,30,32,45,50, 147, 160, 497,503,731,738

Witten, LH.,463,472 Wohl, J.G., 678,689,706,709 Wolf, C., 462,472 Wolinsky,A.,111, 121 Woodard, E. A.,675,709 Woods, D. D., 311,337 Woods, W. A.,742,760 Woodson, W. E., 265,309 Wooler, S., 674,710 Worley, D. R., 675,678,682,698,702, 707,708 Worrell, M,, 485,499,503,506 Wroblewski, D., 461,470 Wu, G., 767,774,782 Wulf, W. A.,761,783 Wynne, B. E., 475,495,502,504

Y Yakemovic, K. C., 462,472 Yanco, H., 413,446 Yang, Q., 87,90 Yankelovich, N., 404,408,546,548,554 Yates, J., 1,4,28,48,51,65,729,738 Ye, M.,602,621 742,748,759 Yim, T,, Yin, R., 293,309 Yoder, E. A.,405,407 Yost, G., 602,605,621 Young, F., 484,506 Young, R. M., 488,504,564,583 Yourdon, E., 282,309 Yu, K. C., 33,47, 138,139,154,160 Z Zakay, D., 674,710 Zhang, Z., 588,589,592,593,594 Zigurs, I., 669,672 Zirk, D. A., 674,675,708 Zlotkin, G., 94,95,99, 101, 104, 118, 119,113,115,118,121,123,124 Zmud, R. W., 484,506,5 10,531 Zysman, J., 5 10,531

This Page Intentionally Left Blank

A activity model, 317-319 ACTS theory, 595 adaptive structuration theory, 476478,483,498,643 adaptive systems, 73-78 adhocracies, 9,27,30,559 Advanced Integrated Manufacturing E n ~ ~ (Awrr), e n 508-530 ~ agency theory41,517,526 agents, 68-89,137, 142-144,165, 311,464465,535-536,588-593 intelligent, 125-126,139,595-597 mobile, 79 alternative offers model,94 analogies, 9/11, 34-35 annotations, !j48-550, 772-773 Artifa~-BasedCollaborative System (ABC), 328-329,392 artifacts, 320,460-461 artificial intelligence, 625 Aspects, 562 Automatic Post Office(MO), 487, 490 awareness, 271-272,279 B Bio~uest,751-755 blackboards, 30 boundary object, 63 browsing, 342,343,740 business process =engineering, 624, 626,767 C cash register, 664-667 cause map,55/62

cen~~~ation-decen~alization, 21, 28-29,51,369,372-374, 596,599600,633 CEO or Top Manager 162-168,588590,592-593,599 client-server, 342, 346-347, 371,400 Cobweb project, 365 CODE framework,677-682,686-695 coding of transcripts, 775-780 cognitive capabilities, 600-601,696697 Colab, 449 Collaboration Technology Suite (CTS), 572-573 collaborative editing, 269-271 collaboratories, 3,392,715,717-719, 724-725, 727, 729, 732-734, 761-762 Rockefeller workshop,3,718 Collaboratory Builders~ ~ ~ (CBE), 767,769 Colorado River basin,652,657-6 common knowledge, 23 communication effort, 163-165 communication overload, 698-700 c o ~ ~ c a t i otypes n , of,424426, 628-629 communities of practice, 52,717 community, 35-36,713-724, 732-734 community electronic, 714 computer c o ~ e r e n 27,133,729 ~n~ computer science, 14,40-41 publications 1958-1990,198, 229 Computer-~tegratedM a n u f a ~ g (CM),511-513,519 concurrency control,268,272,279, 392-393,400 conflict, 68 Constructive Consensus Approa 486

e

SUBJECT INDEX

consume, 416,422-423,429-434,436437,439-441 contingen~theory, 474,516,526, 599-601,624-625,634 4 ,72,104,116 contract nets,3 coordination, 7,9,10-23,39,317, 319,320-321,343,507-530,588, 595-597,617-619,627,673-708 implicit vs. explicit, 674-682, 702-708 costs, 26,38,162-163,373,384, 517-518,560,633,674 processes, 11-13, 36-38,634-635 science, 2,535,538 theory, 7, 70, 80-81,90,411,508, 635,651,653,655,667-671 randomized, 369-389 Coordination Theory & Collaboration T e ~ o l (CTCT), o ~ 2/68! 93 Coord~ator,32,731,152-153 COSMO, 323-335 creativity discovery, 195196,204-209 crisis decision making, 108-115, 119-120

differencing, 268 digital library, 715,719,727,739741,755-758, 762 Dining Philosophers, 354-358 dispute systems, 652, 656 distributed cognition, 51-64,535, 562,732 DistView, 773 domain knowledge, 321-322 E Earth Science Data andI n f o ~ a t i o n System,106-108 economics, 13, 193-195,200,203 EIES, 478-479,483,490-493,729 electronic b r a i n s t o ~ i n747 ~ electronic mail, 1,27,69, 128-133, 450-451,721 negative effects, 1, 41-42 electronic preprints, 726,730 ELM agents, 601-604 empirical evaluationof systems, 132-133,293,300-302, 329-335, 552-553,571-581,774-782 enterprise modeling, 624-627,634636, 6 3 8 - ~ 6 evaluation,24-25 exception ~ a n d l i n ~ expert systems, 452 expertise, 205-206, 4

firm size, 28

SUBJECT INDEX

goals, 17-18 conflict, 25,651-671 a grammatical analysis,567 graph algorithms,740 graze, 416, 423424, 429434,437 group calendars,509-510 group decision making,22-23,289, 296

iterative design,411 *

J journal publication,721,725 just in time inventories,9, 15-16, 514

*

group decision support system (GDSS), 473475,501-502 group^^, 729 groupware, 8/29! 266-280,638 architectures, 274276 GROVE, 449 H heuristics, 54, 72, 108 hierarchies, 21,36,51,73,76,82,87,

139-140,317,344,518,596, 613614,618-619,635,674 HITEC experiment,682-686, 692-695 Hopfie~dne& 746 Hypercard, 343 hy~ermedia,391-407 hypertext, 341,343,344,364-365, 452-453, 501-502 l

incentives, 33 incomplete i~ormation,94 indexing, auto~atic,742,755 Indus~ial Revol~tion, 714,716 i~ormationcondensa Information Lens,30, i ~ o r m a t i o n o ~ e r l1,600,739, oa~,

K knowledge acquisition,452-453 knowledge, model 194-199,201-204,209-227 tacit, 453,461 L leadership, 496 learning automata,382-388 learning by doing, 198-199 level of analysis,617 LiveBoard, 562 load balancing,384-3~8 Lotus Notes,33,154

M m a n u f a ~ r i n g1,5 Man~factu~n Reso g

741

i ~ r a s t ~ c ~715,725-728,730-734 re,

mental model,679,690, 696 mentorin 781

SUBJECT INDEX N narrative, 57/63 Nash equilibrium, 94-95 National ~ o ~ a t i Infrastructure o n (NII), 715, 740 NCSA Mosaic,727,751 negotiation, 93-121,655-664 nested models, 89 network architecture, 726-728 NeXTStep, 762,766,768-769,774 Nominal Group Technique, 474, 478

0 object-oriented analysis& design, 767 open systems, 2, 51 operating systems, 14 organizations, 8,21,26,51 theory 10, 13-14/42, 167-190 types, 617 learning, 616-617, 642-643 simulation, 595-619,623-647 organizational wind tunnel,626,647 Oval (also Object Lens), 138-157, 158-159,328 P parallel processing, 17,31,33-34/69, 84,86,106 participatory design, 16 perspective taking, 63 Petri nets, 31,341-342,344 planning, 86 plural incrementalism, 661 population ecology, 595 PREP editor, 537-553 prepared mind, 197, 207-208 privacy 72, 77-78,393,728 process losses, 473 production function, 194, 199, 227244 project management, 156-157 public good, 195,643

R radar task, 597-599 radical tailorability, 125,136-138, 145 reactive control, 419-420 relevance, 740 replay, 779-781 resource allocation,35,37,68,96104,164-165,377-381 resource dependency theory, 516517 robots, mobile, 82-87,411-443 role differentiation, 719-721 rules, 130-131,142 S

SA", 478 SAMPEX,312-316 scaling, 158,393,400,755,769 search biases, 75-76 second order effects,1/26-27 semiformal systems, 125,134-136, 343 Sequoia 2000,723 session management, 267 SHOR framework, 678-679 ShrEdit, 561, 569-570,579,581 situated action, cognition,463,733 Soar, 601-602, 604-609,640 software engineering, 263-265,343, 500501 inspection,280-303 m~agement,157 system design, 564-569 Sondrestrom Upper Atmospheric Research Facility, 762, 764-765 speech act theory,30,731 Spider, 55 standards, 16/42 s ~ c ~ r a l i s 595 m, subsumption architecture, 419 Suite,277-278

SUBJElCT INDEX

v

T task a s s i ~ e ~ t14,19-21,104-105 s, task circumplex,476,479,483,485 taxonomy tasks, 562,597,628,644 tools, 31-33 work, 315-316 TCB /IP, 726-727 technology change, model,193-249 third order effects,26-27 ~ a ~ a costs, ~ o13,41,517,526 n Trellis, 341-365,405 a , 347-358 X, 359-361 mst, 71-78 U

uncertain^ absorption, 614 undo, 268,272 Upper Atmosp~ericResearch Collaboratory (UARC),761-782 urbanization, 714 usability, 16’23 use cases, 767 user-centered design,560

views, 140-142,146 Virtual Classroom,499-500 Virtual Design Team (VDT),623632,636-646 visual programming,343 vocabulary problem,739-741,755 voice mail,449-450 voting tools,484-485 W workflow, 157,669-671 World Wide Web,365,719,726-727, 740,763,781 Worm C o ~ u nSystem, i ~ 722-724, 74-751 writing processes drafting, 542-544 planning, 538-541 reviewing, 544-547 writing, collaborative,535-554 WYSIWIS, 272,288 X xxx

see electronic

preprint