From: AAAI Technical Report WS-93-07. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved.
Using an organisation model to predict
effects on design team productivity
J. C. Kunz, G. P. Cohen and R. E. Levitt
Dcpartmcntof Civil Engineering
Stanford University
Stanford CA 94305
USA
In engineering,winninga bid for a designproject is a first majorsuccess, andit leads to a first
majorproblem.Giventhe contract, an organization must be built to performthe design. This
research focuses not on the design of artifacts, but on the design of organizations that will
design (or build) an artifact. Wereport our approachto the organization design problem
building symbolic models of design organizations m and discuss our initial results in
simulating organization performance.
1 Introduction
Designis often consideredto include a series of phases, including Requirementsdefinition;
Synthesis; Analysis; Evaluation;and Acceptanceor varyingold designs based on the evaluation
of performance (Levitt 91). Synthesis is inherently difficult. However,some areas
engineering design, such as design of manyphysical systems, nowhave a tradition of
analytical modelsthat are useful for analysis. For these areas, design involves repeated
creation, analysis and modification of analytical models,and the design process has become
quite formalized.Actualartifacts are nowbuilt and bench-testedonly after extensive modeling
and computeranalysis. Becauseof the success of recasting the inherently difficult synthesis
problemas an analysis problem, design has becomehighly formalized in manyengineering
domains, and tremendous progress has been madeboth in understanding the behaviors of
interest in each domainand in generating interesting newsyntheses. Analytical modelinghas
greatly extending the range of artifacts that can be designed and manufacturedsafely and
economically.
In contrast, the use of modelsto analyzesocial systemsis very limited. In practice, whena bid
is won, a project managerusually designs a neworganization based on past experience and
discussions with a few interested principals. Whilethe project managercan expect to use
modelsto analyzethe artifact prior to building it, the managerdoes not nowhavethe luxury of
using modelsto evaluate alternative organizational designs. For example,(Tatum84) reports
that managersof large design and construction projectsmlike most managersdesigning large
scale organizations to carry out complextasks~stiil rely on adaptation of past organization
structures, rather than on systematicgenerationandevaluationof alternatives, in designingtheir
organizationstructures.
Theoreticalworkin the field of OrganizationalBehavioris useful, but it has beendescriptive,
e.g., (Galbraith 77), (Simon76), (Thompson
67). Mostorganizational behaviorsof interest
scientists or managerscan only be representedas discrete, nominalor ordinal variables, leading
to a mismatchbetween these theories and the continuous, quantitative models suited to
83
traditional simulation techniques. Computationalmodelsof organizational behavior are only
nowbeginning to emerge(Masuch89), (Carley 90), (CohenG
Organizationsnormallycreate subteamsto distribute the specialized workof solving complex
problems.Theythen organize the specialized subteamsin somesort of hierarchy to centralize
overall project responsibility. In addition to information that flows up and downreporting
hierarchies, project teams also allow lateral communicationamongpeers to help groups to
obtain missing information with minimumcommunication.The degree of formalization of
communication
within a hierarchy is both a practical and a theoretical issue in organizational
design.
Wehave synthesized a description of the theory of organizations and implementedthis
synthesis in a computational symbolic model. The theoretical frameworkincludes actors,
projects, subteams,and communication
technologies (e.g., telephone, voice mail). Teamsare
related by formal organizational reporting hierarchies and informal information-sharing
networks. Communication
betweensubteamscan follow both vertical hierarchical links and
horizontal information-sharing links. This frameworkforms the basis for our Virtual Design
Team(VDT)model;the modelexplicitly represents organizational entities, and it reasons about
the (stochastic) behavior of processing nodes, communicationsnodes and channels, and
exception processing. Given a description of a project, and an organization and the
communication
tools it uses, the simulation modelcomputesproject duration. Wedescribe the
generic modeland the specific examplemodelsthat we have built and tested for realistic
industrial design projects. Whilewe do not claim to have calibrated it so that the absolute
predicted project durations will be accurate, we have found that changesin organizational
structure or in communicationtools can cause change in simulated project duration. The
changesin simulation results are consistent with predictions of theory and of experienced
project managers.
This paper reports the initial developmentand testing of an AI symbolicmodelof someaspects
of the behaviorof an engineeringdesign organization. Themodelexplicitly represents concepts
fromorganizationtheory, the social science discipline that describesthe waysthat differentiated
groups communicateand function in integrated waysin a business organization. Wediscuss
aspects of that theory and howwe represented it in the model. The VDTmodelis a formal
symbolic computational discrete event simulation modelbased on organization theory. It
includes a high level of detail about tasks, actors and communications
tools, describes static
relationships amongthese entities, and uses simulationto predict their dynamicbehaviorand to
predict project performance.
2 ReDresqntation and Reasonina
This section introducesthe top-level entities that the VDT
modelrepresents. Figure 1 showsan
overviewof the VDTentities and someof their relationships. As discussed below, the objectoriented VDT
modelrepresentsmostentity attributes qualitatively.
84
TASK NETWORKOF ACTIVITIES
ORGANIZATIONAL
HIERARCHY
finish
"IN
TRAY"
"OUT TRAY"
ACTOR
COMMUNICATIONS
FROM OTHER ACTORS
COMMUNICATION
TOOLS
COMMUNICATIONS
TO OTHER ACTORS
Figure 1. VDTmodels the design task, actors, organization structure, and communication
tools. The organization hierarchy consists of actors that are modeled as information
processors. Actors have "attention rules" for selecting communicationswaiting for the actor’s
attention in an "in tray," and selection rules for choosing which communicationtool to employ
for sending communicationsto other actors via an "out tray." The organization structure is
defined in terms of communicationpaths between actors, and the level of the hierarchy at which
reviews and approvals can be made.
Activities are the major processes that create the project deliverables. Activities are processed
as multiple chunks of information termed "communications" in a stochastic, discrete event
simulation. Activities have a numberof attributes including:
¯ Complexity of technical requirements (one of" high, medium,low): relative engineering
difficulty associated with the input to the engineeringworkof this activity;
¯ Interdependence (one of: high, medium,low): the degree to which activity performance
depends on analyses madeby other activities;
¯ Natural idiom (one or more natural idioms for the activity)
¯ Precedence relationships that namePredecessor and Successor activities;
¯ Reciprocal activities: other activities with this activity must workin close concurrence;
¯ Responsible actor
85
¯ Solution complexity(one of: high, medium,low): the relative engineering difficulty
associatedwith the solutioncreatedby this activity;
¯ Tasks: a project is divided into activities that in turn are divided into tasks (with
priorities), that in turn are divided into subtasks. Actualworkis performedby an actor
doing a subtask; tasks, etc. simplyprovidewaysof organizingsubtasks.
¯ Times,(start, end, early start, early end); computed
duration;
¯ Uncertainty (one of: high, medium,low): the degree to whichinputs and solutions are
well understood, stable and manageable.
¯ Workvolume(arbitrary quantitative units)
Actorattributes include:
¯ Coordinates-with: other actors with whomthis actor will share information with
informally. This relationship specifies an informalrelationship within the organization
hierarchy.
¯ Location
¯ Role in the team (one of: project manager,team leader, teammember)
¯ Skill type:the disciplineor craft of this actor;,
¯ Skill degree(one of: high, medium,low):
¯ Speed(a numericpercentageof nominal)
¯ Supervises and supervisedby the formal relationship that defines the organization
hierarchy;
¯ Task experience(one of: high, medium,low):
Actors have informationprocessing capability. TheVDTdoes not nowmodelthe quality of the
engineering judgmentof actors, i.e., their product, but it does modeltheir information
processing behavior. Actors perform a numberof functions including the following, all
implemented
as object-oriented methodsattached to appropriateobjects:
¯ Select messagesfroman "in-tray". Attention, or the informationselection process, is a
crucial feature of organizations. VDTuses stochastic "attention rules," implemented
as
methodson actor objects, to select communications
fromthe in-tray.
¯ Process information. Timeto process a messagedepends (stochastically) on the task
features, nominalduration, degreeof variability, and the degreeof the matchbetweenthe
attributes of an actor and a message.
¯ Sendmessagesto an "out-tray"for distribution. Distribution may(stochastically) be by
tool that is suitable for the particular type of communication
and the workloadof related
actors and tools. Actorshave methodsthat assign a priority to outgoingmessages,based
on the priority of the task.
¯ Choosea communicationtool for an outgoingmessage;considering task priority, tool
capabilities (e.g., synchronicity, natural idiom) and difference betweenactor and tool
locations.
¯ Generateactivities to coordinateactors, based on the need for exceptions to obtain or
share additional information, as suggested by (Galbraith); requirementsfor periodic
percentage-completion
updates; and requirementfor milestonereviewdictated by project
policy. VDTgenerates exceptions based on actor capability and degree of matchbetween
actor capability to processa particular messageand the messageattributes.
Actors use communicationtools to exchangeinformation with other actors. The VDTmodel
includes meetings, telephones, voice mail, electronic mail, file sharing, etc. The VDT
framework
represents each communication
tool in termsof a set of qualitative attributes that we
theorizedwouldaffect both the choiceof tool andthe results of that choice.
Communication
tool attributes include:
¯ Capacity(volumeof messagesthat can be transmitted concurrently);
¯ Location,usedto checkthe proximityof the actor and the tool;
86
¯ Naturalidiomssupportedby this tool;
¯ Recordability (whether or not a permanentrecord of the communicationis available
routinely);
¯ Relative speed for communicationsinvolving different natural idioms (e.g., text,
schematics, 3-Dgeometry);
¯ Synchronicity (one of: synchronous,asynchronous,partially synchronous)
Wehave performedobservations and built modelsfor several real engineering design tasks.
For example,one case (Case-1 in Figure 2) modeleda particular realistic design task and
baseline set of actor and communication
tool capabilities. Actors had a set of communications
tools including telephone, fax, voice mail, etc. Theactors had decentralized control. In the
simulation, computedproject duration was862 simulation units with a standard deviation of 8
units. This actual project duration lasted approximately485days, so the 862simulation units
correspondsto about 485 days for the assumptionsused in this project. Case-2held the task
and actor parameters constant, but actors could not use voice mail as allowed in Case-1.
Theorypredicts that the overall project duration in Case-2shouldbe marginallyincreased than
Case-l, i.e., duration "(Increased)" in the table. Bracketed numbersshowthat simulated
project duration wasgreater than in Case-1,as predicted.
Theresults are statistically significant, as suggestedby the relatively smallstandardvariations.
Case-3 used the sametask, tool and actor descriptions as in Case-l, except that actors had
centralized decision-makingcontrol. In each of Cases 2-4, there was three-way consistency
about change in duration with respect to Case-1 amongpredictions madeby the project
managerof the modeledproject, our predictions based on theory, and the simulationresults.
Information Processin 8 Technolosies
Voicemall
NoVoice Mall
Distributed. Decentralized
1. (Baseline)
[862, SD81
2. (Increased)
[888, SD1]
Hierarchical. Centralized
3. (Increased)
[930, SD10]
4. (Increased)
[970, SD16]
Decision-Makin
8
Figure 2. Qualitative changein simulatedproject performance
in different organizationalcases.
Parenthesizedremarksindicate the theoretically predictedchangein project duration of Cases-2,
3 and 4 with respect to the baseline of Case-1. Bracketednumbersshownumberof simulation
time units required to performthe project in each case, followedby the standard deviation on
those estimates. Weconclude that VDTcan consistently represent the informationprocessing
patterns of engineeringorganizationsdoingroutine tasks.
4. Conclusion
The simulation modelcan be "run" relatively quickly. Thus, it can serve as a facility to
formulate and test large numbersof precise conjectures regarding howchangesin management
87
structure or use of newcommunicationstools will affect the organization’s performance.
Engineeringdisciplines have long had mathematicaland, morerecently, computationalmodels
in support of analysis and optimizationof physical systems. This workshowsthat AI symbolic
modelingcan be usedto expressand to test social sciences modelsapplied to real organizations.
This work demonstrates the applicability of AI symbolic modeling techniques for making
modelsthat describe the detailed structure and predict the behaviorof social theory and human
organizations.
References
[Carley 90] Carley, K., Kjaer-Hansen, J., Prietula, M., Newell, A., "Plural-Soar: A
Prolegomenonto Artificial Agents and Organizational Behavior", GSIAWorkingPaper, Center
for the Management
of Technology,Carnegie MellonUniversity, December,1990.
[CohenG92] Cohen, G.P., "The Virtual Design Team:Aninformation processing model of
coordination in project design teams," Ph.D. dissertation, Stanford University Departmentof
Civil Engineering,1992.
[CohenM86] Cohen, M. D., Artificial Intelligence and the DynamicPerformance of
Organizational Designs," in Ambiguityand Command:
OrganizationalPerspectives on Military
Decision Making, March, J.G., and Weissinger-Baylon, eds., Pittman, Boston, MA,1986,
pp. 53-71.
[Galbraith 77] Galbraith, Jay., OrganizationDesign, Addison-Wesley,Reading, MA,1977.
[Levitt 93] Levitt, R.E., Cohen,G.P., Kunz,J.C., Nass, C.I., Christiansen, T., Jin, Y., "The
’Virtual Design Team’: Simulating HowOrganization Structure and Information Processing
Tools Affect TeamPerformance"in ComputationalOrganization Theory, Kathleen Carley and
MichaelPfietula, Eds., LawrenceErlbaumAssociates, Hillsdale, NewJersey, 1993, in press.
[Masuch 89] Masuch, M., LaPotin, P., "Beyond Garbage Cans: An AI Model of
OrganizationalChoice,"AdministrativeScience Quarterly, 34, pp. 38-67, 1989.
[Simon76] Simon, H.A., Administrative behavior: a study of decision-makingprocesses in
administrative organization, NewYork: Free Press, 1976.
[Tatum1984] Tatum, C.B., "DesigningProject Organizations: HowManagersDecide," Journal
of Construction Engineering and Management,ASCE,110 (3), September, 1984.
[Thompson67] Thompson,J.D., Organizations in Action, McGraw-Hill,NewYork, 1967.
88