From: AAAI Technical Report SS-94-07. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved. Task EnvironmentCentered Design of Organizations * Keith Decker and Victor Lesser Department of Computer Science University of Massachusetts Amherst, MA 01003 Email: DECKER@CS.UMASS.EDU Abstract The design of organizationsor other coordinationmechanismsfor groupsof computationalagents, either interacting with one another or with people, dependscrucially on the task environmentof whichthey are a part. Suchdependenciesinclude the structure of the environment(the particular kinds and patterns of interrelationships that occur betweentasks) and the uncertainty in the environment(both in the a priori structure of any episode within an environmentand in the outcomesof an agent’s actions). Designingorganizations also dependson properties of the agents themselves--butthis has been studied morethoroughlyby other researchers. Thecentral idea is that the designof coordinationmechanisms cannot rely on the principled construction of agents alone, but must rely on the structure and other characteristics of the task environment--forexample, the presence of uncertainty and concomitanthigh variancein a structure. Furthermore,this structure can and should be used as the central guide to the design of coordination mechanisms,and thus must be a part of any comprehensivetheory of coordination. This workingpaper will briefly describe our modeling framework,TASMS, for representing abstract task environments. Wewill also briefly describe a family of domain-independent,team-oriented coordination algorithms called Generalizedpartial Global Planning (GPGP).Havinga family of algorithms allows us tailor the algorithmto the environment(or evena specific situation). Wewill give an exampleof an analysis inspired by Burton and Obel’s workon organizational structure and technologydecomposability. Introduction Wehave developed a task environment-oriented modeling frameworkcalled T/EMS (Task Analysis, EnvironmentModeling, and Simulation)that features careful attention to the quantitative computationalinterrelationships betweentasks, to what information is available (and when)to update *This workwas supported by DARPA contract N00014-92-J1698, Office of NavalResearchcontract N00014-92-J-1450, and NSFcontract CDA 8922572.The content of the informationdoes not necessarilyreflect the positionor the policyof the Government andno official endorsement shouldbe inferred. 32 agent’s mentalstate, and to the general structure of the task environmentrather than single-instance examples[Decker and Lesser, 1993d; Decker and Lesser, 1994b]. Our task environmentmodelscan be used for both the analysis and simulation of coordination algorithms, and also to design organizational structures that are well-adaptedto particular task environments. The formof the TJEMSframework is moredetailed in structure than manyorganizational-theoretic modelsof organizational environments,such as Thompson’s notions of pooled, sequential, and reciprocal processes [Thompson, 1967], Burton and Obel’s linear programs [Burton and Obel, 1984], or Malone’s queueing models [Malone, 1987]. It is influenced by the importance of environmental uncertainty and dependencythat appear in contingency-theoretic and open systems views of organizations [Lawrenceand Lorsch, 1967; Galbraith, 1977; Scott, 1987; Stinchcombe,1990]. TASMSallows the quantitative operationalization of many organizational-theoretic environmentalconcepts, especially various dependenciesand uncertainties (see the decomposability example, from [Burton and Obel, 1984], later in this paper) that are the basis of both contingency theoretic and transaction cost economic(i.e. [Williamson,1975; Moe,1984]) viewsof organizations. Theobservation that no single organization or coordination mechanism is ’the best’ across environments,problem-solvinginstances, or even particular situations is also common in the study of cooperative distributed problemsolving [Fox, 1981; Durfeeet al., 1987; Durfee and Montgomery,1991; Decker and Lesser, 1993b]. Keyfeatures of task environmentsdemonstratedin both these threads of work that lead to different coordination mechanisms include those related to the structure of the environment(whatwe will call task interrelationships) and environmental uncertainty. Uncertainty. Less uncertainty in the environmentmeans less uncertainty in the existence and extent of the task interdependencies, and less uncertainty in local scheduling-therefore the agents need less complexcoordination behaviors (communication, negotiation, partial plans, etc) For example, one can have cooperation without communication [Geneserethetal., 1986]if certain facts are knownabout the task structure by all agents. Sometimes agents will not have a priori informationabout the structure of an episode, but they will be able to get information about a subset of the structure after the start of problemsolving--nosingle agent workingalone will be able to construct a view of the entire problemfacing the group. Someof this uncertainty mayarise fromdisparities in the objective (real) task structure and the subjective(agent-believed) task structure, either initially or the situation evolvesovertime. Taskinterrelationships. Taskinterrelationships include the relationships of tasks to the performance criteria by whichwe will evaluatea system,to the control decisionstructuresof the agents which makeup a system, and to the performanceof other tasks. TheT2EMSobjective subtaskrelationship indicates the relationship of tasks to systemperformance criteria; the subjective mappingindicates the initial beliefs available to an agent’s control decision structures; various non-local effects such as enables and facilitates indicate howthe execution of one task affects the duration or quality of another task. Eachrelationship is quantitatively defined. Whena relationship extendsbetweenparts of a task structure that are subjectivelybelievedby different agents,wecall it a coordination relationship. The basic idea behindGeneralizedPartial Global Planning (GPGP)is to detect and respond appropriately to these quantitative coordinationrelationships. The mainthrust of this workingpaper is to present TASMS as a frameworkthat can be used for computationalorganization design; the remainderof this paper will give a brief overviewof TA~MS, a simple architecture for agents working in a TEEMSenvironment,a brief overviewof the GPGP family of team-orientedcoordination algorithms, and an exampleof a simple analysis. Finally wewill discuss howwe can extend our analyses to more complex,hierarchical organizational structures. Short Overview of TASMS Theprinciple purposeofa TASMS modelis to analyze, explain, or predict the performanceof a systemor somecomponent. WhileT/EMSdoes not establish a particular performance criteria, it focuseson providingtwokindsof multi-criteria performanceinformation:the temporalintervals of task executions, and the quali~y of the executionor its result. Quali~yis an intentionally vaguely-definedterm that mustbe instantiated for a particular environmentand performancecriteria. Examplesof quality measures(whichcan be vectors) include the precision, belief, or completeness of a task result. T/EMSmodels describe howseveral quantities--the quality producedby executing a task, the time taken to performthat task, the time whena task can be started, its deadline, and whether the task is necessaryat all--are affected by the executionof other tasks. A TA~MSmodelof environmentaland task characteristics has three levels: objective, subjective, and generative. The objectivelevel describesthe essential, ’real’ task structure of a particular problem-solving situation or instance over time. Thesubjectivelevel describesthe agents’viewof the situation. A subjective level modelis essential for evaluatingcoordination algorithms, becausewhile individual behaviorand system performancecan be measuredobjectively, agents must makedecisions with only subjective informationJ Finally, the generativelevel describesthe statistical characteristicsof 1 In organizational theoreticterms,subjectiveperceptioncanbe usedto predictagentactionsor ouqauts,whileunperceived, objective environmental characteristicsaffect performance (or outcomes) [Scott, 1987]. 33 the objective and subjective situations in a domain.A generative level modelconsists of a description of the generative processes or distributions from whichthe range of alternative probleminstances can be derived, and is used to study performanceover a range of problemsin an environment. A probleminstance (called an episode E) is defined as a set of task groups, each with a deadline D(T), such E = (TI,’T2, ¯ ¯ ¯, Tn). The task groups(or subtasks within them)mayarrive at different times. Whiletask groupsare in2, the tasks within dependentof one another computationally a single task group are in general not independent.Figure 2 showstwo objective task groups, and Figure 3 showsagent A’s initial subjective view of that task group. A task group consists of a set of tasks related to oneanotherby a su btask relationship that formsan acyclic graph (here, a tree). The circles higher up in the tree represent various subtasks involvedin the task group, and indicate precisely howquality will accrue dependingon what leaf tasks are executed and when.Tasksat the leaves of the tree (withoutsubtasks) represent metho&,whichare the actual computationsor actions the agent will execute(in the figure, these are shownas boxes). Thearrows betweentasks and/or methodsindicate other task interrelationships wherethe execution of somemethodwill havea positive or negative effect on the quality or duration of another method.The presenceof these interrelationships makethis an NP-hardscheduling problem; further complicating factors for an agent’s local schedulerinclude the fact that multiple agents are executingrelated methods,that some methodsare redundant(executable at morethan one agent), and that the subjective task structure maydiffer from the real objective structure. This notation and associated semantics are formally defined in [Deckerand Lesser, 1993d; Deckerand Lesser, 1994b]. Hospital Patient Scheduling Example Let’s look at a brief exampleofa T/EMStask structure model in termsof its ability to reasonaboutorganizationaldecision making. The following description is from an actual case study [OFetal., 1989]: Patients in GeneralHospitalreside in units that are organizedby branchesof medicine, such as orthopedicsor neurosurgeryEachday, physiciansrequest certain tests and~ortherapyto be performedas a part of the diagnosis andtreatmentof a patient. [... ] Tests are performedby separate,independent,anddistally located ancillary departmentsin the hospital. Theradiologydepartment,for example,providesX-rayservices andmayreceive requests ~oma numberof different units in the hospital. Furthermore,each test mayinteract with other tests in relationships such as enables, requires-delay (must be performedafter), and inhibits (test A’s performance invalidates test B’s result ifA is performedduringspecified time period relative to B). Notethat the unit secretaries (as scheduling agents) try to minimizethe patients’ stays in the hospital, while the ancillary secretaries (as schedulingagents) try maximizeequipment use (throughput) and minimize setup times. 2 Exceptfor the useof the processor(s) or otherphysicalresources [DeckerandLesser, 1994b]. fNu rsing Unit 1 "~ fNursing Unit 1 method(executable task) (~ task with quality accrual function min task already communicated to ancillary subtaskrelationship ....... ~. enablesrelationship ,---~ ~-~-~ requires delay inhibits 1 . Figureh High-level, subjectivetask structurefor a typicalhospitalpatientscheduling episode.Thetoptask in eachancillaryis really the same objectiveentityas the unit task it is linkedto in the diagram. Figure 1 showsan subjective TAZMS task structure corresponding to an episode in this domain,and the subjective viewsof the unit and ancillary schedulingagents after four tests have been ordered. Note that quite a bit of detail can be captured in just the ’computational’aspects of the environment--inthis case, the tasks use peoples’ time, not a computer’s. However,TA~MS can modelin more detail the physicalresourcesand job shopcharacteristics of the ancillaries if necessary[DeckerandLesser, 1994b].Suchdetail is not necessary for us to analyze the protocols developedby [Ow et al., 1989], whoproposea primaryunit-ancillary protocol and a secondaryancillary-ancillary protocol. Weuse min (AND)to represent quality accrual because general neither the nursing units nor ancillaries can change the doctor’s orders--all tests mustbe doneas prescribed. We have added two newnon-local effects: requires-delay and inhibits. The first effect says that a certain amount8 of time must pass after executingone methodbefore the second is enabled. The second relationship, A inhibits B, means that B will not produceany quality if executedin a certain windowof time relative to the execution of A, and can be defined in a similar manner. Analysis in TASMS The methodology we have been building uses the TA~MS frameworkand other DAIformalismsto build and chain together statistical modelsof coordinationbehaviorthat focus on the sources of uncertainty or variance in the environment and agents, andtheir effect on the (potentially multi-criteria) performanceof the agents. For example, we have used this methodologyto develop expressions for the expected value of, and confidenceintervals on, the time of terminationof a set of agents in any arbitrary simple distributed sensor networkenvironmentthat has a static organizational structure and coordination algorithm [Decker and Lesser, 1993b]. We have also used this modelto analyze a dynamic,one-shot 34 reorganization algorithm (and have shownwhenthe extra overheadis worthwhileversus the static algorithm) [Decker and Lesser, 1993c]. In each case we can predict the effects of adding moreagents, changingthe relative cost of communication and computation, and changing howthe agents are organized (in this case, by changingthe range and overlap of their capabilities). Theseresults wereachievedby direct mathematicalanalysis of the modeland verified throughsimulation in T/EMS. Wewill omit the details here and refer the reader to our papers. An Agent Architecture The T/EMSframeworkmakesvery few assumptions about the architecture of agents--agentsare loci of belief and action. Agents have somecontrol mechanism that decides on actions giventhe agent’scurrentbeliefs. Thereare three classes of actions: methodexecution, communication,and information gathering. Methodexecution actions cause quality to accrue in a task group (as indicated by the task structure). Communicationactions are used to send the results of method executions(whichin turn maytrigger the effects of various task interrelationships) or meta-level information.Information gatheringactions add newlyarriving task structures, or newcommunications, to an agent’s set of beliefs. Formally, we write BtA(X)to meanagent A subjectively believes x at time t [Shoham,1991]. Wewill shorten this to B(x) when wedon’t have a particular agent or time in mind. The GPGPfamily of coordination algorithms makes stronger assumptionsthan TASMS about the agent architecture. Mostimportantly, it assumesthe presence of a local scheduling mechanismthat can decide what method execution actions should take place and when, which will be described in the next section. It assumesthat agents do not intentionally lie and that they believe whatthey are told. The Local Scheduler Each GPGPagent contains a local scheduler that takes as input the current, subjectively believed task structure and produces a schedule of what methods to execute and when. It chooses these methods in an attempt to maximizea pre-definedutility measure;in this paper the utility function is the sum of the task group qualities: U(E) ~7"~EQ(T, D(T)), where Q(T, t) denotes the quality of 3T at time t as defined in [Deckerand Lesser, 1993d]. Besidethe subjective task structure, the local scheduler should accept a set of commitmentsC from the coordination component. These commitmentsare extra constraints on the schedules that are producedby the local scheduler. For example, if method1 is executable by agent A and method 2 is executable by agent B, and the methodsare redundant, then agent A’s coordination mechanismmaycommitagent A to do method1 both locally and socially (commitmentsare directed to particular agents in the sense of [Shoham,1991; Castelftanchi, 1993]) by communicating this commitment to B (so that agent B’s coordination mechanismrecords agent A’s commitment,see the description of non-local commitments NLCnext). The final piece of informationthat is used by the local scheduler is the set of non-local commitments madeby other agents NLC. This information can be used by the local scheduler to coordinate actions betweenagents. For example the local scheduler should havethe property that, if method Mxis executable by agent A and enables method M2at agent B (and agent B knowsthis), then if agent A makes commitmentto B to communicatethe results of M1by time t, B will only scheduleM2after time t. 4 Thuswe maydefine the local scheduler as a function LS(E, C, NLC)returning a set of schedules S = {$1, $2,... , S,m,). Moredetailed information about this kind of interface betweenthe local scheduler and the coordination componentmaybe found in [Garveyeta/., 1994]. and Lesser, 1992]. Module1 exchangesuseful private views of task structures; Module2 communicatesresults; Module 3 handles redundant methods; Modules4 and 5 handle hard and soft coordination relationships. Moremoduleshave been designed, such as a load-balancing module. The modules are independentin the sense that they can be used in any combination. The GPGP Coordination Modules The GPGP algorithmfamily specifies three basic areas of the agent’s coordination behavior: howand whento communicate and construct non-local views of the current problem solving situation; howand whento exchangethe partial results of problemsolving; howand whento makeand break commitments to other agents about whatresults will be available and when.5 The coordination mechanismsupplies nonlocal viewsof problemsolvingto the local scheduler,including non-local commitments about what non-local results will be available locally, andwhenthey will be available. The five modulesdescribed in [Deckerand Lesser, 1994a] forma basic set that providessimilar functionalityto the original partial global planningalgorithmas explainedin [Decker 3Notethat qualitydoesnot accrueafter a task group’sdeadline. 4Ofcoursein generala local schedulercannotbe optimal,and this holdsfor human agents as well. Wecan exploreother models of local scheduling, prioritization,attention,style, etc. [March and Simon,1958].Someworkis alreadybeingdonein this area [Lin and Carley,1993;J.C. Kunz,1993]. 5Theuse of commitments in the GPGP familyof algorithmsis basedonthe ideas of many researchers[CohenandLevesque,1990; Shoham, 1991;Castelfranchi,1993;Jennings,1993]. 35 Simulation Example Simulationis a useful tool for learning parametersto control algorithms, for quickly exploring the behavior space of a newcontrol algorithm, and for conductingcontrolled, repeatable experimentswhendirect mathematicalanalysis is unwarranted or too complex. The simulation system we have built for the direct execution of modelsin the TtEMS frameworksupports, for example,the collection of paired responsedata, wheredifferent or ablated coordinationor local schedulingalgorithmscan be comparedon identical instances of a widevariety of situations (generatedusing the generative level of the model). Wehave used simulation to explore the effect of exploiting the presenceof facilitation betweentasks in a multi-agent real-time environmentwhereno quality is accrued after a task’s deadline [Deckerand Lesser, 1993a]. The environmentalgenerative characteristics here included the meaninterarrival time for tasks, the likelihood of one task facilitating another, and the strength of the facilitation (Od)" In contrast to the previousanalytical work,in this section wewill considerthe use of T/EMS as a simulatorto explorehypotheses about the interactions betweenenvironmentaland agent-structural characteristics. Weuse as an examplea question exploredby Burtonand Obel:is there a significant difference in performance due to either the choiceof organizational 6? structure or the decomposabilityof technology Weequate a technologywith a TASMS task structure, instead of a linear program.Taskstructures allow us to use a clear interval measurefor decomposability,namelythe probability of a task interrelationship (in this exampleenables, facilitates, and overlaps). Wedefine a nearly decomposable task structure to havea baseprobabilityof 0.2 for these three coordinationrelationships and a less decomposable task structure to havea base probability of 0.8 (see Figure 2). Wewill continuein this exampleto look at purely computationaltask structures, althoughthe use of physical resourcescan also be represented (see [Deckerand Lesser, 1994b]). Burton and Obel were exploring the difference in M-form (multidivisional) and U-form(unitary--functional) hierarchical structures; we will analyze the GPGP family of teamoriented coordinationalgorithms. For our structural variable we will vary the communicationof non-local views (GPGP Module1). Informally, we will be contrasting the situation where each agent makes commitments and communicates results based only on local information(no non-local view) with one wherethe agents freely share task structure information with one another across coordination relationships (partial non-local view). Figure 3 showsan example--note 6 Technology is usedhere in the management sciencesenseof"the physicalmethodby whichresourcesare convertedinto productsor services" or a "meansfor doingwork"[Burtonand Obel, 1984; Scott, 1987]. ~ / ~ ~ method ~ k minj~ ~~ (executable ~ task ~ ~ " "kmax] ~ task) with quality accrual function rain 1,~ enablesrelationship .... facilitates relationship ~:-, ~~~ ~ Less Decomposable (p=0.8) MoreDecomposable (p=0.2) Figure2: Example of a randomly generatedobjectivetask structure, generatedwiththe parameters in the previoustable. that in neither case does the agent have the global view of Figure2. Burtonand Obeluseda profitability indexas their performancemeasure,derived from the percentageof optimal profit achieved. In general, the schedulingan arbitrary T/EMStask structure is an NP-hardproblemand so we do not have access to optimal solutions. Instead we compareperformance directly on four scales: the numberof communicationactions, the amountof time spent executingmethods,the final quality achieved, and the termination time. Simulationruns for each of the four combinationsof non-local view policy and level of task decomposabilitywere done in matched sets--the randomlygenerated episode was the samefor each combinationwith the exception of morecoordination relationships (including moreoverlappingmethods)being added in the less decomposable task structures. FollowingBurton and Obel, we used the non-parametric Friedman two-way analysis of variance by ranks test for our hypotheses. The assumptions of this test are that each block(in our case, randomlygenerated episode) is independentof the others and that each block contains matchedobservations that maybe ranked with respect to one another. The null hypothesis is that the populationswithin each blockare identical. Wegenerated 40 randomepisodes of a single task group, each episodewas replicated for the four combinationsin each block. Weused teamsconsisting of 5 agents; the other parameters used in generatingthe task structures are summarized in Table1 and a typical randomlygeneratedstructure is shown in Figure2. Figure3 showsthe differencein the local viewof one agent with and without creating partial non-localviews. Wefirst tested twomajor hypotheses: Hypothesis1: There is no difference in performancebetween agentswith a partial non-localview andthose without. For the communicationand method execution performance measures,wereject the null hypothesisat the 0.001level. Wecannotreject the null hypothesisthat there is no difference in final quality and termination time. Teamsofcom- Parameter Value MeanBranchingfactor (Poisson) MeanDepth(Poisson) MeanDuration(exponential) Redundant MethodQAF Number of task groups TaskQAF distribution Decomposition parameter HardCRdistribution Soft CRdistribution 1 3 10 Max 1 (50%rain) (50%max) p = 0.2 or 0.8 (p enables)((l-p) (p facilitates) (10% hinders) ((.9-p) none) P Chanceof overlaps(binomial) Table1: Parametersusedto generatethe 40 random episodes putational agents that exchangeinformation about their private, local views consistently exchangemoremessages (in this experiment,a meanincrease of 7 messages)but less work(here, a meandecrease of 20 time units of work, probably due mostly to avoiding redundancy). Hypothesis2: Thereis no difference in performance due to the level of decomposabil#yof technology. For the communication and methodexecution performance measures, we reject the null hypothesisat the 0.001 level. Wecannot reject the null hypothesisthat there is no difference in final quality and termination time. Teamsof computational agents, regardless of their policy on the exchange of private, local information communicatemoremessages (in this experiment, a meanincrease of 47 messages)and do more work (here, a meanincrease of 24 time units) when faced with less decomposablecomputational task structures (technology). Again following Burton and Obel, we next test for interaction effects betweennon-local view policy and level of technologydecomposabilityby calculating the differences in performanceat each level of decomposability,and then test- 36 ing across non-localviewpolicy. This test wassignificant at the 0.05 level for communication,meaningthat the difference in the amountof communication under the two policies is itself different dependingon whethertask decomposability is high or low. This difference is small (two communication actions in this experiment)and was not verified in a second independenttest. 5 NoA~J-Local Vitwat AgentA (p*0.2) Figure3: Example of the local viewat AgentAwhenthe teamshares privateinformationto create a partial non-localviewandwhenit doesnot. Conclusions & Future Directions TA~MS is a frameworkfor describing complextask environments. Whencombinedwith traditional DAItools for describing coordination algorithms and agents, it provides a basis for analysis and/or simulationon either the TI Explorer Lisp machine or the DECAlpha AXP.Whenanalyzing an existing or proposedorganizational design or coordination algorithm, we advocatean approachthat focuses first on behavior due to coordinationrelationships in certain situations and then expandingthis modelto incorporate the sources of uncertainty that are present. Sucha process mayiteratively refine the organization or algorithm--with a parameterized algorithm one might approach this as a pure parameter optimization problem.In [Deckerand Lesser, 1993b] we also discuss howthe modelof performancein a particular episode can be used with meta-level communication to choosea good organizational design dynamically wheninformation about the current situation becomesavailable to the agents [Stinchcombe, 1990]. On the other hand, when trying to design an organization or coordination algorithm for a given environment, one can also start with both the coordination relationships and the uncertainties present, and add features to deal with each explicitly (our implementationof the GPGP algorithm, whichfeatures several independent’plug-in’ modules to deal withdifferent classes of interrelationships, is a case in point). The two-level agent architecture that we use with the GPGPfamily (with its separate coordination and local schedulingfunctions) is not necessaryfor using TASMS, whichviews agents only as consumersof subjective beliefs and producersof actions. Representing Organizational Roles Future workwill include the analysis of moretraditional hierarchical organization forms. The addition of the concept of ’organizational role’ can be accomplishedthroughthe use of non-local commitments(and their accompanyingexpectations). In its simplestsense, agentA’s organizational role is a set of continuingnon-local commitments to certain classes of tasks. By making continuous commitmentsagents can avoid communicatingabout every episode anew--assuming the future structures are somewhatpredictable. This reasoning leads naturally to learning organizationalroles and to open systems conceptsof temporarily settled questions (the current set of continuous commitments)and algorithms to re-open and re-settle these questionsas the task environment changes. The developmentof such algorithms can then lead to exploring conceptions of agents as nothing more than convenient bundles of organizational roles [Gerson, 1976; Gasser et al., 1989], and to whenand whyorganizations do not act like ’big agents’ [Allison, 1971].Wewill also the expansionof commitments to recta-level roles (i.e., comitments to the coordinationparametersin effect for certain classes of tasks). The Tower of Babel Upto this point, we have focusedon the interrelationships and the uncertaintyarising directly fromthe generative model of the environment,but wewouldalso like to explorethe uncertainty and variancearising fromthe differencein the objective and subjective models.Different relationships between these modelswill lead to different organizationalstructures. For example,in the case of uncertainty arising fromthe generative modelof an environment, we showedthat the wide variance in performanceof a systemof agents with static organizationsin different episodesled to the use of recta-level communication to reorganize the agents to adapt to the particular episode at hand [Deckerand Lesser, 1993c]. Weare also looking to expandthe T/EMS conceptionof environments to encompassmore dynamicsituations: another important source of environmentaluncertainty. For example, in the Towerof Babel problem[Ishida, 1992], agents try to pile numberedblocks into a tower in numerical order. When manyagents try to solve this problemwithout centralized coordination, they tend to bottleneck aroundthe tower itself, bumpinginto one another. Ishida solves this problem by dividing the agents into two groups:one to collect blocks and bring them near (but not too near) the tower, and one group to stack the tower. Fromour point of view, the task structure at lowlevels is changingrapidly, and no agent can (or needsto) keeptrack of all the changesin relationships betweenagent and block positions. The rapid changein the environmentmeansthat agents’ subjective views are always out of date. By dividing the agents into two groups, the bottleneck resource is controlled by organization--removing it fromthe uncertainties facing the agents. Stackingagents knowthere are not enoughof themto saturate the bottleneck, and other agents no longer use the resourceat all. The uncertainty is still there, but it no longer has an impacton the agents’ decision-makingprocess. A similar exampleis Gasser and Ishida’s workon self-organizing productionsystems [Gasser and Ishida, 1991] where dynamicchanges in the environmentallowed the systemto reorganize itself for more efficient performance. Wecould potentially analyze either of these systems. References Allison, Graham T. 1971. Essenceof Decision:Explainingthe CubanMissle Crisis. Little, Brownand Company,Boston. 37 Burton, Richard M. and Obel, Borge1984. DesigningEfflcient Organizations:Modellingand Experimentation.North Holland, Amsterdam. Castelfranchi, C. 1993. Commitments:from individual intentions to groupsand organizations. In Prietula, Michael, editor 1993,AI andtheories of groups&organizations:Conceptual and Empirical Research. AAAIWorkshop.Working Notes. Cohen,Philip R. and Levesque,Hector J. 1990. Intention is choice withcommitment. Artificial Intelligence 42(3). Decker, Keith S. and Lesser, Victor R. 1992. Generalizing the partial global planningalgorithm.International Journal of Intelligent andCooperativeInformationSystems1 (2):319346. Decker, Keith S. and Lesser, Victor R. 1993a. Analyzinga quantitative coordination relationship. GroupDecision and Negotiation2(3): 195-217. Decker, Keith S. and Lesser, Victor R. 1993b. Anapproach to analyzing the need for meta-level communication.In Proceedingsof the Thirteenth InternationalJoint Conference on Artificial Intelligence, Chamb~ry, France. Decker, Keith S. and Lesser, Victor R. 1993c. A oneshot dynamiccoordination algorithm for distributed sensor networks.In Proceedingsof the EleventhNationalConference on Artificial Intelligence, Washington. 210-216. Decker,Keith S. and Lesser, Victor R. 1993d. Quantitative modelingof complexcomputational task environments.In Proceedingsof the EleventhNationalConference on Artificial Intelligence, Washington.217-224. Decker, Keith S. and Lesser, Victor R. 1994a. Designinga family of coordination algorithms. Submittedto AAAI-94. Decker,Keith S. and Lesser, Victor R. 1994b. Quantitative modelingof complexenvironments.InternationalJournalof Intelligent Systemsin Accounting,Finance,andManagement. To appear in special issue on "Mathematicaland Computational Modelsof Organizations:Modelsand Characteristics of AgentBehavior". Durfee, E. H. and Montgomery,T. A. 1991. Coordination as distributed search in a hierarchical behaviorspace. IEEE Transactionson Systems, Man,and Cybernetics21(6): 13631378. Durfee, EdmundH.; Lesser, Victor R.; and Corkill, Daniel D. 1987. Coherent cooperation amongcommunicating problemsolvers. IEEETransactionson Computers 36(11):1275-1291. Fox, MarkS. 1981. Anorganizational view of distributed systems. IEEETransactionson Systems, Man,and Cybernetics 11(1):70-80. Galbraith, J. 1977. OrganizationalDesign.Addison-Wesley, Reading, MA. Garvey, Alan J.; Decker, Keith S.; and Lesser, Victor R. 1994. A negotiation-based interface betweena real-time scheduler and a decision-maker. Submittedto AAAI-94. Gasser, Les and Ishida, Toru 1991. A dynamicorganizational architecture for adaptiveproblemsolving. In Proceedings of the NinthNationalConference onArtificial Intelligence, Anaheim. 185-190. 38 Gasser, Les; Rouquette, N. E; Hill, R. W.; and Lieb, J. 1989. Representingand using organizational knowledgein distributed AI systems. In Gasser, L. and Huhns,M. N., editors 1989,DistributedArtificial Intelligence,Vol. II. Pitman PublishingLtd. Genesereth,M. R.; Ginsberg, M.L.; and Rosenschein,J. S. 1986. Cooperation without communication.In Proceedings of the Fifth NationalConference on Artificial Intelligence, Philadelphia, PA. 51-57. Gerson, Elihu M. 1976. On ’quality of life’. American Sociological Review41:793-806. Ishida, Toru 1992. Towerof Babel: Towardsorganizationcentered problemsolving. In Proceedingsof the 1 lth Workshopon Distributed Artificial Intelligence, TheHomestead, Michigan. 141-153. J.C. Kunz,R. E. Levitt G. R Cohen1993. Modelingeffects of organizational structure and communicationtools on design teamproductivity. In Prietula, Michael,editor 1993, AI and theories of groups&organizations: Conceptualand Empirical Research. AAAIWorkshop.WorkingNotes. Jennings, N. R. 1993. Commitmentsand conventions: The foundation of coordination in multi-agent systems. TheKnowledgeEngineeringReview8 (3) :223-250. Lawrence, Paul and Lorsch, Jay 1967. Organization and Environment. HarvardUniversity Press, Cambridge,MA. Lin, Z. and Carley, K. 1993. Doesagent style matter in organizational decision making?In Prietula, Michael,editor 1993, A/andtheories of groups&organizations: Conceptual and Empirical Research. AAAIWorkshop.WorkingNotes. Malone, ThomasW. 1987. Modelingcoordination in organizations and markets. Management Science 33:1317-1332. March, J. G. and Simon, H.A. 1958. Organizations. Wiley, NewYork. Moe,Terry M. 1984. The new economicsof organization. AmericanJournalof Political Science 28(4):739-777. Ow,P. S.; Prietula, M. J.; and Hsu, W1989. Configuring knowledge-based systems to organizational structures: Issues and examplesin multiple agent support. In Pau, L. E; Motiwalla, J.; Pao, Y. H.; and Teh, H. H., editors 1989, Expert Systems in Economics,Banking, and Management. North-Holland, Amsterdam.309-318. Scott, W.Richard 1987. Organizations:Rational, Natural, and OpenSystems. Prentice-Hall, Inc., EnglewoodCliffs, NJ. Shoham, Yoav 1991. AGENT0:A simple agent language and its interpreter. In Proceedingsof the Ninth National Conferenceon Artificial Intelligence, Anaheim. 704-709. Stinchcombe,Arthur L. 1990. Information and Organizations. Universityof California Press, Berkeley,CA. Thompson, James D. 1967. Organizations in Action. McGraw-Hill, NewYork. Williamson,Oliver E. 1975. Marketsand Hierarchies:Analysis andAntitrust Implications. The Free Press, NewYork.