From: AAAI Technical Report SS-98-05. Compilation copyright © 1998, AAAI (www.aaai.org). All rights reserved. TowardUbiquitous Satisficing Agent Control * Thomas Wagner and Victor Lesser Computer Science Department University of Massachusetts Amherst, MA01003 Email: { wagner, lesser} @cs.umass.edu Abstract generally non-trivial, requiring either exponential workor, in practice, a sophisticated solution schemethat controls the algorithmic complexityand makesagent control a tractable problem. Our most recent work in agent control, namely Designto-Criteria (Wagner, Garvey, & Lesser 1998; 1997a) scheduling and GPGP(Decker 1995; Prasad & Lesser 1996) multi-agent coordination is directed at addressing such agent control issues. Webelieve that the key to operating in these open environmentsis the ability to satisrice ubiquitously, and dynamically,to adjust problemsolving to the changing environmentand problem solving context. Oneview of this idea, that also appears in our work, is the notion of resource-bounded-reasoning,i.e., solving a problemin different amountsof time or with different allotments of computationalresources, perhaps trading-off solution quality and resource consumption. This is classically illustrated by an anytime (Zilberstein 1996) curve and typically translates into parameterizingthe size of the solution space that is searched during problemsolving (or the level of abstraction at which problemsolving occurs). However,this view is only part of the picture. Satisficing, and the ability to satisfice, is intertwinedwith the ability to approachproblemsolving froma flexible perspective, to attack problemsusing different techniques or using different problem solving parameter settings, or consumingdifferent resources or quantities of resources. Flexibility enables satisficing behavior. The existence of choices or alternatives begets the ability to performdifferently in different circumstances- this is whythe ability to satisfice is important. In different circumstances, different problemsolving behaviors are appropriate. The choice of whichsatisficing behavior to employis situation specific, i.e., dependenton the problemsolving context at hand. These three concepts, satisficing, flexibility, and situation specificity are all interrelated and the boundarybetweenthese concepts blur when they are examined closely. Through our work, we have cometo see and partially understand the relationships between these concepts and the importance of addressing all of these notions whenbuilding agent control systems. Un- The dynamicsand unpredictability of open environments pose a challenge to agent development.Agentsoperating in these environmentsmustbe able to adapt processingto the availableresourcesandthe currentproblemsolvingcontext. Theability to satisfice ubiquitously,in all aspectsof agent problemsolving,is the key to existing in these environments.Our workon satisficing agent control problem solving,i.e., multi-agentcoordinationandagentscheduling, servesas the foundationfor our current focus ona holistic approachto satisficing agentcontrol. Introduction With the advent of open computing environments adaptability in softwareapplications is critical. Since openenvironmentsare less predictable, applications must be able to adapt their processing to the available resources (hardware, software, time, money,databases, etc.) and the different goal criteria set by different clients. In agent basedsystems, where software applications are persistent, autonomous, goal oriented, and function in real-time, this requirement is doublyimportant. In this context agents must be able to reason about their local problemsolving activities, interact with other agents, plan a course of action, and carry out the actions in the face of limited resources and uncertainty about action outcomesand the actions of other agents, all in real-time and within real resource and cost constraints. Furthermore, in dynamicopen environments new tasks can be generated by existing or new agents at any time, thus an agent’s deliberation must be interleaved with execution - rescheduling, replanning, and recoordination with other agents is the norm, not the exception. To make matters even more difficult the planning and scheduling tasks are "Thismaterial is baseduponworksupportedby the National Science Foundationunder Grant No. IRI-9523419and the Departmentof the Navyand Office of the Chief of NavalResearch, underGrantNo.N00014-95i- 1198.Thecontent of the information doesnot necessarilyreflect the positionor the policyof the Government or the National ScienceFoundationand no official endorsement shouldbe inferred. 8O (Oather-Purchase-Data-on-Nissan-Maxima) (Gather-Reviews) ~sum0"~ [Edmund’s-Reviews]]Heraud’s-Test-Ddves ] ~" Q (20% 0)(80% O (2% 0X98%I0) D (5(~, 120X25%13(I) (25’~ 14(}) D(4(1%240)(601~[,300) Sublask R©latlon ~c’~" u,,Ji,y, c,~,, "’\\ Enable~N1.EImem°°l D = Dttratkm (Find-lnvoice-Price-Data) Intelichoice ] L.~ ~"’--"’’"~ max() "0 (I(K)~, 17) \ c (100% $4.95) " NA D(5(r~,48O)(5O% [Edmund’s-Price-Guide[ ~’c’6~bl~;->~ Q (5% 0X95% .(l(xI1) C(IIXE~,1)) D (5(1%3(I)(50% Q (5% 0X95%12) Q (11N1%24) C(11X)’-:~, (I) C (1110%$9.95) D (5(}9[, 2411)(50%2611) D (51)% 120)(25% (25% 14) Figure 1 : T,’EMS TaskStructure for GatheringAutoPurchaseInformation derstandingthe full ramificationsof these conceptsis taking the problemone step further. Wewill return to this issue in the architecture section, but clearly satisficing on one aspect of problem solving normally has direct consequences to other aspects of problemsolving, or problemsolving that occurs downstreamtemporally. fects), are also representedin T,’EMS and the effects of the interactions are reasonedabout statistically during scheduling and coordination. T/EMSmodels are the grounding element and mediumof exchange for Design-to-Criteria (Wagner, Garvey, & Lesser 1997a; 1997b) and Design-toTime (Garvey & Lesser 1995) scheduling, and and multiagent (Decker 1995) coordination research, and are being used in Cooperative-Information-Gathering(Lesser et al. 1997), collaborative distributed design (Decker &Lesser 1995), distributed situation assessment (Carver &Lesser 1995), and surviveable systems (Vincent et al. 1998) research projects. A simplified exampleof a Th~MS task structure for gathering auto purchase information via the Webis shownin Figure 1. The oval nodes are tasks and the square nodes are methods. The top-level task is to Gather-PurchaseData-on-Nissan-Maximaand it has two subtasks GatherReviews and Find-Invoice-Price-Data. The top-level task accumulatesquality according to the sum_all() quality accumulationfunction (qaf) 1 so both of its subtasks must be performed to satisfy the objective. The Gather-Reviews task has two methods, query Edmund’s-Reviewsand query Heraud’s-Test-Drives. These methods are governed by a sum() qaf thus the power-set of the methods minus the emptyset maybe performedto achieve the tasks, i.e., Edmund’s maybe queried, Heraud’s maybe queried, or both maybe queried. The Find-Invoice-Price-Datataskhas three subtasks, two of type methodand one of type task, governed by the max() qaf which is analogous to an ORrelationship. Note the decompositionof the obtain invoice via AutoSite task into two methods, one that locates the URLand one that issues the query. The enables NLEbetween the URL finding methodand the query method, in conjunction with the low quality associated with the URLfinding method, indicate that finding the URL is necessary for task achievementbut that it contributes very little to achievingthe task relative to the methodthat actually obtains the pricing report. Before delving into the details of our satisficing agent control work, let us ground further discussion in a highlevel overview of the modeling frameworkused in these research projects. Our research focuses on a class of computational task structures wherethere are typically multiple goals and multiple different actions for performinga particular task, whereeach action has different statistical performancecharacteristics, and the outcomesof actions are highly uncertain. Wemodel these problem solving activities using the T,’EMS(Decker 1995) domain-independent hierarchical task modeling framework. In T,’EMS,primitive actions, called methods,are modeledstatistically via discrete probability distributions in three dimensions,quality, cost, and duration. Quality is a deliberately abstract domain-independent concept that describes the contribution of a particular action to the overall problemsolving objective. Duration describes the amountof time that a method will take to execute. Cost describes the financial or opportunity cost inherent in performing the action modeledby the method.It is no accident that problemsolving activities are modeledin these four (uncertainty via the probability distributions) dimensions.Satisficing systems must operate in multiple dimensionsto address the multi-dimensionality of the environments in which they exist. Problemsolving issues are not typically limited to time, other items are also important, e.g., resources, financial costs, robustness, and precision. Theability to adapt problemsolving for different multi-dimensionalsets of goal criteria, amenableto future extension,is critical. Returning to T,’EMS,as with most hierarchical representations the high-level task is achieved by achieving some combinationof its subtasks. Accordingly, since different methodshavedifferent quality, cost, duration, and certainty trade-offs, different solutions and partial solutions also have different characteristics and different trade-offs. Hardand soft interactions betweentasks, called NLEs(non-local ef- 1Qafsdefinehowa giventask is achievedthroughits subtasks or methods.Thesum_edlO qaf meansthat all of the subtasksmust be performed andthat the task’s qualityis a sumof the qualities achievedby its subtasks. 81 ProblemSolver (D~rnainPn~blernSolving) Enumrratex I)omaiJ andSy.vtrrn Con.vtraintx DomainPlans * Resource Constrainta * Client Goal Criteria * * Schedule Sdectedfi*r Execution * Envelopc.~ Defining Ranges of Acceptable Scheduler ( L~v.:alContru]) CommitraenL~ to Other Agents* CiwarnitmenL~frum Other AgenL~* CommitmentImportance Criteria Ctamlination Module (Non-ta~cal C(mccrns) tives. In our work, agents generally exist in a multi-agent environment and each agent consists of three primary components: a domainproblemsolver, a scheduler that determines which domainactions to perform and in what sequence, and a coordination modulethat coordinates interactions with other agents. A high-level view of these three components is shownin Figure 2. In these systems, the domainproblem solver describes domainproblemsolving options in the T,’EMSlanguage and emits the generated task structures along with goal criteria that specifies the desired quality, cost, duration, and certainty of a path throughthe task structure that achieves the high-level task. Thescheduler inputs the task structures and goal criteria and constructs a schedule customdesigned to achieve the task while adhering to the criteria specified by the problemsolver. The coordination moduleworksin conjunction with the scheduler to handle interactions with other agents by giving and receiving commitmentsto perform work at certain times. The importance of the non-local interactions is weighedagainst local problem solving concerns during scheduling; the resulting schedule is returned to the problemsolver for execution. The schedule is annotated with performance envelopes and monitoring points define whenrescheduling is necessary due to unexpected (or unprobable) execution results. Note that each agent has its ownlocal scheduler and local coordination module - agents are autonomousand there is no centralized locale where all scheduling or coordination takes place. Agents can behave in a more aggregate fashion, but this is through the use of organization rather than fixed centralized control. Specification of Local Pml,lem Solving Actions Best Suited for thr Current Lot~#l Context Mtululaled by the Non-l~Jt’tdConlex! ] Ramificationsof Lx,cal Control to Non-Lt~’alContext * Satisfied Commitments \Data Flow * Violated CommitmenL~ * ScheduleSelccW..dfi,r Execution Figure 2: Key Agent Components T/EMSmodels several important task features that facilitate a satisficing approach to problemsolving. The existence of alternative ways to perform tasks gives T~MS based problem solvers (schedulers, coordination algorithms) options about howto achieve tasks. Another feature of the modelthat supportssatisficing is the statistical characterizationsof primitive actions. Thedifferent statistical characteristics of alternative waysto achieve tasks gives us the ability to explore different possible solution paths, and consider the different trade-offs of each path, to find one that best satisfices to meet the current problemsolving context. This is the T.’EMSscheduling problem. For a given task structure and a given problemsolving context, find a way(in real-time) to achieve the high level task that is in keepingwith the context and the client’s resource constraints. Wewill discuss our Design-to-Criteria satisficing scheduling workin more detail later in the paper. Modeling task interactions or non-local-effects is also an important feature of T?EMS. Explicit representation of both the quality of interactions and the quantified ramificationsof interactions provides Th~MS clients with information that can be used to determine the relative importanceof workingto resolve interactions. Froma scheduling perspective, this meansreasoningabout the benefits of leveraging task interactions relative to other constraints and to alternative ways of achieving the overall task. Froma coordination perspective, agents can reason about the benefts of coordinating interactions with other agents versus the costs of such coordination. Wewill return to the issue of coordinationshortly. Learningis also implicit in T/EMS.The probability distributions associated with primitive actions can be learned a priori or be refined over time via learning. Domainindependence is also an important feature of T,’EMS.The expense of customdesigning agent componentsis serving to inhibit widespread agent development; moving toward a flexible and adaptable domainindependent agent control system is critical for the future of agent systems.As wediscuss in the architecture section, integrating domainindependentagent control componentsis one of our primary near-term objec- Design-to-Criteria Scheduling The scheduler is the heart of satisficing agent control. An agent equipped with a Design-to-Criteria scheduler is able to adjust its problemsolving activities to meet changesin resource requirements and to reason about the trade-offs of different courses of action. The central objective in Designto-Criteria scheduling is to cope with the combinatorialexplosion of possibilities while reasoning about a particular set of client goal criteria and the trade-offs presented by different solutions and partial solutions. In other words, schedulerclient applications or users specify the designcriteria and the scheduler designs a schedule to best meet the criteria, if possible given the task model. Figure 3 shows a set of satisficing schedules producedby the Design-toCriteria scheduler, for the sampletask structure (Figure 1), using four different sets of design criteria. ScheduleA is constructed for a client interested in a fast, free, solution with any non-zero quality. Schedule B suits a conservative client whois interested primarily in certainty about quality achievement.Schedule C is designed for a client whowants high quality information, is willing to wait a long time for 82 ScheduleA: Fast and Free [Edmund’s-Reviews[ Edmund’s-Price-Guide I Q (~(14, 0)(5% IOX2% 12)(934, c (l(X)% D (254, 240)(25,#, 2~)X3 I% 260)( 124,270)(6% ExpectedQ: 21 Q Certainty: 93% ExpectedC: 0 C Certainty: 100% ExpectedD: 255 scconds D Certainty: 50% SchcduleC: G~x,dQuality, M~vJerate Cost, Slow IEdmund’s-ReviewslHeraud’s-Te.st-Drives I lntelichoice] Q (~04, 17X20% 27)(2%34)(78% C (I(K)4, $4.95) D (20% 840X 19’,~, 9lX))(314, 920)( 19%980X I 1% ExpectedQ: 40 Q Certainty: 78% ExpectedC: $4.95 C Certainty: 1(10% ExpectedD: 920 seconds D Certainty: 70% ScheduleB: HiJ/h QualityCertainty,MtnlerateCost lEdmund’s.Reviewsilntelichoice ] Q (24,17)(98% c (100%$4.95) D (25% 6(X))( 124, 620)(3 I%fiX())( ExpectedQ: 26 Q Certainty: 98% ExpectedC: .$4.95 C Certainty: 100% Expected D: 647 ~conds D Certainty: 50% ScheduleD: HighQuality, HighCost, ModerateDuration [Edmund’s-Revicws]Heraud’s-Test-Drives]Get.AutoSite.URLllssue.AutoSite.Request ] Q (1%0)(44, 27)(19%34)(24, 41)(74% C (1004, $9.95) D (20%630)(31% 690)(24% 72(l)(19% 740)(6% ExpectedQ: 46 Q Certainty: 74% ExpectedC: $9.95 C Certainty: I(X)% ExpectedD: 698 seconds D Certainty:514, Figure 3: FourSatisficing Schedules an answer, and is only willing to spend $5 on the search. Schedule D is meets the criteria of a client whowants the highest possible quality, is willing to spend $10, and wants the gathered data in 15 minutesor less. The fundamental premise of our scheduling work is that the goodnessof a particular solution is entirely dependent on a particular client’s complexobjectives and that different client’s have varying objectives. Thusthe scheduling process must not only consider the attribute trade-offs of different solutions, but must also do so dynamically.Furthermore, the scheduling process must be efficient - because of the inherent uncertainty in the domains, where actions mayfail or have unexpectedresults, scheduling activities are interleaved with planning and execution. Thus scheduler inefficiencies are multiplied manytimes during a problem solving instance. In general the upper-bound on the number of possible schedules for a task structure containing m methods is ~--~ir=~ o (r~)i! and the w(2m) and o(mTM) combinatorics of our scheduling problemprecludes using exhaustive search techniques for finding optimal schedules. Design-toCriteria copes with these explosive combinatoricsby satisricing with respect to the goal criteria and with respect to searching the solution space. This satisficing dualismtranslates into four different techniquesthat Design-to-Criteria uses to reduce the search space and makethe scheduling problemtractable: mitmentsmadewith other agents and so forth. Heuristic Error Correction The use of approximation and heuristic decision makinghas a price - it is possible to create schedules that do not achieve the high-level task, or, achieve it poorly. A secondary set of improvementheuristics act as a safety net to catch the errors that maybe correctable. Design-to-Criteria thus copes with computational complexity by using the client goal criteria to focus processing, reasoning with schedule approximations rather than complete schedules, and using a heuristic, rather than exhaustive, scheduling approach. This methodologyis effective because several aspects of the scheduling problemare soft and amenableto a satisficing approach. For example, portions of the client goal specification (Wagner, Garvey, Lesser 1997a) express soft client objectives or soft constraints. Solutions often do not need to meet absolute requirements because clients cannot knowa priori what types of solutions are possible for a given task structure due to the combinatorics.Similarly, soft task interactions also represent soft constraints that can be relaxed, i.e., they can be leveraged or not dependingon the situation. Finally, though the TJ~MSscheduling problem is more complex than many traditional scheduling problemsbecause of its representation of multiple approachesfor task achievement,it is also moreflexible. If we viewthe schedulingactivity as a search process, typically there is a neighborhoodof solutions that will meet the client’s goal criteria and the lack of exhaustive search, i.e., search by focused processing and approximation, does not necessitate scheduling failure. Several illustrations of Design-to-Criteria at work, and morealgorithmic details, can be found in (Wagner,Garvey, &Lesser 1998). Criteria-Directed FocusingThe client’s goal criteria is not simply used to select the "best" schedule for execution, but is also leveragedto focus all processingactivities on producingsolutions and partial solutions that are most likely to meet the trade-offs and limits/thresholds definedby the criteria. ApproximationSchedule approximations, called alternatives, are used to provide an inexpensive, but coarse, overviewof the schedule solution space. Heuristic Decision MakingThe action ordering scheduling problem also suffers from large combinatorics. We cope with this complexityusing a group of heuristics for action ordering. The heuristics take into consideration task interactions, deadlines, resource constraints, com- GPGPMulti-Agent Coordination Satisficing in the GPGP(generalized partial global planning) multi-agent coordination workto date takes a different from than satisficing in Design-to-Criteria scheduling. In coordination, flexibility with respect to computational resources is derived from a modularizedcoordination approach that enables different modules,with different over- 83 head costs, to be applied independently from one another. Whenresources are tight, only the most rudimentary coordination is used. Whenresources are less scarce, or the beneft of coordinating outweighs the cost of the coordination overhead, agents coordinate over "optional" interactions like soft interactions in T/EMS.The GPGPmodularized coordination approachfacilities a range of cost/benefit options for any problemsolving episode as different subsets of coordination modulesmaybe applied. All coordination modules in GPGPcoordinate through the use of commitments, that is inter-agent contracts to perform certain tasks by certain times. Commitments are made through the coordination mechanismsand considered by the scheduler whenthe local course of action is decided. GPGPdefines the following coordination mechanisms(for the formal details see (Decker1995)): Share Non-LocalViews This most basic coordination mechanismhandles the exchangeof local views between agents and the detection of task interactions. CommunicateResults This coordination mechanismhandles communicatingthe results of methodexecution to other agents, at various levels of detail and coverageas defined by different communication policies. Avoid Redundancy This mechanism deals with detected redundancyby committingan agent at randomto execute the redundant methodin question. Handle Hard Task Relationships - The enables NLE pictured in Figure 1 denotes a hard task relationship. This coordination mechanismdeals with such hard, nonoptional, task interactions by committingthe predecessors of the enables to performthe task by a certain deadline. Handle Soft Task Relationships Soft task interactions, unlike hard interactions like enables, are optional. When employed,this coordination mechanismattempts to form commitments on the predecessors of the soft interactions to performthe methodsin question by a certain deadline. As mentioned above, the GPGPcoordination module modulates local control by placing constraints, commitments, on the local scheduler. The commitmentsgenerally fall into three categories: 1) deadline commitments are contracts to performworkby a certain deadline, 2) earliest start time commitmentsare agreements to hold-off on performing certain tasks until a certain time has passed, and 3) do commitmentsare agreements to perform certain tasks without a particular time constraint. GPGPachieves flexibility through a modularized approach to coordination. The question then becomeswhich combinations of mechanismsare most effective. Empirical analysis (Decker 1995) has shownthat no single coordination mechanism,or sets of mechanisms,is most effective for all situations. Instead, as shownin Prassad and Lesser 84 (Prasad & Lesser 1996), coordination is best viewedas situation specific activity; whichset of mechanisms is most effective is dependenton the types of tasks in question and the overhead costs of communications.In different environments, different mechanismsare most effective. The Future: A Holistic Satisficing Architecture Agent Both GPGPand Design-to-Criteria are able to adapt processing to a given situation. GPGP,uses a modularizedapproach to provide different degrees of coordination, each with its ownoverhead costs and benefits to problemsolving. Design-to-Criteria uses a satisficing methodologyto control the combinatoricsand to produceschedules in realtime. This illustrates an important point; satisficing can take manyforms, e.g., performingless search (Design-toCriteria), or using less context to makedecisions (GPGP), or working from default assumptions (situation specific GPGP).In fact, the implementationof GPGP is itself an illustration of satisficing by sacrificing context. GPGP forms commitmentsbetween agents through a one-shot mechanism where two agents form an agreement about task performanceindependently of the other agents in the system. This mechanismreduces the overhead required for coordination at the expense of making the commitmentswithout understanding more of the group context. In contrast, if the commitmentforming context is broadened to include a groupof problemsolving agents, to support a multi-step negotiation process, the overhead of commitmentformulation will increase but so will the amountof information used to make commitmentdecisions. In addition to meetingdeadlines, Design-to-Criteria also builds customschedulesto suit a particular client’s multidimensional goal criteria. This latter Design-to-Criteria feature represents one next step in the evolution of GPGP and our multi-agent coordination work. Currently, though GPGPprovides a modularized toolbox of coordination mechanisms, it does not reason about which mechanisms to use to meeta particular client’s needs(or goal criteria). The workin learning which coordination algorithms to use in particular situations is a step in the right direction, but this work also does not address meeting dynamic multidimensional goal criteria. Like Design-to-Criteria, GPGP mustbe able to adapt its processingautomaticallyto a given problemsolving context and resource constraints. Despite Design-to-Criteria’s strengths, neither aspect of our agent control work is complete. Neither system accounts for the actual cost of the control problemsolving activity. In other words, thoughthe scheduler copeswith hard combinatorics to produce schedules in interactive time, it does not account for the resource costs of the scheduling processitself, i.e., Design-to-Criteriais fast, but it doesn’t reason about the time required to find a good schedule and accountfor this time, and adjust its activities accordingly. Throughthe focusing mechanismit can adjust its ownproblem solving to different resource requirements, but right now,the focus is definedby the client or hardwireddefaults. The right approach for coordination and scheduling, and overall agent control, is to accountfor the both the control and domainproblemsolving costs, to reason about resource allocations and the cost/benefits of such allocations, and to control the agent accordingly. The benefits of spending time doing further coordination or morerefined scheduling versus time spent problem solving must all be compared and reasoned about. This meta analysis (Russell & Wefald 1991) approachto agent control in conjunctionwith the satisficing abilities of Design-to-Criteriaand GPGP, will result in agents beingable to satisfice all activities and adapt to a wide range of dynamicproblem solving environments. However,this is an oversimplification of the problem.It is importantto note that the issue is not to simplydetermine a resource allocation to control and domainproblemsolving, but to grapple with the downstream ramifications of the allocations and satisficing behaviors employedor considered. Control problemsolving and domainproblemsolving are interdependentactivities. Considera case wheretime is allocated betweena domainproblemsolver and a scheduler, and the domainproblemsolver faithfully describes its problem solving options in a T/EMS task structure and then asks the scheduler to determinea course of action. If the scheduler satisfices, to meetits time resourceconstraint, it will do so by exploring a smaller part of the schedule solution space. Onepossible outcomeof this is that the scheduler will fail to generate a very goodsolution and will instead generate and return a mediocresolution. The domainproblem solver will then execute the schedule and obtain lower quality results than it mighthaveif less time had beenallocated to domain problem solving and more time had been allocated to scheduling. In this example, the downstream effect of control satisficing wasnot generatinga goodresult in the time alloted. In other cases, the ramifications range fromthe introduction of additional uncertainty to being unable to find any solution because one of the components was over constrained. The previous exampleis cast in terms of time allocations, but as mentionedearlier, satisficing can be in manydimensions like precision, robustness, and cost, and it can be in terms of multiple dimensions simultaneously. This is what motivated our recent workin multi-dimensional goal criteria for Design-to-Criteria scheduling (Wagner, Garvey, Lesser 1997a). But even that work needs to be pushed to another level - satisficing to meeta particular set of multidimensional goal criteria is only part of the problem. In most problemsolving situations, there are actually multiple sets of goal criteria, i.e., there is no single evaluation function. Considera case wherea client tells the scheduler 85 to producea schedule that will achieve the task while minimizing time and incurring no cost. If, for the given task structure, staying within these constraints meansgenerating a very poor result, the client maywant to modifythe constraints and ask for another schedule, rather than execute the schedulethat was returned for the first set of goal criteria. Agent control components,and the holistic agent control mechanism,must reason about multiple sets of goal criteria, or provide a meansfor negotiation or iterative refinementto generate a set of criteria that is suited for the problemsolving context and the task at hand. Giventhe complexissues surrounding the issue of satisricing, our goal of a satisficing domain-independent agent control architecture seems quite ambitious. The satisficing cohesion neededfor this architecture will require that all componentshave the ability to meta analyze their problem solving activities and to estimate the local (as everything is intertwined) costs/benefits of resource allocations. Fromthe scheduling and coordination perspective, this is an achievable objective. For the domainproblem solver, meeting this requirement is application dependent. However, to the extent that the domainproblemsolver can enumerate its problemsolving options as a T/EMStask structure, the scheduler can then reason about the domainproblemsolver’s activities explicitly. In fact, we maymovetowardrepresenting all major control and domainactivities in T/EMStask structures, and even estimating and modeling (via nles) the downstreaminteractions of resource allocations, and analyzing it using the Design-to-Criteria scheduler. If this approachis used, wemust then also accountfor the time required to perform the meta-analysis. However, for certain classes of task structures the schedulinganalysis time is highly predictable. For large problemsolving tasks, we also plan to explore changing the problemsolving horizon, the distance betweenthe current point in time and a point at whichallocations will be reexamined. Satisficing, flexibility, and multi-dimensionaltrade-off behavior, are critical for agents in complexdynamicopen environments. Our future work is centered on a holistic, domainindependent, agent control architecture designed to achieve these goals. It is our hopethat with the architecture agent developers will be able wrap or embedtheir domain problemsolvers and create satisficing and flexible multiagent systems. Acknowledgments Wewouldlike to thank Professor Keith Decker, Professor Alan Garvey, Professor NormanCarver, and Dr. Nagendra Prassad, for their contributions to this ongoingwork. References Carver, N., and Lesser, V. 1995. The DRESUN testbed for research in FA/Cdistributed situation assessment: Exten- sions to the modelof external evidence. In Proceedingsof the First International Conferenceon MultiagentSystems. Decker, K. S., and Lesser, V. R. 1995. Coordinationassistance for mixed humanand computational agent systems. In Proceedings of Concurrent Engineering 95, 337-348. McLean,VA: Concurrent Technologies Corp. Also available as UMASS CS TR-95-31. Decker, K. S. 1995. Environment Centered Analysis and Design of Coordination Mechanisms.Ph.D. Dissertation, University of Massachusetts. Garvey, A., and Lesser, V. 1995. Representing and scheduling satisficing tasks. In Natarajan, S., ed., Imprecise and Approximate Computation. Norwell, MA: Kiuwer AcademicPublishers. 23-34. Lesser, V.; Horling, B.; Kiassner, F.; Raja, A.; Wagner,T.; and Zhang, S. X. 1997. Information Gathering as a Resource BoundedInterpretation Task. UMASS Department of ComputerScience Technical Report TR-97-34. Prasad, M. N., and Lesser, V. 1996. Learning situationspecific coordination in generalized partial global planning. In AAAI Spring Symposium on Adaptation, Coevolution and Learning in Multiagent Systems. Russell, S., and Wefald, E. 1991. Principles of metareasoning. Artificial Intelligence 49. Vincent, R.; Horling, B.; Wagner,T.; and Lesser, V. 1998. Survivability simulator for multi-agent adaptive coordination. In Proceedingsof the First International Conference on Web-BasedModeling and Simulation. To appear. Also available as UMASS CS TR-1997-60. Wagner, T.; Garvey, A.; and Lesser, V. 1997a. Complex Goal Criteria and Its Application in Design-to-Criteria Scheduling. In Proceedings of the Fourteenth National Conference on Artificial Intelligence, 294-301. Also available as UMASS CS TR- 1997-10. Wagner,T.; Garvey, A.; and Lesser, V. 1997b. Leveraging Uncertainty in Design-to-Criteria Scheduling. UMASS Department of Computer Science Technical Report TR97-11. Wagner,T.; Garvey, A.; and Lesser, V. 1998. CriteriaDirected Heuristic Task Scheduling. International Journal of ApproximateReasoning, Special Issue on Scheduling. To appear. Also available as UMASS CS TR-97-59. Zilberstein, S. 1996. Usinganytimealgorithmsin intelligent systems. AI Magazine17(3):73-83. 86