Centralized Task Distribution in the Presence of Uncertainty and Time Deadlines

This paper focuses on de’adlim.-ha~d nmLtiagent task decomposition/scheduling of itierarchical task structures in an environment in which one agent is responsible for assigning subtasks and their execution order to a set of agents. The overMI task to be decomposed is specified in aa extended version of "I’AEMS task modeling framework in which task quality and execution time are uncertain before execution and are represented in the form of a probabilily distribution. The master agent generates and assigns an initim subtask sched..le to aot’.ls, and it’ necessary will revise ~gent schedules whvnev(.r agents ntake rescheduliug requests. These requests are based on the degree of deviation t.inal Ina.~ occurred in expected quality anti duration as a result of at’tuaI method execution. ]"valualiolns of scheduler performance indicate a phase transilion of reschedalivg efficiency with rt’spect to probh:m difficulty. Specifically, while frequent rescheduling was ineffective for both wry ea.~y and very difl:icult problems, it was extrcmvly effective in the transition rt’gi()rt (mid-range) of difficulty. Introduction This work focuses on the problem of dcadlineha~d mull,iagent task decomposition and scheduling (MADS) of medium grain l,roMem-solving activities in an environment in which one (master) agent is responsible for assigning sul)tasks and their execution order to a set of (slave) agents. ’The master agent decides not only how to partition subl.asks anloltg the slave agents but also how marly agents are required in order to meet the perfortnancp goals associated with the high-level trek. The master agent is also responsible for rescheduling the slave agout.s wlnen tilt’ master agent is notified that agent cxet’ul ion of activities does llOt conforlll to a priori expectations, anti integrating tim completed partial result of agents into a overall solution to the lfigh-level ttLsk. Slave agents, wherc necessary: transmit intermediate results to other slave agents so as to enable or facilitate execution of their activities. Figure I shows th,’ interaction among the master and slave agents. ’fhis approach to task decomposition is obviously different from a negotiation model of task decomposilion such as the contract net approach (Smith 1980} where the task executing agents haw; more autonomy in deciding what activities they will execute. It also differs from that approach bccause part of an agent’s task assignment may require tile agent to transmit intermediate results not to the contracting agent but rather to another agent before a specific time. This possibility was foreseen in a later paper by Davis and Smith (Davis & Smith 1983) in which, once a task decontposition is estM*lished there can be a "result shariLtg" phase, but to our knowledge this result sharing arid its tinting were never a consideration in any of the work on act ual implement ations of t he contract, net algorithm. Another approach to task decomposition is tile one developed by Durfet, and Montgomery (Durfee & Montogomery 199 l ). ’rlney represent the overall task using a hierarchical behavior lnodel and are concerned with finding appropriate abstractions in the behavior space that motivate the decompLxsitkm process. Our model of task decomposition is prohahly closest to the work by l)urfee and Lesser on the implementation of hierarchical organizational structures in PGP (Durfee & Lesser 1991). liowever, in that work, the decomposition of activities to agents is given a priori and the focus is on how to appropriately schedule the activities and to whom intermediate results should be sent. In our work, we not only do that but also decide on which activities to assign to specific age.nts and determine how many agents are required, t ~A~’ also base these decisions on a more complex anti doruaiil-independem, model of a task structure than the one used in l.heir work. A key aspect, of this model is the explicit declaration of the expected distribution of the time for subtask completion and of the quality achieved. This information together with time deadaAnotiwr way of viewing this work is that the MADS "algorithm is another coordination module that can he incorporated into the (;PGP arcititecture (Decker &. Lesser 1995a). From that perspective, it extends the range of agent organizational structures that can be constructed using GPG P. Fujita 87 From: Proceedings of the Second International Conference on Multiagent Systems. Copyright © 1996, AAAI (www.aaai.org). All rights reserved.