47 From: AAAI Technical Report WS-94-04. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved. Hierarchical Conflict Management David Goldstein North Carolina Agricultural & Technical State University Greeensboro, NC27411 USA Voice: (910) 334-7245 Fax: (910) 334-7244 goldstn@ncat.edu Keywords: Real-time reasoning, Assumption-based reasoning, Hierarchical, Conflict resolution Abstract conflicting constraints in a hierarchy of Twodistinct concerns seem apparent in decision makers. conflict management strategies, one reflecting real-time decision-makingand Introduction the other constrained deliberation. This We have been examining paper attempts to provide more detail resolution concerning our experiments in this area chaining systemsat the rule level and (2) and provide a more unified conflict managementwith respect to the model of (1) in real-time conflict forward- conflict resolution as the management of hypotheses proposed by interacting resources in constrained environments. agents. While these two areas initially This paper is structured in the seem disparate, they both consider following manner. First, we describe the choosing between multiple alternatives scope of the problemswhich interest us. based upon constraints. Wethen describe moreinformation added to either system, our approaches to address these problems. Finally, describe the implementation we of our the better the methods should be to perform. This paper considers, research and our goal of developing macro level, mechanisms applicable concepts to resolving Further, the the questions: at a (1) what are common to these two approaches and (2). commonalties lead these can to a mechanism for addressing helping to meet an application’s general deadlines. Morespecifically, we develop conflict a partial ordering of actions reflecting a partial ordering of rule priorities. resolution in a hierarchical manner. We then use these relative rule priorities, Our Approaches along with resource dependencies, to Our initial develop time-varying priority functions. time research stemmedfrom real- knowledge-based Initially, systems. we attempted to devise clever These functions are then converted into spline functions for efficient computation mechanismsfor "conflict resolution" to (Figure 1). We have modified our determine the most appropriate chain of inferencing mechanismsto efficiently reasoning to pursue amongthe multiple computethese priorities at run-time[ 1 ]. We examine employing such a chains currently available to a forwardchaining system. Morespecifically, given conflict n satisfied rules in the "conflict set" of a considering the generation of plausible forward-chaining reasoning system, how courses of action and selecting one such might we choose the most appropriate course prior to some time t. Wediscuss rule for execution. Commonly used determining the parameters driving the recency and dynamicpriorities in [1 ]. Weintend to approaches, including resolution algorithm means-endsanalysis, essentially ignore eventually the real-time characteristics of problems. reasoning must take place (measured by Our approach individual agenda items (i.e., how much counting and weighing the tokens in the to conflict resolution modifies the priorities also consider by’ of rules and RETE network) priorities. to modify the rule By tying the contents of the their matching data elements) during RETEnetwork to the rule priorities, each reasoning cycle. Weprioritize rules hope to be able to extend an approach based upon the likelihood of the actions we 49 such as that in [2] to predict near-future management); agents can reason about behavior of a system. other agents if their goals require such knowledge,but if an agent G’s goals are The above approach seemingly addresses only the lowest levels conflict disjoint of from other agents then processing can proceed (with literally management in a system. Therefore, we now address developing hundreds of agents being dynamically mechanisms for conflict created or destroyed) without any other management agent having knowledge of G. While amongmultiple, interacting agents. The mechanisms numerous conflict we have previously used for conflict management have been were rather autocratic negotiation, [3]. However, management models proposed reflecting stock markets, our current experiments in Distributed playing, Artificial exploring is based upon transferring Intelligence deal with the etc., game- the approach we are extremeof distributed control (and so we procedural knowledge and assumptions are devising newstrategies for conflict amongagents. Scenario-Specified Task Criticality Priority / ’~ Resource Estimation (obtainable from Directed Acyclic Graph) ~ P(Task will Complete) Output Task ~cheduled Useful Output Required Output Critical Task Executed or No Longer InstantaneousPriority (Salience) Figure 1: Prioritizing a Rule via a Spline Function 50 Weview agents as a collection of given that certain constraints must be engaged in a dialogue in a met for this knowledgeto be executed, entities canonical language. contradictory Periodically, statements (plans) are our conflict management mechanism removes the low-priority stated (created). Based upon the actual from world, at least one statement fromthe set corresponding of all constraining assumptions. Werepeat this statements involved in the consideration knowledge and the constraints and conflicting set of statements must be process until no conflicts are present. retracted. Weuse the assumptions to Wethen choose the remaining statements determine which piece of the procedural (decide in favor of the remainingagent’s knowledgeshould be invalidated first. courses of action.) Given a set of resources with some value prioritized per unit and a set of goals, procedural knowledge Weview removing assumptions as taking one of two forms, either (1) literally discarding the hypotheses related to the low priority has a cost associated with achieving each assumptionsor (2) generating alternative specific goal. Eachgoal similarly has an hypotheses via employing domain- associated, independent mechanisms to achieve computable benefit. The difference between these two quantities alternative can be viewedas a net gain in potential have not yet implementedour library to resources for having achieved a goal generate alternatives (note: one should not be able to utilize proposed plans, we hope to develop a negative resources domain-independent rule-base nor do we allow implicit borrowingof resources.) The net potential resource gains form an ordering of procedural knowledge to be executed. However, plans as per [4]. While we to inconsistent, to map domain-specific concepts and classes of objects to generally techniques. applicable 51 These two conflict The underlying resolution meta-models strategies mayseemdisparate initially, (semantics) of data items is an important but they do have some commonroots; consideration. they both examine resource conflicts the Product Data ExchangeSpecification using rule-based (PDES) for our meta-models [7]. develop techniques. As we Weuse an extension of Our meta-modelsexhibit our philosophy that our domain independent techniques for conflict resolution, we basic concepts hope to better understand the conflict application in a system must first resolution described. process so as to perform understood by each be Because PDES includes agenda managementin a more efficient representations of quantities (such as manner. Further, we hope to eventually force, be able to developa hierarchy of conflict concepts (such as behaviors, phenomena, management schemes from the inner- and physical objects) can be represented. distance, and hue) diverse agent to society-wide levels such that one can model the resolution processes Conclusion using similar mathematicsat each level. We have presented two conflict management schemes operating at Status different levels in a hierarchy of conflict The agent architecture has been realized resolution. as the Distributed Artificial Intelligence more rigorously in the future to better Toolkit understandthe essential characteristics of (DAIT), a series of tools including an enhanced version NASA’s C Language Integrated Production of System (CLIPS) [5][6]. conflict Weintend to model these managementin a hierarchy of decision makers. Weintend to use these schemes to develop domain-independent Agents can communicate via facts conflict resolution mechanismswhich can (mathematical templates be employedat any level in the hierarchy tuples), (database records) and objects. 52 by mapping general mechanisms Expert, Miller-Freeman Publishing, January, 1994, pp 3034. tO domain-specific objects and concepts. [7] References [1] D. Goldstein, "A Fault-tolerant Agent-based Architecture for Manufacturing", Concurrent Engineering: Research and Applications, to appear 1994. [2] F. Barachini and H. Mistelberger, "Run-Time Prediction for Production Systems", in Proceedings of the Tenth National Conferenceon Artificial Intelligence, American Association for Artificial Intelligence, Menlo Park, CA, July 1992, pp. 478 - 485. [3] D. Goldstein and J. Tiernan, "Domain-independent Conflict Managementvia Agent-supplied Metrics", 1993 IJCAI Workshop on Conflict Management, International Joint Conferenceon Artificial Intelligence, France, September 1993. [4] M. Klein, "Supporting Conflict Resolution in Cooperative Design Systems", in Proceedings of the 10th International Workshopon DistributedArtificial Intelligence, Bandera, TX, October 1990. [5] J. Giarrantano, "CLIPS User’s Guide", NASA/Johnson Space Center, Houston, USA,1991 [6] D. Goldstein, "The Distributed Artificial Intelligence Toolkit", AI International Standards Organization, Product Data ExchangeSpecification, Part 32, National Institute of Standards and Technology.