From: Proceedings of the First International Conference on Multiagent Systems. Copyright © 1995, AAAI (www.aaai.org). All rights reserved. Distributed Symbolic-SubSymbolic Agent Architecture Configuring Power Network Faults for R. Khosla T. Dillon Expert and Intelligent SystemsLaboratory, Applied ComputingResearch Institute, and Department of ComputerScience and ComputerEngg., La Trobe University, Melbourne, Victoria, Australia e-mail: khosla@laicsl.laf.oz.au Tel: -I-61.3.4793034, Faz: +61.3.4704915 Extended Abstract The distributed computational level architecture is derived through integration of task structure level Intelligent Agentsamongother aspects represent task architecture with an object-oriented model and an unix operating system process model. The compuabstraction. Tasks and methods (used to accomplish the tasks) can be considered as mediating concepts tational level architecture consists of preprocessing in problem solving. The symbolic and ¢onnectionist and decompositionagents at the global level and conmethodsbecauseof their ditferent philosophical, cogtrol, decision, and postprocessing agents at the local level. At the computational level we determine nitive and computational underpinnings impose their the knowledgecontent of the symbolic-connectionist ownconstraints on the quality (e.g time, resources, no. agents. The symbolic-connectionist knowledgecontent of levels of abstraction/solution hierarchy, reliability, etc.) of task accomplishment. Further, these conincludes amongother aspects the planning knowledge, straints can lead to task generation dependingupon belief knowledge,a priori learning knowledge (for Artiwhichglobal constraints have been violated by a certidal Neural Networks), execution knowledge,commutain method,and the degreeof tolerance of the violated nication knowledge,and validation (in terms of time constraints. A proper integration of symbolicand conand context) knowledgefor preprocessing, decomposinectionist methodsfor task accomplishmentcan thus tion, control, decision, and postprocessingagents. improve the quality of task accomplishmentand reThe knowledgecontent in general represents three strict task generation. Moreso, such a integration can forms of knowledge, i.e. symbolic and inforprovide moreflexibility to an agent in terms of multims] (conceptual and henristical), sub-symbolic(subplicity of methodsfor accomplishinga task. conceptual), and symbolic and formal (logic, mathemstical models) knowledge. The integration of In this direction our work involves developmentof the task structure level architecture with an objectmulti-agent symbolie-subsymbolicarchitecture at the oriented model and unix operating system process task structure level, computationallevel, and progrsm modelalso enables us to satisfy constraints like dislevel for complexknowledgeintensive domains. tributed processing, inheritance, knowledgesharing Thearchitecture at the three different levels is based and reuse. The computational architecture has been on pragmatic constraints like cognitive compatibility successfully used for developinga real-time distribu (in terms of information processing),neurobiologlcal ted, alarm processing system for configuring network considerations, forms of knowledge,problemcomplexfaults in a powersystem control centre with encouragity, search time, temporalreasoning, learning, incoring performanceresults. The objectives of the alarm rect and incompleteinformation, reliability, generalprocessing system are a) isolate the event, b) isolate ization, distributed processing, global and local parthe cause of the event, and c) isolate the cause of the allelism, global and local competition, cooperation, event in four seconds for simple faults and maximum of knowledgeshying/reuse, inheritance, scalability, and thirty secondsfor severe faults. Real time alarm data cost effectiveness. from the powersystem control centre has been used in Thetask structure level architecture is a product of the alarm processing system. the tasks, constraints (i.e. global andlocal constraints) on the tasks and the methods used to accomplish the tasks. Thetasks are defined within five phases of information processing, namely, preproceesingphase, decompositionphase, control phase, decision phase, and postprocessing phase. The division of responsibility between symbolic and connectionist methodsis based on the satisfaction of various task constraints. Khosla 451