A Probabilistic Framework for Cooperative Multi-Agent Distributed Interpretation and Optimization of Communication

Multiply sectioned Bayesian networks for single-agent systems are extended into a framework for cooperative multi-agent distributed interpretation systems. Each agent is represented as a Bayesian subnet. We show that the semantics of the joint probability distribution of such a system is well defined under reasonable conditions. Unlike in single-agent systems where evidence is entered one subnet at a time, multiple agents may acquire evidence asynchronously in parallel. New communication operations are thus proposed to maintain global consistency. It may not be practical to maintain such consistency constantly due to the inter-agent ‘distance’. We show that, if the new operations are followed, between two successive communications, answers to queries from an agent are consistent with all local evidence, and are consistent with all global evidence gathered up to the last communication. During a communication operation, each agent is not available to process evidence for a period of time (called off-line time). Two criteria for the minimization of the off-line time, which may commonly be used, are considered. We derive, under each criterion, the optimal schedules when the communication is initiated from an arbitrarily selected agent.