Distributed multi-agent probabilistic reasoning with Bayesian networks

Abstract

Main stream approaches in distributed artificial intelligence (DAI) are essentially logic-based. Little has been reported to explore probabilistic approach in DAI. On the other hand, Bayesian networks have been applied to many AI tasks that require reasoning under uncertainty. However, as commonly applied, a single-agent paradigm is assumed in Bayesian networks.

This paper extends multiply sectioned Bayesian networks (MSBNs) for single-agent systems into a framework for multi-agent distributed interpretation systems. Each cooperative agent is represented as a Bayesian subnet that consumes its own computational resource, gathers its own evidence, and can answer queries. Unlike in single-agent systems where evidence is entered one subnet at a time, multiple agents may acquire evidence in parallel. We add to the set of single-agent MSBN operations new belief propagation operations which, when performed, regain global consistency. Due to the inter-agent ’distance’ and the associated communication cost, global consistency can not be maintained constantly. We show that, when the proposed operations are followed, between two successive communications, the answers to queries from an agent are consistent with all local evidence gathered so far, and are consistent with all global evidence gathered up to the last communication.