A New Framework for Inference in Distributed Bayesian Networks for Multi-Agent Sensor Interpretation

Multi-agent systems (MAS) are groups of interacting intelligent software agents. An important application is sensor interpretation (SI) in sensor networks. SI domains are frequently modeled with Bayesian networks (BNs), and distributed versions of these problems can be modeled with distributed Bayesian networks (DBNs). The multiply sectioned Bayesian network (MSBN) framework is the most studied approach for inference in DBNs, in an MAS setting. However, we do not believe the MSBN framework is well suited for large-scale MAS-based SI. This paper describes an alternative framework for inference in DBNs that we have developed to support efficient, approximate MAS-based SI. Compared to the MSBN approach, our approach supports more autonomous and asynchronous agents, and more focused, situation-specific communication patterns. Our analyses show that this framework can be used to produce acceptable interpretations at substantially lower cost than the MSBN.