Efficient Distributed Bayesian Reasoning via Targeted Instantiation of Variables

This paper is focusing on exact Bayesian reasoning in systems of agents, which represent weakly coupled processing modules supporting collaborative inference through message passing. By using the theory on factor graphs and cluster graphs we (i) analyze the suitability of the existing approaches to modular inference with respect to a relevant class of domains and (ii) derive methods for construction of modular systems, which support globally coherent Bayesian inference without compilation of secondary probabilistic structures spanning multiple modules. In the proposed approach dependencies between inference modules are reduced through targeted instantiation of variables, which is based on the analysis of cluster graphs.