Cooperative triangulation in MSBNs without revealing subnet structures

Multiply sectioned Bayesian networks (MSBNs) provide a coherent framework for probabillstic inference in a cooperative multilagent distributed interpretation system. Inference in MSBNs can be performed effectively using a compiled representation. The compilation involves the triangulation of the collective dependency structure (a graph) defined in terms of the union of a set of local dependency structures (a set of graphs). Privacy of agents eliminates the option to assemble these graphs at a central location and to triangulate their union. Earlier work solved distributed triangulation in a restricted case. The method is conceptually complex and the correctness of its extension to the general case is difficult to justify. In this paper, we present a new method that is conceptually simpler and is efficient. We prove its correctness in the general case and demonstrate its performance experimentally.