Iterative Compilation of Multiagent Probabilistic Graphical Models

Abstract

Multiply sectioned Bayesian networks (MSBNs) support multiagent probabilistic inference in distributed large problem domains. Inference in MSBNs can be performed effectively using their compiled representations. The compilation involves cooperative moralization and triangulation of the set of local graphical structures that collectively defines the dependencies among domain variables. Privacy of agents prevents us from compiling MSBNs by first assembling graphical subnets at a central location and then compiling their union. In earlier work, agents perform compilation in a limited parallel via a depth-first traversal of the local structures organized in a tree structure (called hypertree). Agents need some synchronization with each other. In this paper, we present an iterative method, by which multiple agents compile MSBNs asynchronously. Compared to the traversal method, the iterative one is self-adaptive and robust.