Conditioned Belief Propagation revisited (extended version)

Abstract

Belief Propagation (BP) applied to cyclic problems is a well known approximate inference scheme for probabilistic graphical models. To improve its accuracy, Conditioned Belief Propagation (CBP) has been proposed, which splits a problem into subproblems by conditioning on variables, applies BP to subproblems, and merges the results to produce an answer to the original problem. In this work, we propose a reformulated version of CBP that exhibits anytime behavior and allows for more specific tuning by formalizing a further aspect of the algorithm through the use of a leaf selection heuristic. We propose several simple and easy to compute heuristics and demonstrate their performance using an empirical evaluation on randomly generated problems.