Decisiveness in loopy propagation

Abstract

When Pearl's algorithm for reasoning with singly connected Bayesian networks is applied to a network with loops, the algorithm is no longer guaranteed to result in exact probabilities. We identify the two types of error that can arise in the probabilities yielded by the algorithm: the cycling error and the convergence error. We then focus on the cycling error and analyse its effect on the decisiveness of the approximations that are computed for the inner nodes of simple loops. More specifically, we detail the factors that induce the cycling error to push the exact probabilities towards over- or underconfident approximations. © 2007 Springer.