On the Convergence Error in Loopy Propagation

Reasoning with a Bayesian network amounts to computing probability distributions for the network’s variables. Pearl’s propagation algorithm provides for efficient reasoning with singly connected networks. When applied to a multiply connected network, the algorithm no longer yields exact probabilities, yet has been reported to result in good approximations. In this paper we provide some theoretical background to the algorithm’s performance on multiply connected networks. We indicate the two types of error that can arise in the probabilities computed by the algorithm and analyse the different factors that govern one of these errors.