Approximate Expectation Maximization

Abstract

We discuss the integration of the expectation-m axim ization (EM) algorithm for maximum likelihood learning of Bayesian networks with belief propagation algorithms for approxim ate inference.Specifically we propose to combine the outer-loop step of convergent belief propagation algorithms with the M-step of the EM algorithm.This then yields an approxim ate EM algorithm th a t is essentially still double loop, with the im portant advantage of an inner loop th a t is guaranteed to converge.Simulations illustrate the m erits of such an approach.