An empirical analysis of likelihood-weighting simulation on a large, multiply connected medical belief network

We are developing a probabilistic reformulation of the Quick Medical Reference (QMR) system. Our current probabilistic model of the QMR knowledge base of internal medicine consists of a two-level, multiply connected, belief network. Because of the size and connectivity of this belief network, most exact algorithms for calculating the posterior marginal probabilities of diseases are not applicable. In this paper, we analyze the convergence properties of an approximation algorithm, called likelihood-weighting simulation, on the QMR-DT belief network. Specifically, on two difficult diagnostic cases, we examine the effects of Markov blanket scoring, importance sampling, and self-importance sampling, demonstrating that the Markov blanket scoring and self-importance sampling significantly improve the convergence of the simulation on our model.