Empirical Analysis of Likelihood-Weighting Simulation on a Large, MultiplyConnected Belief Network

摘要

We analyzed the convergence properties of likelihood weighting algorithms on atwo-level, multiply connected, belief-network representation of the QMR knowledge base of internal medicine. Specifically, on two difficult diagnostic cases, we examined 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 of our model. The Quick Medical Reference (QMR) program is a decision-support tool for diagnosis in internal medicine that was developed at the University of Pittsburgh as the successor to INTERNIST-1 (Miller, Pople, et al., 1982). Designed to assist a physician in making difficult diagnoses, QMR is built on one of the largest knowledge bases (KBs) in existence. We are developing the foundation for a decision-theoretic version of QMR, which we call QMR-DT for Quick Medical Reference-Decision Theoretic. Our research to date has focused on building the QMR-DT KB, a probabilistic reformulation of the QMR KB, and on developing a method for inference on the QMR-DT KB. In this paper, we concentrate our discussion on likelihood weighting as a method of inference. Before describing the likelihood-weighting algorithm, we briefly examine the QMR-DT KB. (kr)