Learning in Probabilistic Expert Systems

Abstract

Abstract Probabilistic expert systems use a directed graphical structure to express conditional independence relationships, and conditional probability tables to summarise quantitative knowledge. We explore the consequences of assuming these probabilities to be parameters, where beliefs about those parameters are updated as data accumulate. A simple approximate Bayesian procedure is shown to be related to those used in ‘unsupervised learning’ and is investigated by simulations and applied to a difficult real example. The procedure has reasonable properties, although for certain missing data configurations the approximations used are clearly somewhat extreme, and further work is required to handle induced dependencies between parameters.