Learning Accurate Belief Nets

Bayesian belief nets (BNs) are typically used to answer a range of queries, where each answer requires computing the probability of a particular hypothesis given some spec-iied evidence. An eeective BN-learning algorithm should, therefore, learn an accurate BN, which returns the correct answers to these speciic queries. This report rst motivates this objective, arguing that it makes eeective use of the data that is encountered, and that it can be more appropriate than the typical \maximum likelihood" algorithms for learning BNs. We then describe several different learning situations, which diier based on how the query information is presented. Based on our analysis of the inherent complexity of these tasks, we deene three algorithms for learning the best CPtables for a given BN-structure, and then demonstrate empirically that these algorithms work eeec-tively.