An extension on learning Bayesian belief networks based on MDLprinciple

摘要

Bayesian belief network (BBN) is a framework forrepresentation/inference of some knowledge with uncertainty. Since theprocess of constructing a BBN manually by experts is time-consuming ingeneral, some method supporting the task is needed. We proposed analgorithm for acquiring some BBN automatically from finite examplesbased on minimum description length (MDL) principle. This paperaddresses an improvement which relaxes a constraint that the originalscheme held on the representation. In BBNs, attributes and stochasticdependencies between them are expressed as nodes and directed linksconnecting them, respectively, where each attribute may be a predicate,a numerical data, etc., and each dependence is numerically expressed asthe conditional probability of one attribute given other attributes iftheir dependence exists. Therefore, in general, BBNs are represented interms of the network structure and the conditional probabilities