信度网中条件概率表的学习

Learning CPT in Belief Networks

一、引言信度网B的学习包括结构B(?)的学习和条件概率表B_p的学习。因果马尔可夫条件原理表明:如果图形G是一个随机变量集合X的因果图,那么图形G也是该随机变量集合的联合概率分布所对应的信度网的结构图。根据这一原理,在实际应用中。

Learning with a belief network is to build its two basic components from data: a network structure and a set of conditional probability tables. It is of crucial importance in the practical application of belief networks,so is one of the current hot research topics. In this paper,we discuss approaches to learn conditional probability tables from a set of data,given a fixed network structure. We will analyze two representative exact algorithms and three widely used approximate algorithms:maximum likelihood estimation, maximum a posterior, EM algorithm, gradient ascent algorithm and Gibbs sampling algorithm,in order for users to choose based on their advantages and shortcomings.