信度网推理——方法及问题(上)

Inference in Belief Network-Methods and Problems (1)

基于概率知识表达的信度网,已成为人工智能非精确知识表达与推理领域近10几年来研究的热点.目前国外的许多研究机构都对信度网进行了深入的研究.这些研究主要集中在以下三个方面:基于信度网的推理、基于信度网的学习和基于信度网的应用.其中基于信度网的推理一般分为:精确推理(即精确计算概率值)和近似推理(近似计算概率值)两个部分,主要研究高效的推理算法[13,6];基于信度网的学习一般分为参数学习和结构学习两个内容,同时根据样本数据的不同性质每一部分均包括:实例数据完备、实例数据不完备两个方面[7,6];基于信度网的应用,主要包括:基于信度网的知识表达、相应的软件工具开发、基于信度网的实例应用等.目前这些研究都取得了丰硕的成果,正逐步走向实际应用.信度网的提出人Pearl教授也于1999年被授予IJCAI杰出研究成果奖.

Belief network(BN) ,as a kind of knowledge representation based on the probabilistic theory, has become the main branch in the area of non-deterministic knowledge representation and inference in artificial intelligence (AI) in this decade. BN is being applied in many areas such as industry, military, medical treatment, commerce and so on ,and its representative systems involve expert system, voice recognition,fault diagnosis in space crafts,causality mining based on probabilistic syntax,etc. This paper introduces the primary methods and the current problems in inference of BN. After the introduction of the concept ,representation and goal of inference of BN,the paper summarizes the basic idea and problems in the principal algorithms of exact inference in BN,which include polytree propagation algorithm, clique tree propagation algorithm, graph reduction algorithm, and combination and optimization algorithm.