贝叶斯网学习中一种有效的爬山算法  

Efficient Hill-climbing Algorithm for Learning Bayesian Networks

作　　者：单冬冬[1] 吕强[1,2] 李亚飞[1] 王磊[1] 

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215006 [2]江苏省计算机信息处理技术重点实验室,江苏苏州215006

出　　处：《小型微型计算机系统》2009年第12期2457-2460,共4页Journal of Chinese Computer Systems

基　　金：国家教育部博士点基金项目(20060285008)资助;江苏省自然科学基金项目(BK2003030)资助

摘　　要：提出在学习贝叶斯网下的一种行之有效的爬山算法,HCBest算法.该算法在学习网络结构形成环时,选择删除能提高打分值最多的边,直到没有环为止.实验证明,HCBest既可以作为一种独立的贝叶斯网学习方法,又可以作为其它复杂元启发方法的局部搜索算法.HCBest学出的网络在打分质量和结构上都比较好.在算法的简洁性和稳定性方面,HCBest的表现也令人满意.This paper presents an efficient hill-climbing approach for learning Bayesian Networks, HCBest algorithm. When a circle emerges in the process of evolving networks, the algorithm delete the edge that can best improve the score of the learned network until there is no circle. The experiments show that, HCBest not only can be used as an independent approach for learning Bayesian networks, but also can be applied in other complicated meta heuristic approach as a local search algorithm. HCBest can achieve good quality network structure with higher metric score. In the aspects of simplicity and robustness, the algorithm also performs well.

关 键 词：学习贝叶斯网 爬山算法 打分制 元启发 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]