一种利用不可行解的贝叶斯网学习算  

An Approach to Learning Bayesian Network by Using Infeasible Solutions

作　　者：李小琳[1] 何湘东[2] 陈传明[1] 

机构地区：[1]南京大学管理学院,江苏南京210093 [2]南京大学网络信息中心,江苏南京210093

出　　处：《同济大学学报（自然科学版）》2010年第5期744-748,共5页Journal of Tongji University:Natural Science

基　　金：国家自然科学基金资助项目(60803055);教育部人文社会科学一般资助项目(08JC630041);中国博士后科学基金资助项目(20080441031);江苏省博士后科研资助计划(0801038C);南京大学人才引进培养基金资助项目

摘　　要：现有的基于打分搜索的贝叶斯网学习方法都是利用满足有向无环图的可行解进行学习.在搜索过程中遇到不可行解时,这类算法简单地去除不可行解或将不可行解转化为可行解.然而,有的不可行解中往往蕴含着有价值的信息.本文提出一种新的贝叶斯网学习方法ISEC,同时利用可行解和不可行解学习贝叶斯网络,并提出针对不可行解的选择策略,在学习过程中可以有效地利用不可行解中的有用信息.实验结果表明,ISEC能够比仅利用可行解的方法更快地学习到更优的贝叶斯网.Existing Bayesian network learning approaches based on search and scoring usually work with feasible solutions which satisfy directed acyclic graph.This kind of approaches often removes infeasible solutions or converts infeasible solutions to feasible solutions when the solutions are infeasible.However,some infeasible solutions are much informative.This paper proposes the ISEC method for learning Bayesian network by using feasible and infeasible solutions synchronously based on an infeasible solution selection strategy.Then,the method can take advantage of the information in the infeasible solutions.Experiments show that the proposed approach can achieve better performance in less time than the approaches which use feasible solutions only.

关 键 词：机器学习 贝叶斯网 结构学习 最小描述长度 进化计算 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]