基于约束最大信息熵的贝叶斯网络结构学习算法  

Algorithm for Bayesian Networks Structure Learning Based on Information Entropy

作　　者：张少中[1] 王秀坤[2] 

机构地区：[1]浙江万里学院电子信息学院,浙江宁波315100 [2]大连理工大学计算机科学与工程系,辽宁大连116023

出　　处：《小型微型计算机系统》2005年第6期983-987,共5页Journal of Chinese Computer Systems

基　　金：国家科技部"九七三"专项(2001CCA00700)资助.

摘　　要：贝叶斯网络的学习可分为结构学习和参数学习.基于约束最大信息熵的结构学习算法是一种以搜索最高记分函数为原则的方法.本文以KL距离、相互信息以及最大相互信息为基础,通过附加合适的约束函数降低变量维数和网络结构的复杂度,提出了一种附加约束的最大熵记分函数,并结合爬山法设计一种贝叶斯网络结构学习的启发式算法.通过与著名的K2和B&B_MDL算法的实验比较,结果表明该算法在时间和精度上都具有较好的效果.The automated creation of Bayesian networks can be separated into two tasks, Structure learning, which consists of creating the structure of the Bayesian networks from the collected data, and parameter learning, which consists of calculating the numerical parameters for a given structure. A score function for optimization based on maximum mutual information entropy with odditional restriction is proposed. The entropy is based on KL distance, mutual information and maximum mutual information. A hill-climb algorithm is used in Bayesian networks structure learning. Some experimentation on K2, B&B -MDL and MMI -L are given. The result indicates that the heuristic algorithm based MMI -L has more efficiency and precisely in cost and precision than K2 and B&B -MDL.

关 键 词：贝叶斯网络 结构学习 最大信息熵 约束函数 启发式算法 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论] TP18[自动化与计算机技术—计算机科学与技术]