基于混合方式的贝叶斯网络结构学习  

Structural Learning Bayesian Network Based on a Hybrid Method

作　　者：张燕[1] 朱明敏[1] 宋苏鸣 

机构地区：[1]西安电子科技大学数学系,陕西西安710071

出　　处：《电子科技》2014年第10期115-118,共4页Electronic Science and Technology

摘　　要：基于最大主子图分解技术和遗传算法,提出了一种混合方式的贝叶斯网络结构学习算法。该算法首先根据领域知识和观察数据构造网络的无向独立图,并对其进行最大主子图分解,再利用遗传算法学习每个子图的结构,同时进行合并修正得到最优的贝叶斯网络结构。分解过程将一个学习大网络问题转化为小子图的学习问题,降低了搜索空间。仿真结果表明,新算法的学习效果与运行效率均有明显提高。A hybrid algorithm for structure learning of Bayesian network which based on maximal prime decomposition technology and genetic algorithm is proposed. The algorithm first constructs the undirected independence graph of a BN according to domain knowledge and observation data. Then it performs MPD to decompose the undirected graphs. The genetic algorithm is used to learn the local structure and combine the subgraphs then correct them to obtain the final BN. The decomposition splits the problem of learning a large network into some problems of learning small subgraphs. Experimental results show that the learning ability and performance of novel algorithm are improved significantly.

关 键 词：贝叶斯网络 Markov边界 最大主子图分解 遗传算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]