基于进化MCMC的DBN学习算法

摘要

Dynamic Bayesian Network(DBN) learning with evolutionary MCMC algorithm is presented.Parameter learning with absent data is done with EM algorithm.In structure learning, multiple Bayesian network chromosomes are generated as candidates and are processed with mutation and crossover.The structures are learned with MCMC function which is obtained with temperatures and Bayesian Information Criterion(BIC) scores of corresponding Bayesian networks.In each generation, the Bayesian network with the maximal BIC score is selected as the result of the stracture learning.Experimental results proves the stability of the method's performance.%提出利用进化MCMC算法进行动态贝叶斯网络(DBN)学习的方法.在数据缺省情况下利用EM算法进行贝叶斯网络参数学习,结构学习部分生成多条备选的贝叶斯网络染色体,对染色体进行变异操作和交叉操作,在遗传操作中根据温度参数和贝叶斯网络及贝叶斯信息准则来构造MCMC函数,并利用MCMC函数进行贝叶斯网络学习.每一代进化后,将贝叶斯信息评分最大的贝叶斯网络作为结构学习的结果.实验结果验证了该方法性能的稳定性.