基于动态约束模型的贝叶斯网络结构优化算法  

Bayesian Network Structure Optimization Algorithm Based on Dynamic Constraint Model

作　　者：李沛然 刘琨[1] 张育 任佳[1] 崔亚妮[1] Li Peiran;Liu Kun;Zhang Yu;Ren Jia;Cui Yani(College of Information and Communication Engineering,Hainan University,Haikou 570228,China)

机构地区：[1]海南大学信息与通信工程学院,海南海口570228

出　　处：《海南大学学报（自然科学版）》2021年第2期132-140,共9页Natural Science Journal of Hainan University

基　　金：国家自然科学基金(61961160706);海南省自然科学基金创新研究团队项目(620CXTD434);澳门科技发展联合基金(0066/2019/AFJ)。

摘　　要：提出一种基于动态约束模型的贝叶斯网络结构优化算法.通过构建结构搜索空间约束模型,一方面利用互信息和独立性测试双重检验手段辨识节点依赖关系,限制马尔科夫覆盖的候选节点集合,缩小结构搜索空间规模,另一方面通过改进基于粒子群算法的结构搜索方式,设计粒子动态更新方程和局部搜索跳出机制,可以达到保持结构多样性,避免陷入局部最优的效果.采用2种标准网络对算法性能进行测试,实验结果表明该方法在收敛效果和学习精度上均有良好表现.The search space of the structure optimization can increase exponentially in large scale Bayesian Network(BN) learning task when applying global optimization algorithm to BN. The large search space can increase the computational complexity and reduce the accuracy due to premature optimization. In the report, a BN structure optimization algorithm based on dynamic constraint model and particle swarm algorithm was proposed. On the one hand, the mutual information and independence testing were used to identify node dependencies, which can restrict the set of candidate nodes covered by Markov for a reduced search space;on the other hand, the search method for the particle swarm algorithm was improved, a particle dynamic update equation and a local search jump-out mechanism were designed, which can maintain the structural diversity for global optimum. The results of the simulation experiments, in which the two standard networks were used to test the characteristics of the algorithm, indicated that the method has the better convergence and accuracy.

关 键 词：贝叶斯网络 结构学习 动态约束 粒子群算法 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]