基于双尺度约束模型的BN结构自适应学习算法  

BN Structure Adaptive Learning Algorithm Based on Dual-scale Constraint Model

作　　者：戴晶帼 任佳[1] 董超 杜文才[1,3] DAI Jing-Guo;REN Jia;DONG Chao;DU Wen-Cai(School of Information and Communication Engineering,Hainan University,Haikou 570228,China;Key Labora-tory of Marine Environmental Survey Technology and Applica-tion,Ministry of Natural Resources,Guangzhou 510310,China;Institute of Data Science,City University of Macao,Macao 999078,China)

机构地区：[1]海南大学信息与通信工程学院,海口570228 [2]自然资源部海洋环境探测技术与应用重点实验室,广州510310 [3]澳门城市大学数据科学研究院,澳门999078

出　　处：《自动化学报》2021年第8期1988-2001,共14页Acta Automatica Sinica

基　　金：国家国际科技合作专项(2015DFR10510);国家自然科学基金(61562018);国家海洋局南海维权技术与重点实验室开放基金(1704);海口市重点科技计划项目(2017041)资助。

摘　　要：在无先验信息的情况下,贝叶斯网络(Bayesian network,BN)结构搜索空间的规模随节点数目增加呈指数级增长,造成BN结构学习难度急剧增加.针对该问题,提出基于双尺度约束模型的BN结构自适应学习算法.该算法利用最大互信息和条件独立性测试构建大尺度约束模型,完成BN结构搜索空间的初始化.在此基础上设计改进遗传算法,在结构迭代优化过程中引入小尺度约束模型,实现结构搜索空间小尺度动态缩放.同时,在改进遗传算法中构建变异概率自适应调节函数,以降低结构学习过程陷入局部最优解的概率.仿真结果表明,提出的基于双尺度约束模型的BN结构自适应学习算法能够在无先验信息的情况下保证BN结构学习的精度和迭代寻优的收敛速度.In the absence of prior information, the size of the search space for Bayesian network(BN) structures grows exponentially with the increasing number of nodes, resulting in the great difficulties of BN structure learning. To solve this problem, BN structure adaptive learning algorithm based on dual-scale constraint model is proposed. The maximum mutual information and conditional independence tests are first used in the proposed algorithm to build a large-scale constraint model, completing the initialization of the search space for BN structures. Based on this, an improved genetic algorithm is presented, introducing a small-scale constraint model during the iteration to realize the dynamic adjustment of the search space. At the same time, an adaptive control function of mutation probability is constructed in order to reduce the probability of getting trapped in a local optimum. The simulation results show that BN structure adaptive learning algorithm based on dual-scale constraint model can guarantee the accuracy of learning BN structure and the convergence speed of the iteration without prior information.

关 键 词：贝叶斯网络 结构学习 约束模型 遗传算法 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]