基于节点排序的贝叶斯网络结构学习算法  

Bayesian Network Structure Learning Algorithm Based on Node Ordering

作　　者：姚洁[1] 朱响斌[1] 宋新方 李广龙[3] 邱慧玲[1] 

机构地区：[1]浙江师范大学数理与信息工程学院,浙江金华321004 [2]横店集团东磁股份有限公司,浙江东阳321118 [3]山东省曹县第一中学,山东曹县274400

出　　处：《计算机工程》2017年第5期317-321,共5页Computer Engineering

摘　　要：针对K2算法学习贝叶斯网络结构时需要确定节点顺序的问题,提出一种混合贝叶斯网络结构学习算法。在给定数据集的情况下,利用MMPC算法获得网络的初始结构图,应用广度优先搜索的方式对此初始结构图进行搜索,从该图中入度为0的节点出发,按层次依次访问图中的邻接点,获得优化的节点顺序。将该节点顺序作为K2算法的初始节点顺序,再利用K2算法对空间进行搜索,找到全局最优解。实验结果表明,与K2算法和限制性粒子群算法相比,该算法在相同的样本数据集下产生多边、少边和反边情况的概率更低,并且可学习到更准确的贝叶斯网络结构,收敛速度快、求解精度高。Due to the problem that K2 algorithm requires node ordering in learning Bayesian network structure, this paper proposes a hybrid Bayesian network structure learning algorithm. In the situation of a given data set,it uses Max- rain Parents and Children（MMPC） algorithm to obtain the initial network structure and utilizes the way of Breadth First Search （BFS） to search the initial network structure. It startly searchs from the node whose in-degree is zero and visits in turn the adjacent points in the figure according to level, thereby it can gain the node order and make it as the initial node order of K2 algorithm. Then,it uses K2 algorithm to search the experimental results show that compared with K2 algorithm network space to find out and Restricted Particle the global optimal solution. The Swarm Optimization （ RPSO ） algorithm, the new algorithm has lower probability of multi-edge, lack-edge and reverse-edge under the same sample data set. It can learn more accurate Bayesian network with faster convergence speed and higher precision.

关 键 词：贝叶斯网络 结构学习 MMPC算法 K2算法 广度优先搜索 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]