基于评分缓存的节点序空间下BN结构学习  

Bayesian network structure learning based on score cache in node ordering space

作　　者：高晓光[1] 闫栩辰 王紫东 刘晓寒 冯奇 GAO Xiaoguang;YAN Xuchen;WANG Zidong;LIU Xiaohan;FENG Qi(School of Electronic and Information,Northwestern Polytechnical University,Xi’an 710129,China)

机构地区：[1]西北工业大学电子信息学院,陕西西安710129

出　　处：《系统工程与电子技术》2024年第12期4091-4107,共17页Systems Engineering and Electronics

基　　金：国家自然科学基金(61573285,62003267)资助课题。

摘　　要：针对大规模贝叶斯网络结构学习容易陷入局部最优的问题,提出一种节点序空间下迭代局部搜索算法。在局部搜索环节,设计评分缓存的选择插入算子和次优解的容忍策略,评估自适应的纵向插入邻域,攻克由盲目搜索导致的邻域受限问题。在迭代重启环节,采用等价类结构和深度优先遍历的转换机制,避免由随机扰动导致的评分退化问题。通过相融实验分别验证搜索和迭代算法的有效性。实验结果表明,相较于现有的主流方法,迭代局部搜索算法能够精确地学习大规模网络结构。Aiming at the problem that large-scale Bayesian network structure learning falls into local optima easily,an iterative local search algorithm in node ordering space is proposed.During the local search step,the selective insertion operator based on score cache and the tolerance strategy for suboptimal solutions are designed.The adaptive longitudinal insertion neighborhood domain is evaluated to overcome the limited neighborhood domain problem caused by blind search.During the iterative restart step,the conversion mechanism of equivalent class structure and depth-first search(DFS)is adopted to prevent score degradation problem caused by random disturbances.After verifying the effectiveness of the search and iterative algorithms through fusion experiments,the experimental results show that compared with existing mainstream methods,the iterative local search algorithm can learn large-scale network structures accurately.

关 键 词：贝叶斯网络 结构学习 节点序 局部搜索 迭代重启 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]