基于遗传算法联姻策略的贝叶斯网络结构学习  

Bayesian network structure learning method based on genetic algorithm marriage strategy

作　　者：朱宇 王慧玲 徐苗 綦小龙 Zhu Yu;Wang Huiling;Xu Miao;Qi Xiaolong(School of Network Security and Information Technology,Yili Normal University,Yining,835000,China)

机构地区：[1]伊犁师范大学网络安全与信息技术学院,伊犁835000

出　　处：《南京大学学报（自然科学版）》2024年第3期396-405,共10页Journal of Nanjing University（Natural Science）

基　　金：新疆维吾尔自治区自然科学基金(2021D01C467);南京大学计算机软件新技术国家重点实验室项目(KFKT2022B30)。

摘　　要：针对基于进化方法的贝叶斯网络结构学习易陷入局部最优和寻优效率低的问题,提出一种利用遗传算法联姻策略学习贝叶斯网络结构的技术.首先设计了“同”联姻策略,两个种群使用相同的搜索策略和评估模型完成贝叶斯网络结构学习.对学习到质量最好的子代个体进行联姻,将所得的质量最佳的子代个体共同返回两个种群中进行迭代.由于联姻的子代保留了另一个种群的片段,对种群中基因的多样性起到很好的保障,有效规避了近亲繁殖造成的缺陷.针对同代理模型的联姻策略无法同时兼顾网络结构质量及学习效率的问题,提出基于集成的遗传算法联姻策略,两个种群分别使用不同的代理模型和搜索策略进行学习,对各自学习到的当代最优个体进行联姻迭代.实验表明,提出的算法在小、中和大规模网络上的学习精度和有效性都优于对比算法.The study of Bayesian network structure based on evolution methods suffers from local optimality and low efficiency.To address this issure,a Bayesian network structure learning method based on genetic algorithm marriage strategy is proposed.First of all,the"same"marriage strategy is designed.That is,two groups use the same search strategy and evaluate models to complete the Bayesian network structure learning.Then we marry the best quality sub⁃generation individuals,and iterate the best quality sub⁃generation individuals.Because the sub⁃generation of the marriage retains another group of segments,it has a good guarantee for the diversity of genes in the population,and effectively avoids the defects caused by the reproduction of close relatives.In response to the marriage strategy of the same agent model failing to ensure the quality of network structure and learning efficiency at the same time,we propose integrated genetic algorithm marriage strategies.Specifically,two groups use different agency models and search strategies to learn,and then the best⁃quality individuals in each group are married and iterated.Experiments show that the learning accuracy and effectiveness of the proposed algorithm on all⁃scale networks are better than the comparative algorithm.

关 键 词：遗传算法 联姻策略 代理模型 贝叶斯网络结构 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]