一种基于柯氏复杂度的因果网络定向方法  

A causal network orientation method based on Kolmogorov complexity

作　　者：韩梦瑶 鲁云军 金乙乔 刘乾 陈克斌 HAN Meng-yao;LU Yun-jun;JIN Yi-qiao;LIU Qian;CHEN Ke-bin(College of Information and Communication,National University of Defense Technology,Wuhan 430019,China;Department of Defense Economics,Army Logistical University of PLA,Chongqing 400030,China)

机构地区：[1]国防科技大学信息通信学院,武汉430019 [2]陆军勤务学院国防经济系,重庆400030

出　　处：《控制与决策》2021年第9期2241-2248,共8页Control and Decision

基　　金：军委科技委理论科研项目(19JSLLKY015)。

摘　　要：因果网络定向问题实质是一个"多对多"因果关系发现过程,传统的V-结构定向方法只能确定一组马尔可夫等价类而非最终的因果关系.为解决该问题,从柯氏复杂度的因果推断原理视角出发,利用贝叶斯链式法则推导出局部网络因果定向规则,并在此基础上提出高维全局网络因果定向方法.同时,将前者运用于改进基于局部条件独立信息搜索学习马尔可夫毯典型算法,后者运用于改进基于约束的因果网络结构学习典型算法.实验结果表明,改进后算法在保证较高准确率的同时可有效提升执行效率.The nature of causal network orientation problems is a"many-to-many"causal discovery process.The traditional V-structure method can only determine a set of Markov equivalent classes rather than the final causal relationship.In order to solve this problem,based on the Kolmogorov complexity,a causal orientation rule of local networks is deduced using the Bayesian chain rule,thus a high-dimensional global network causal orientation rule is proposed on this basis.At the same time,the former is used to improve the Markov blanket typical algorithm based on the local condition independent information searching;the latter is used to improve the constraint based causal network structure learning typical algorithm.The experimental results show that the improved algorithm can effectively improve the execution efficiency while ensuring high accuracy.

关 键 词：因果网络 因果定向 柯氏复杂度 最小描述长度 随机复杂度 马尔可夫毯 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]