一种高效的贝叶斯网络敏感性分析方法  

Effective Strategy for Sensitivity Analysis of Bayesian Networks

作　　者：覃飙[1] 王秋月[1] 李超[2] 

机构地区：[1]数据工程与知识工程教育部重点实验室(中国人民大学),北京100872 [2]中国政法大学商学院,北京100088

出　　处：《小型微型计算机系统》2016年第4期732-737,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61472425)资助;中国政法大学青年教师学术创新团队项目(2014CXTD06)资助

摘　　要：我们经常需要修改网络参数以在局部信任度改变时集成新的概率信息,贝叶斯网络的敏感性分析就是研究网络参数与输出概率之间的关系.首先剪枝掉那些不能到达证据的节点,然后采用联合树算法对剪枝后的贝叶斯网络进行的敏感性分析,该算法在每一簇中缓存中间结果以加速计算P(e)相对于下一变量P(xi|ui)的一阶偏导;并且能够在各个簇中并行地计算导数.我们利用偏导的计算结果对两类网络参数进行敏感性分析:调整单个参数和调整整个条件概率表,并分别提出算法来进行处理.最后,实验结果证实了我们的算法在进行敏感性分析时的高效性.We often need to revise the network parameters to incorporate newprobabilistic information in the form of local belief changes.Sensitivity analysis in Bayesian networks is relevant to the relationships between network parameters and output probabilities.In this paper,we first prune the nodes which cannot reach the evidence and then perform sensitivity analysis of the pruned Bayesian network using the jointree algorithm.In each cluster,our jointree algorithm caches an intermediate result to accelerate the computation of the first derivatives of P（ e） with respect to the next variable P（ xi｜ ui）.In all clusters,our algorithm can compute the derivatives in parallel.Using the results of partial derivatives,we perform sensitivity analysis of Bayesian networks by changing network parameters in two types：a single parameter change and a single conditional probability table change.We propose two algorithms to process them respectively.Finally,our preliminary experimental results demonstrate the efficiency of our algorithms for performing sensitivity analysis of Bayesian networks.

关 键 词：贝叶斯网络 敏感性分析 联合树 一阶偏导 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]