基于PS—EM算法和BP神经网络的影响图模型选择

A Model Selection Method of Influence Diagrams Based on PS-EM Algorithm and BP Neural Network

影响图模型选择中存在数据依赖性、计算复杂性和非概率关系问题.通过对影响图结构进行分解,提出PS-EM 算法对影响图的概率结构部分进行模型选择.给出一种 BP 神经网络,通过对局部效用函数的学习实现效用结构部分的模型选择,并引入权重阈值来避免过拟合.PS-EM 算法是在 SEM 算法中引入一种融合先验知识的MDL 评分标准来降低传统 MDL 评分对数据的依赖性,并通过将参数学习和结构评分分开计算提高计算效率.算法比较的结果显示 PS-EM 比标准 SEM 的时间性能好、对数据依赖性小,且效用部分的结构选择易于实现.

In the model selection of influence diagrams（IDs）, the problems of the data dependency, the computation complexity and non-probability relation are discussed. Based on the structure decomposition of IDs , a PS - EM algorithm is presented . A BP Neural Network is introduced by learning local utility function of each utility node, and the overfitting is avoided by inducing the threshold of weights. To reduce the data dependency, a new MDL scoring is presented which includes the prior knowledge of network structures. Based on SEM algorithm, PS-EM algorithm induces the new MDL scoring, and separates parameters learning from structures scoring to improve the computation efficiency. Compared with SEM algorithm, the performances of both the computation complexity and the data dependency of PS-EM algorithm are improved, and the model selection of the utility part is easy to achieve.