基于结构分析的局部Gibbs抽样自动推理算法  

A Local Gibbs Sampling Automatic Inference Algorithm Based on Structural Analysis

作　　者：王浩[1] 曹龙雨[1] 姚宏亮[1] 李俊照[1] 

机构地区：[1]合肥工业大学计算机与信息学院,合肥230009

出　　处：《模式识别与人工智能》2013年第4期382-391,共10页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金资助项目(No.61070131;61175051)

摘　　要：提出一种基于结构分析的局部Gibbs抽样的贝叶斯网络推理算法(S-LGSI).S-LGSI算法基于联合树算法的概率图模型分析思想,对贝叶斯网络进行精确分解,然后根据查询结点和证据结点生成具有强相关性的局部网络模型,进而对局部网络模型进行Gibbs抽样推理.与当前基于抽样的其它近似推理算法相比,该算法降低推理的计算维数.同时,由于局部抽样模型包含了与查询结点相关的重要信息,因此该算法保证局部抽样推理的精度.算法分析和在Alarm网的实验结果表明,S-LGSI算法较显著降低时间复杂度,同时也提高推理精度.S-LGSI算法应用于上海证券交易所股票网络的推理结果与实际情况基本一致,表现出较强的实用性.In this paper, a local Gibbs sampling inference algorithm of Bayesian networks （S-LGS!） is proposed. Firstly, the S-LGSI algorithm precisely decomposes Bayesian networks based on the analytic idea of junction tree algorithm. Secondly, the suitable local model is chosen by the query node and the evidence node. Then, Gibbs sampling inference algorithm for local network model is utilized. Compared with other current approximate sampling algorithms, the S-LGSI algorithm significantly reduces the calculation dimension. The sampling inference in the local model avoids the statistics of joint sample series and greatly reduces the calculation dimension. The proposed algorithm guarantees the inference precision, as the local model contains important information about the query node. Algorithm analysis and experimental results on Alarm network show S-LGSI significantly reduces the complexity and improves the inference precision. The proposed algorithm has strong practicability, because the inference results of S-LGSI algorithm are basically consistent with the real situation on the Shanghai Stock Exchange network

关 键 词：自动推理 贝叶斯网络 马尔科夫蒙特卡洛 吉布斯抽样 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP391.3[自动化与计算机技术—控制科学与工程]