一种基于随机抽样的贝叶斯网络结构学习算法  

Stochastic Sample Based Algorithm for Learning Bayesian Networks

作　　者：胡春玲[1] 胡学钢[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：《计算机科学》2009年第2期199-202,共4页Computer Science

基　　金：安徽省自然科学基金课题(编号050420207)资助

摘　　要：针对贝叶斯网络的结构学习问题,基于并行随机抽样的思想提出了结构学习算法PCMHS,构建多条并行的收敛于Boltzmann分布的马尔可夫链。首先基于节点之间的互信息,进行所有马尔可夫链的初始化,在其迭代过程中,基于并行的MHS抽样总体得到产生下一代个体的建议分布,并通过对网络中弧和子结构的抽样产生下一代个体。算法PCMHS收敛于平稳分布,具有良好的学习精度,而该算法又通过使其初始分布和建议分布近似于其平稳分布,有效提高了马尔可夫链的收敛速度。在标准数据集上的实验结果验证了算法PCMHS的学习效率和学习精度明显优于经典算法MHS和PopMCMC。Based on the ideas of parallel stochastic sampling, this paper put forward an algorithm PCMHS for learning Bayesian networks. The PCMHS algorithm runs multi parallel Markov chains converging to Boltzmann distributions. The algorithm PCMHS, based on the mutual information between nodes, initializes all Markov chains. In the process of iteration, the algorithm, based on the population from parallel Metropolis-Hasting samplers, generates the proposal distribution for the next generation,and uses arc sample and sub-structure sample to produce the individuals of the next generation. The algorithm PCMHS converges to stationary distribution and its learning accuracy is high. In addition, it designs initial sample and proposal distribution as close as possible to the stationary distribution, and greatly improves the convergence rate. The experimental results on standard data set prove that the learning accuracy and efficiency of the algorithm PCMHS greatly outperform the classical algorithms MHS and PopMCMC.

关 键 词：贝叶斯网络 结构学习 随机抽样 马尔可夫链 建议分布 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] O212.2[自动化与计算机技术—控制科学与工程]