Stochastic Sample Based Algorithm for Learning Bayesian Networks
Xuegang Hu · 2009
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.