Parallelization Of MCMC Based On Multi-chain

Monte Carlo Markov Chain (MCMC) alogrithms are important to probabilistic inference.There are many applications for MCMC in statistics, artificial intelligence, graphics, statistical physics and so on.MCMC algorithms have a strict required to the speed and storage of the computer.Parallel computers have more fast speed and more large storage.The existing softwares for MCMC are almost based on single-process. There has been fewer researchs for Parallel MCMC algorithms, far from practical applications.In view of this situation, after concluded MCMC algorithms and the basic technologies of parallel computing, the paper proposes and implements a medel for applying MCMC algoritms in parallel computer, called simple parallelization of MCMC based on multi-chain. First we propose and anlyse the model.The model has six modules, which is called input module, data preprocessing module, parallel chains sampling module, multiple chains convergence of the diagnosis module, data processing module and export module. Then using JAGS software as sampling engine, we implement the model whith the C programming language and MPI library for parallel programming. Finally we compare our parallel model with BUGS software and JAGS software by some experiments.The experiments show that the simple parallelization of MCMC based on multi-chain model can greatly improve the speed of computing with the same accuracy.