The Research On Inference Algorithm For Bayesian Networks Based On Sampling

Bayesian network provides a powerful graph tool to express domain knowledge based on probability. It has becomes an important methodology for representing and computing uncertain problem of stochastic processes in the field of artificial intelligence, and it has been successfully applied to fault diagnosis, data mining, medical diagnosis, and other fields. Dynamic Bayesian network is a Bayesian network with the expansion on the factor of time, it provides a powerful tool to represent and deal with dynamic uncertain problem of stochastic processes in the field of artificial intelligence. Based on the comprehensive overview of Bayesian network, this thesis focuses on the research of approximate inference algorithm for Bayesian network and dynamic Bayesian network. The main contents of this thesis are as follows:(1) This thesis makes a survey about the research on Bayesian network, including the origin, development, the model, the construction process, the type and the application of Bayesian network. On basis of this, dynamic Bayesian network is introduced. And, the main research contents of Bayesian network are summarized. Moreover, the inference algorithms for Bayesian network and dynamic Bayesian network are introduced in detail.(2) In the approximate inference of Bayesian network, to deal with the existing problems of Markov Chain Monte Carlo (MCMC) inference algorithm, taking Gibbs sampling as an example, a parallel MCMC (PMCMC) inference algorithm for Bayesian network is proposed. To improve the accuracy, larger samples are used in PMCMC when generating the sequence of Markov chain. Moreover, to guarantee the inference speed, the algorithm has been implemented using the Master-Slave parallel programming model with the support of MPI. The experiment results on three different Bayesian networks show that PMCMC can achieve higher accuracy without the loss of time performance.(3) In the approximate inference of dynamic Bayesian network, to improve the performance of particle filtering (PF) inference algorithm, discrete particle swarm optimization (DPSO) technique is introduced to form a new filter called the evolutionary particle filter (EPF), in which the iterative search and optimization procedure of DPSO is used to redistribute particles closer to the true posterior density. The experimental results on two different dynamic Bayesian networks show that EPF can achieve better accuracy with fewer particles.