The Hybrid parameter learning algorithm development for dynamic Bayesian network in the context of the Metropolis — Hastings approach

Abstract

Bayesian networks are an efficient tool for analyzing complex stochastic processes. The generic methods of analysis based on Bayesian networks are probabilistic inference procedures that aimed at solving filtering, prediction and smoothing problems. An essential stage in the probabilistic inference methods implementation is network structure and parameters learning carried out on the special training samples obtained as a result of passive observation of the analyzed process and/or active experiment. Effectiveness of the Bayesian networks parameters learning algorithms directly depends on the preprocessing and organization of the training sample processing, which optimizes the classical learning algorithms capabilities. In the context of research introduced in this article, a hybrid algorithm for the parameters learning of dynamic Bayesian networks with observed and hidden variables has been developed. Algorithm allows increasing the efficiency of probabilities calculation procedure for the initial state and the probabilities of transition between the networks slices involving the variety of special methods for organizing training sample preprocessing. The questions of convergence of the proposed algorithm are also considered.