An improved EM algorithm for Bayesian networks parameter learning

The automated creation of Bayesian networks can be separated into two tasks, structure learning, which consists of creating the structure of the Bayesian networks from the collected data, and parameter learning, which consists of calculating the numerical parameters for a given structure. EM algorithm is a normal method for parameter learning in incomplete data. The traditional EM algorithm has some shortages such as that could not deal with large data sets, convergence is slow and easily results in local maximum. This paper is based on E step and M step respectively. It divides large data set into several small blocks and optimizes them in the small ones. An improved simulating anneal algorithm is used in E step. Policy of dynamic iteration, starting temperature for simulating anneal and ending condition is proposed. It adopts Cauchy as the chronology to generate adjacent value. Experimental results indicate that the improved EM algorithm proposed in the paper has more advantages than the standard EM.