The improved MC(su3) algorithm and Bayesian network learning
Hui-Feng Shi, Milan Xing · 2008
In this paper, A new method of learning Bayesian network is presented. This method Improves the popular Markov Chain Monte Carlo (MC) method for structural learning in graphical models. In the improved learning algorithm, mutual information is used to determine the conditional independence of two variables. The bayesian network obtained by this approach is considered as the the initial status in the Markov Chain. Using the network operators(Adding, Deleting and conversing), we can get a new Bayesian network which looked as a new status of the Markov Chain. Iterating this new algorithm for given times, the latest status of the Markov Chain is obtain used as the Bayesian network structure. The result of the experiment shows that convergence velocity of the improved MC3 algorithm is faster than the ordinary MC(su3) algorithm’s, and the Bayesian network structures learned by two algorithm are similarly.