Research On Structural Learning Based On Heuristic Search In Bayesian Networks

Bayesian network(BN) is graphical mode based on probability theory and graph theory, it has unique advantages in dealing with uncertain knowledge, it has been widely applied to many fields, such as of artificial intelligence, finance, medical and military. But the traditional method of building a Bayesian network only by experts priori knowledge is unreliable, it is unable to meet our needs, so, how to construct a Bayesian network from data has attracted the attention of many scholars, and become the research point in the filed.This paper research basic theoretical knowledge of Bayesian network and the existing structure learning method of Bayesian network, in order to weakness of existing algorithm, combining with knowledge of other areas, we propose new structure learning method for constructing Bayesian network, and research work of this paper is as follows:First of all, this paper detail introduces background and significance and basic theories knowledge of Bayesian network, and then describes the research status and common structure learning methods of Bayesian network.Secondly, this paper introduces MMPC algorithm and particle swarm optimization algorithm, and on this basis we propose the new Bayesian network structure learning method based on particle swarm optimization. The new algorithm integrate the ideal of two algorithms, optimizes the generate way of the initial population by experts experiences and mutual information knowledge, and find the global optimal solution by searching the neighbourhood space using particle swarm algorithm. The results of implements show that the new algorithm has better learning performance, faster convergence speed and higher solution quality.Then, this paper introduces K2 algorithm,and on this basis we propose the new Bayesian network structure learning method based on the order of nodes. The new algorithm take full advantage of breadth-first search algorithm to search the original network structure generated by MMPC algorithm, then we get the optimize order of the nodes, and put it as the initial order of K2 algorithm. Experiments show:the new algorithm has obvious advantages, reflecting the better learning performance.Finally, summarize the research contents of this paper, prospects further research work.