Learning Algorithm For Bayesian Network Structure Based On The Kl Distance

The Bayesian Networks is based on probability theory, statistical theory andgraphical theory, it has solid mathematical foundation. The knowledge of probabilityand statistic give an effective method for uncertain knowledge reasoning, the graphicalmodel of Bayesian Networks can express the cause and consequence problems in realworld intuitively and do accordingly reasoning, Bayesian Networks provides a naturalmethod for expressing causal relationship, it has powerful probability reasoning,semantic clarity and easy to understand, it is suitable for expression and analysis thenature of uncertain and probability things, and can make effectively reason forincomplete, imprecise or uncertain knowledge or information, it is one of the besteffective theoretical model and research hotspot in artificial intelligence at present.Based on these advantages, Bayesian Networks is widely used in many fields, and itsfeasibilityand validityhave been verified in applications.The application of the Bayesian Networks is based on the Bayesian Networks'learning. The Bayesian Networks'learning includes two factors: structure learning andparameter learning. The parameter learning is on the specific network structure, sostructure learning is the core of the Bayesian Networks'learning. The effective structurelearning methods and algorithms are the base to construct the best Bayesian Networks.It is the difficulty in graphical model learning (include Bayesian Networks) thathow to get knowledge from data in order to achieve machine-learning, and correctlyexpress the valuable information which inherent in data. Bayesian Networks is adirected acyclic graph model, which expresses is non-causal loop knowledge and it onlycan do non-causal loop reasoning. In Bayesian Networks'learning, to avoid causal loopis how to avoid producing directed ring.The main purpose of this paper is to explore Bayesian Networks'learning, thepaper analyzed the origin and development, the application and research of the BayesianNetworks; described the basic application pattern of graphic model in uncertainknowledge representation and reasoning; focused on the research of BayesianNetworks'learning theory, described the main content and the main method adopted forBayesian Networks structure learning, studied the mechanism and methods for BayesianNetworks'learning; analyzed the influence of nodes order in Bayesian Networks structure learning. By analyzing the mechanism of Bayesian Networks structurelearning and relationship between nodes in directed graph, a hierarchical parentsconcept was proposed, and then proved that in the Bayesian Networks, the same levelparents are conditional independence. Combining proper scoring function and itsapplication in probability relationship comparison, on the concept of hierarchical parentnodes, proposed a KL distance-based ring deletion algorithm which applied for ringdeletion in process of Bayesian Networks structure learning, thus eliminating thedependence on nodes'order in the process of Bayesian Networks structure learning. Inthe Bayesian Networks structure learning, the idea of the algorithm in the paper isscoring and ring deletion, using the scoring function to search the best parents, whichwill produce rings, then deleted the weaker arcs according to KLdistance to achieve thepurpose of ring deletion. By experimenting analysis and comparison, it showed thealgorithm can learn Bayesian Networks structure and achieve better results in the caseof knowing less priori knowledge, that is the case of does not know order of nodes.Meanwhile, in order to do further research, after analyzing experimental results, thispaper also discussed the shortcomings of the algorithm, and analyzed some possiblereasons which bring the shortcomings.