The Research On Learning Incrementally Of Bayesian Network Structures

Bayesian Network also known as Belief Network, which originates from solving the uncertain problems about artificial intelligence in the mid 1980s, the clearly semantic structure of Bayesian Network reveals the statistical relationship in researching objects, and it is the compact representation of the complex joint probability distributions. Because the Bayesian Network has the properties of natural representation, effectively learning ability and convenient reasoning mechanism, it becomes an important theoretical tool in the field of artificial intelligence. The batch learning algorithms for the structure of the Bayesian Network can not adapt to the incomplete of the initial training data set and the model's dynamic change, which makes the incremental learning algorithms becoming a researching hot spot. This paper mainly focuses on the algorithms for incremental learning the structure of the Bayesian Network, the major work and innovation of this paper are as follows:Based on the TOCO heuristic function and the RSS heuristic function, an algorithm for incremental learning the structure of the Bayesian Network was researched. The TOCO heuristic function can determine when the structure of current network should be updated by analyzing the search path in the Hill-Climbing algorithm. The model will only be updated when there is no new valid data, and the learning algorithms will always use the structure of current network as the initial structure. The RSS heuristic function can reduce the search space by restricting the arguments, thus the efficiency of the algorithm for incremental learning the structure of the Bayesian Network can be improved effectively.Based on the lossless decomposition of Bayesian Networks, an algorithm for incremental learning the structure was be proposed. Through the conception and properties of the junction tree a single Bayesian Network was decomposed into several sub-networks. These sub-networks not only reserve the original independency information but also not introduce extraneous independency information, which assured this method is lossless. For deleting, reversing, adding edges in cliques and adding edges among cliques these four changes of network structure, in the sub-networks the algorithm using restricted maximum likelihood estimation to determine delete, reverse, add edges in cliques these three structure changes; However, between the sub-networks by detecting the independence and the changes to determine the clique pairs which occur add edges changes, then using the score function to determine the structure changes between cliques pairs. Incremental optimization the structure of Bayesian Network has been completed by combining the structure changes of within or between sub-networks at last. The effectiveness of this algorithm has been to verify by the simulation results.