Research On Bayesian Network Structure Learning Methods Based On Knowledge And Data

ith its own unique advantages,Bayesian network has become one of the most efficient theoretically models in the domain of uncertainty knowledge representation and reasoning and is widely used in many fields such as artificial intelligence,fault diagnosis,and so on.Bayesian network learning includes parameter learning and structure learning.Among in these two parts,structure learning is the core task for Bayesian network learning.In this paper,Bayesian network structure learning is studied based on knowledge and data.In this paper,we study the Bayesian network structure learning based on data and propose the pMIC-BPSO-ADR hybrid algorithm.First of all,the algorithm improves the independence-based algorithm based on the Maximum Information Coefficent(MIC)(the first stage algorithm),with the indroduction of punishment factor p1,p2 method to avoid invalid triangle rings.The local network framework is determined by the indipence-based algorithm based on improved MIC(the first stage algorithm).Then,the structure learning is used the score-based algorithm based on BIC scoring function and the search strategy of BPSO+ADR heuristic search algorithm.This paper puts forward the idea of the smallscale search+large-scale search,using BPSO to search the local network framework and using the proposed new ADR heuristic search algorithm to search the global network framework.Experiment results showed that the proposed pMIC-BPSO-ADR hybrid algorithm effectively improved the efficiency and quality of Bayesian network structure learning.In this paper,we study the method of applying knowledge in Bayesian network structure learning and proposed a method of using different types of experts' knowledge in the hybrid algorithms.In this paper,experts' knowledge is divided into two types:complete and incomplete knowledge,subdivided into six kinds of knowledge.In the fisrt stage algorithm,we formulate the rules for adding and deleting edges of different types of experts' knowledge to amend the generation process of the initial network.In the second stage algorithm,we improve the scoring function by using the differernt types of experts' knowledge as the penalizations of BIC scoring function.Make the structure less punishing when it tends to experts' knowledge.So the more consistent with experts'knowledge,the easier it is to select a network.Experiment results showed that the proposed method of incorporating hybrid structure learning algorithm with different experts' knowledge is effective and effectively improved the efficiency and quality of Bayesian network structure learning.