The Research Of Bayesian Network Structure Learning Algorithms Based On Artificial Fish Swarm

Bayesian network is a typical classifier. It can be used to represent uncertaintyinformation combined with knowledge of probability theory and graph theory andhandle data sets with missing data or the presence of noise data. The first task of theBayesian network is to construct Bayesian classification model, i.e. BN structure. Inorder to construct a better data fitting Bayesian classification model by learning fromdata sets instead of by priori knowledge constructed by experts. It has a variety ofalgorithms to optimize Bayesian network structure. There are two types of popularmethods:1Learning methods based on information theory,2Learning methods basedon score and search algorithm. The second method is studied more. This paper also uses aBayesian score and artificial fish-swarm search method to learn Bayesian networkstructure.The artificial fish-swarm is a bionic optimization algorithm proposed in recentyears. Artificial fish individual chooses behavior to adjust its status according toambient so as to achieve the global optimum. It has good robustness, fast convergencecapability advantages, and so on. Using artificial fish-swarm will study the structureof the Bayesian network on the MBNC experimental platform. The main steps includecoding, initialization, bulletin boards, behavior selection, using the Bayesian score asthe fitness value of each artificial fish. In this problem, there is a big differencebetween optimization of Bayesian network structure and optimization of others: itoptimizes a structure, i.e. a series of directed acyclic graph. A series of studies for thespecial nature are carried on in this article: study about mobile strategies of changingthe order of the nodes and of pulsing side, minusing side, reversing side. In this paper,a mobile strategy of changing the order of the nodes is adopted. Future research cantake other mobile strategy to improve the algorithm. The parameters of Artificial fishnumber and iterations number have an influence on the result, which is verified byexperiments. This paper selected the best parameters. Finally, the paper studied the K2learning method and its realization in the BNTtoolbox in matlab. The artificial fish-swarm can optimize the order of the Bayesiannetwork node, then the paper used it as the initial node sequence of the K2algorithm.The experiment was made to compare the three models: AFS_BNS, AFS_K2, K2.From these experiments it was concluded that the use of artificial fish swarmalgorithm to learn Bayesian network structure was feasible and effective and could getthe best results. Therefore, the artificial fish-swarm algorithm can be used as a newapproach to learning Bayesian network structure.