Research On Optimization Algorithms Of Bayesian Network

The purpose of data mining is to find useful rules and schemas which are used for description and forecast from a lot of data by using specific algorithms.Bayesian network is a kind of directed acyclic graph which provides a natural method of representing causal informantion.It is one of the most effective theoretical models in the current fields of uncertain knowledge presentation and inference.And it is often used for classification in data mining.In the structure learning of bayesian network,how to achieve better networks for better classification in data mining by optimizing exsiting search algorithms of structure is a research area.In this paper,researches on the optimizing of bayesian network are as follows:(1)The origin and development of bayesian network were briefly narrated.And the international outcomes and hot research areas of bayesian network were presented in detail.(2)The main research areas of bayesian network were analysed and discussed.The principles and methods of representation,features,structure learning,parameter learning and inference were all included.Particularly in structure learning of network,the principles and characters of methods based on scoring or conditional independence test were analysed and compared.Some typical algorithms and main features were listed.(3)The search algorithms based on scoring of structure learning were studied and discussed.The principles and processes of different common heuristic search algorithms were introduced,and their own advantages and disadvantages were analysed.The principle and specific process of genetic algorithm used for structure searching were presented,and the defect of it were analysed.(4)A kind of genetic search algorithm based on simulated annealing named SA-GS was presented in this paper.Simulated annealing was introduced into the process of genetic search in this algorithm.After every genetic search,temperature is lowered,and then a new search under current tempeture is finished.The process is repeated until reaching final temperature.In every process of the generation of a new parent population,networks of worse score are accepted at certain probabilities.(5)To test the performance of SA-GS search algorithm,the algorithm was encoded in NetBeans first,and then was used in Weka.The performances of bayesian networks which were constructed by using different search algorithms with 4 datasets were tested and compared.The results show that,SA-GS search algorithm achieve better performance than genetic search.The number of correctly classified instances increased 1 or 2.But with the increasing of the number of attributes,how to improve the efficiency need further research and perfection.