Bayesian Network Structure Learning Research Based On Multi-population Bacteria Foraging Optimization

As the battlefield environment becomes more and more complex,the recognition accuracy of the identification method relying on a single sensor is limited,and the multi-sensor information fusion method is usually used to complete the more accurate judgment of the target.Bayesian network is one of the most effective models for multi-sensor information fusion.The Bayesian network model with excellent learning performance from the sample data observed by multi-sensor is the premise of using Bayesian network for information fusion.Therefore,the study of effective Bayesian structure learning algorithm has important theoretical and practical significance.A Bayesian structure learning algorithm is produced in this thesis,which improves the low search efficiency and large structural error of the original structure learning algorithms.Meanwhile,it assists in the construction of the network structure by introducing expert knowledge and applies the results to the field of target information fusion identification.Firstly,the structure learning algorithm based on Bacteria Foraging Optimization(BFO)is improved,which is inefficient due to the fixed search parameters.So a structure learning algorithm based on Multi-population Bacteria Foraging Optimization(MBFO)is proposed in the thesis.Aiming at the problem of large solution space in the original algorithm,the method based on statistical test and Kent mapping is used to generate the initial solution.The initial solution can be uniformly distributed in the solution space while narrowing the search range.Finally,compared with artificial bee colony algorithm,bacterial foraging algorithm and ACOE algorithm,the simulation results show that the improved algorithm has smaller error and better searching speed than other algorithms when the sample data set is larger.Otherwise,the time performance of the improved algorithm is between the original algorithm and artificial bee colony algorithm.Secondly,the structure learning algorithm based on MBFO is combined with expert prior information,and the BIC scoring function of the original structure learning algorithm is improved by setting the adaptive weighted term based on the sample set size and the priori information penalty term based on the entropy theory.In addition,the evidence fusion method based on cosine similarity and entropy is used to avoid the paradox of evidence fusion.Finally,compared with MBFO algorithm without expert prior knowledge and MBFO algorithm with different scoring functions,the simulation results show that the proposed algorithm has smaller structural error compared with other algorithms and has better adaptability to error priors.Finally,the MBFO structure learning algorithm combining prior information is applied in the field of target recognition.Then,the simulation experiment is carried out with the generated test data.The results show that the recognition accuracy proposed structure learning algorithm is better than the structure learning algorithm based on BFO.It shows that the constructed network structure is better,and the effectiveness of the MBFO structure learning algorithm combining prior information is verified.