Research On Structure Learning Algorithm Of Bayesian Network Based On Improved Particle Swarm

Learning the Bayesian network structure from data is an NP-hard problem.At present,the hybrid structure method combining score-and-search(SS)and constraintbased analysis(CB)can make best use of its strengths and avoid its weaknesses,and better learn the Bayesian Network(BN)structure.However,when the global optimization algorithm is used to complete the Bayesian Network learning task,the solution space size of the structure optimization will increase exponentially with the increase of the number of nodes,which is easy to produce problems such as high computational complexity,decreased accuracy,premature optimization and low algorithm stability in the model optimization process.So,this paper first proposed a wake-up optimized particle swarm structure learning method based on search space constraints.Through the Maximum Weight Spanning Tree(MWST)and Mutual Information(MI)constraint edge generation conditions,initial particles with different node sequences can be generated according to different root nodes defined.Then,the idea of Particle Swarm Optimization(PSO)is used to optimize the updating mechanism to search the network structure.Finally,a "wake up" mechanism is added to force the updating and replacing part of the particles after the particle fitness value converges,so as to improve the learning accuracy and prevent the algorithm from falling into local optimal convergence too fast.Secondly,in order to improve the stability of structure learning quality,accelerate the convergence speed,reduce the number of unnecessary search cycles,and improve the efficiency of structure learning,this paper makes an improvement on the basis of the previous method,and proposes a Bayesian Network structure optimization algorithm based on Markov Blanket dynamic constraint model.The algorithm by constructing two phase structure search space constraint model,using the mutual information and independence test double inspection means to identify nodes dependencies,limit the Markov Blanket of the candidate set of nodes,reducing the searching space scale structure,and on the basis of the improvement based on the structure of the particle swarm algorithm search,by designing the particle dynamic updating equation,keep the diversity of structure,increase the stability of the quality of the network structure in the search results,at the same time to speed up the optimization algorithm.Finally,the simulation experiments by using two standard networks to test the performance of algorithm,with the BNC-PSO algorithm,IK2 v MB algorithm,Hill Climbing algorithm,and the random initialization of the structure of the Particle Swarm Optimization algorithm,to prove the superiority of the two algorithms in the constraint stage,and timeliness and stability in the search stage,two algorithms is verified as good effects in learning Bayesian Network structure.