Research On Dynamic Surrogate Models Assisted Particle Swarm Optimization

Real-world optimization problems in practical engineering are almost expensive black box optimization problems,and it often involves computationally expensive experiments or simulations.This characteristic will seriously affects the schedule of product development,so the using of surrogate assisted evolutionary algorithm(SAEA)to reduce the times of experiments or simulations is a research hotspot.However,the surrogates used in evolutionary algorithms still are in low efficiency.To improve the efficiency of SAEA,this paper studies on it and the main contents are given as follows:Firstly,based on the situation that the surrogates in SAEA are only used to predict the fitnesses of individuals,this paper propose an algorithm(EGOAPSO)that the global search ability of Kriging based on EI criterion is used to assist the search process of Particle Swarm Optimization.Meantime,to improve the precision of surrogate and reduce the time of constructing surrogate,the strategy that the area of buliding surrogate is determined by the Particles in PSO is used in this algorithm.Secondly,to improve the the efficiency of SAEA,two-layer multiple surrogates assited PSO(TLMPSO)is proposed on the basis of the EGOAPSO.In term of global scope,multiple surrogates are constructed and the global search abilities of them based on the EI criterion are used to search in the global scope.To some extent,it can cover the shortage of inaccurate information provided by only one imprecise Kriging.In the term of local scope,a local RBF surrogate is constructed to search in the neighbourhood of current optimum solution to modify the current optimum solution.Finally,this paper introduces the statics analysis and fatigue life analysis of the drive shaft of the all-direction propeller,and constructes a structural optimization design model to maximize the fatigue life with the restraines of weight and 6 important sizes.The procedure based on TLMPSO to slove this design model is given later.Through the analysis and comparison of the results of this model,the feasibility and effectiveness of the proposed algorithm in this paper are verified when sloving real-world optimization problems.