Research On Surrogate Models Assisted Particle Swarm Optimization

Many optimization problems in scientific research and engineering practice involve simulations that are computationally expensive.How to effectively solve these computationally expensive optimization problems is still a huge challenge.Recently,surrogate-assisted evolutionary algorithm(SAEA)has been widely studied,which is considered to have the potential to solve such optimization problems.In view of the poor reliability of the surrogate model,the low efficiency of the SAEA,and the poor accuracy of the offline data-driven SAEA when solving high-cost optimization problems in the traditional SAEA,three surrogate-assisted particle swarm optimization algorithms are proposed in this thesis.The main research work is as follows:(1)In order to solve the problems of imbalanced distribution of data samples and evolutionary population,as well as sample uncertainty when constructing the surrogate model,an adaptive lifting surrogate-assisted particle swarm optimization algorithm(ALSAPSO)is proposed.Firstly,the two factor individual evaluation mechanism is used to evaluate the uncertainty and fitness of the candidate solution by using the ensemble surrogate model,so as to avoid the uncertain solution being ignored by the surrogate model.Secondly,a local search method based on uncertain solution and optimal solution is proposed.These uncertain solutions can effectively help to find more uncertain solutions in the local search to add to the data set,so as to improve the surrogate model adaptively and enhance the approximation ability of the surrogate model in a part of the search space.Finally,the performance of the proposed algorithm is compared with that of four state-of-the-art SAEAs on 24 computationally expensive benchmark functions.The experimental results show that the proposed algorithm has an overall superior performance.(2)In order to make full use of the surrogate model to obtain the local features of the search space of computationally expensive optimization problems and improve the efficiency of the SAEA,a local surrogate-guided efficient particle swarm optimization algorithm(SEPSO)is proposed.Firstly,a boot strategy based on global search is proposed,which uses an independent particle swarm optimization as the optimizer to solve the problem of lack of better samples in the initial stage of data set.Secondly,a dynamic partition method based on RBF surrogate model is adopted to obtain the local guidance information of the search space and guide the search direction of the SAEA.Finally,a selective population updating strategy is used to ensure that all the individuals in the population have real fitness values,while reducing the number of additional fitness function evaluations.Comparison experiments on 24 computationally expensive benchmark functions show that the efficiency of the proposed algorithm is significantly better than four state-of-the-art comparison SAEAs.(3)In order to ensure the accuracy and reliability of the offline data-driven SAEA,an offline data-driven particle swarm optimization algorithm based on surrogate knowledge transfer(DDPSO-KT)is proposed.A method of knowledge transfer between surrogate models is proposed.On the one hand,the knowledge of global surrogate is transferred to local search to build a more accurate local surrogate model;on the other hand,the knowledge of local surrogate is applied to global search to ensure the accuracy of candidate solution.The dynamic update of the local surrogate makes it resample the offline data set for many times,so as to fully extract and use the knowledge in the data.Numerical experiments show that the proposed algorithm has certain performance advantages in computationally expensive optimization problems.According to the above research work,there are still some deficiencies that need to be improved.In the future,we can design the algorithm that can enhance the reliability of the surrogate model and improve the search efficiency of the algorithm according to the characteristics of the computationally expensive optimization problem.In addition,we can further explore the performance of the proposed algorithm in high dimensional computationally expensive optimization problems.