Research On Infill Criterion In Surrogate-Assisted Optimization Algorithms For Large-scale Expensive Problems

With the complexity of practical engineering optimization problems increasing,the dimension of decision space becomes much higher.Some problems even have no explicit objective functions,and the time to evaluate the objective functions is very time-consuming,resulting in the difficulty for evolutionary algorithms to solve these problems.Using surrogate models to approximate the objective values is a common method for solving computationally expensive optimization problems.However,more training samples are required to train a good surrogate model when the dimension of the optimization problem becomes higher.Yet it is impossible to obtain sufficient samples in a limited computational budget.Therefore,we propose to decompose a large-scale optimization problem into a number of sub-problems using the random grouping technique to assist in large-scale optimization.Then two infill criteria are proposed to obtain a better solution in a limited computational budget.The main contributions of this thesis include:(1)A surrogate-assisted large-scale evolutionary algorithm with an expected improvement variant(SAEA-EI).A large-scale optimization problem is divided into some sub-problems using the random grouping technique.Some best solutions in the archive are selected to be the training data for training a surrogate model for each sub-problem.The populations of the final generation for sub-problem optimization will be combined into a population for the large-scale optimization problem,and a solution will be selected using the proposed expected improvement(EI)variant strategy to be evaluated using the expensive problems.The performance is tested and evaluated on CEC’2013benchmark problems.Compared to some state-of-the-art algorithms proposed in recent years,the proposed method has better performance for solving expensive large-scale optimization problems.(2)A two-layer surrogate-assisted large-scale evolutionary algorithm(TSAEA).In order to speed up the convergence of surrogate-assisted large-scale optimization algorithms,the degree of contributions of each population for the sub-problem is analyzed.The sub-problems with maximum contributions will be the problem optimized in the lower layer.The problem in the lower layer will be decomposed into some sub-problems.A Gaussian process model will be trained for each sub-problem in the lower layer,and the population of sub-problem optimization will be used to replace the population of the upper layer on corresponding dimensions.Two solutions will be selected according to the expected improvement variant proposed in SAEA-EI from the upper and lower layer,respectively,for expensive objective evaluation.The comparison of the experimental results on CEC’2013 benchmark problems shows that the proposed method is efficient for solving expensive large-scale optimization problems.