Construction And Management Of Surrogate Models In Expensive Multi-Objective Evolutionary Optimization

In high-dimensional and expensive multi-objective optimization problems,with the gradual increase of the scale of the problem,the simulation and evaluation of actual problems require a lot of time and computational cost.In order to achieve the expected results with limited computing resources as much as possible,usually Inexpensive surrogate models are introduced to speed up the search for the optimal solution to the problem and save valuable computing resources.At present,in the study of expensive multi-objective optimization problems,many studies have applied surrogate models,and constructing and using appropriate surrogate models has become an important aspect in the study of expensive multi-objective optimization problems.This paper studies the expensive multi-objective optimization problem,and proposes an adaptive modeling strategy and a filling criterion method based on the characteristics of the objective function.The main research contents of this paper are as follows:(1)Most of the construction methods of surrogate models in the existing research are to build corresponding surrogate models for specific algorithms,but the ability of the models to adapt to other algorithms is not strong.This paper proposes a modeling method based on adaptive model selection.The self-adaptive model of data features effectively improves the applicability of the model in different environments.This method sets the mean value of the variance of the sample fitness value of the objective function as a threshold.When the variance of the sample fitness value of the time-consuming objective function is greater than the threshold,a global surrogate model is constructed to improve the global exploration ability of the algorithm;when the variance of the sample fitness value of the time-consuming objective function is less than at this threshold,a local surrogate model is established to enhance the local exploration ability of the algorithm.Through this threshold,the global model and the local model work together,which improves the processing capability and applicability of the algorithm.In order to verify the effectiveness of the proposed model construction method,the method is applied to the double-archive time-consuming multi-objective optimization algorithm based on Gaussian process assistance and the time-consuming multi-objective optimization algorithm based on Gaussian process assistance reference vector guidance,and tested in DTLZ function to test.The experimental results show that the adaptive model construction method based on data features can effectively solve the time-consuming multi-objective optimization problem.(2)To balance the convergence of the population and its diversity,a cluster-based filling criterion is proposed for a surrogate model-assisted expensive multi-objective evolutionary algorithm.First,all the individuals selected by the environment are clustered according to the individual attributes,and then by judging the distance from the cluster center of each category to the reference point,the individuals to be truly evaluated in each generation are selected.This method will select two individuals at a time.One individual selects an individual with the largest crowding degree from the class closest to the cluster center,and the second individual selects an individual closest to the ideal point from the second closest class to the cluster center.Such an individual selection strategy It not only ensures the diversity of the population but also promotes the convergence speed of the population.By comparing with several existing algorithms on the DTLZ test function,the experimental results show that the clustering-based filling criterion proposed in this paper performs better.