Design And Application Of Adaptive Ensemble Surrogate-based Multi-objective Evolutionary Algorithm

Expensive multi-objective optimization problems such as multi-objective neural architecture search are widespread in the fields of science,engineering and economics.These problems involve optimising multiple trade-off objectives simultaneously,and some of these objectives can only be evaluated in small numbers due to the high cost of their evaluation in terms of time,cost and computing power.Due to the expensive evaluation process,traditional multi-objective evolutionary algorithms have many challenges in solving such problems.The main solution to this problem is to use surrogate models to approximate individual fitness as an alternative to costly real evaluation.However,most algorithms currently use only a single surrogate model to approximate individual fitness,without taking into account the need for different preferred surrogate models at different stages of evolution.In addition,the prediction accuracy of the surrogate models in expensive many-objective optimization problems needs to be further improved as the objective space increases.By analyzing the demand characteristics of different surrogate models in different stages of evolution,this paper proposes a dynamic integrated surrogate model,and further designs a dynamic ensemble surrogate-based multi-objective evolutionary algorithm(DESEA).The dynamic ensemble surrogate model consists of two local surrogate models with different preferences,which dynamically adjusts the dominance of different local surrogate models at different stages of evolution and integrates them into a global surrogate model.In the early stage of evolution,the diversity local surrogate model is dominant to increase the individual differentiation of the surrogate model,while in the later stage of evolution,the proportion of the prediction result of the convergent local surrogate model is improved.To drive the population to converge rapidly to the Pareto front.In order to promote the exploration of the population in a wider space and maintain the balance between different regions,in the selection of sampling solutions,a model management strategy based on global reference vector clustering is designed to guide the evolution of the population.Through a comprehensive comparison with four mainstream surrogate-based multi-objective evolutionary algorithms in DTLZ and WFG test sets,DESEA has achieved obvious advantages in both IGD and HV.In addition,for the severe challenges posed by expensive many-objective opimization problems,this paper designs an adaptive model management strategy based on DESEA,which adaptively selects a small number of high-value individuals and uses their valuable real evaluation data increments to train integrated surrogate models to continuously improve their prediction accuracy.The strategy first uses a clustering approach to cluster individuals in the population according to the degree of dispersion,and then chooses whether to use a homogeneity or diversity sampling strategy to select new high-value individuals for expensive true evaluations,depending on the degree of variation in the number of individuals in different classifications.A comprehensive comparison on the DTLZ,WFG test set and two real engineering problems shows that the adaptive model management strategy designed in this paper effectively improves the diversity and uniformity of DESEA in expensive many-objective opimization problems.In the quality assessment of solution sets for several algorithms,the final solution sets obtained by DESEA with the adaptive model management strategy possess wider and uniform distribution characteristics.