Online Data-driven Evolutionary Algorithm Research Based On Surrogate Model

When solving many real-world multi-objective optimization problems,it is usually difficult to find a suitable evaluation function or the evaluation cost is very expensive.This kind of evolutionary optimization problem can be solved by data-driven methods collected in experiments or daily life,which is called data-driven evolutionary optimization.Datadriven evolutionary algorithm is an effective method to solve expensive optimization problems.Generally,the training agent model is used to approximate the objective function evaluation to reduce the evaluation cost.According to whether the new data generated in the optimization process can be used,data-driven evolutionary optimization can be divided into offline data-driven evolutionary optimization and online data-driven evolutionary optimization.Offline data-driven evolutionary optimization can not use the new data generated in the optimization process,while online data-driven evolutionary optimization is the opposite.Therefore,online data-drive is more flexible than offline data-drive.In order to solve expensive multi-objective optimization problems,this paper proposes four online data-driven evolutionary algorithms based on agent model,and the specific research contents are as follows:(1)Based on the advantages of non-dominated solutions in Pareto dominance,an online data-driven evolutionary algorithm NSUSODDEA updated by non-dominated solutions is proposed.Compared with other solutions,the non-dominated solution has the least goal conflict and can provide a better choice space for the decision-maker.Therefore,NSUSODDEA chooses the non-dominated solutions to update the proxy model.(2)Aiming at the influence of sensitive points on the quality of proxy model,an improved kriging model-assisted two-archive online data-driven evolutionary algorithm KTA2＿add Model4 is proposed.The algorithm is improved according to KTA2.Three kriging models trained in KTA2 are used as agent models,while in KTA2＿add Model4,four kinds of kriging models are trained as proxy models,in which an insensitive model 3 is added that does not have large influence points and small influence points.(3)Aiming at the conflict between convergence and diversity in multi-objective optimization problems,a dual model online data-driven evolutionary algorithm TAODDEA based on two-archive is proposed.In this algorithm,two-archive--convergence archive and diversity archive are used to train two models--convergence model and diversity model.Furthermore,one of the two models is selected as the final prediction model through the decision-making mechanism.(4)In order to give consideration to both convergence and diversity,an online datadriven evolutionary algorithm TAUPODDEA with two-archive update strategy is proposed.In this algorithm,the two-archive--convergence archive and diversity archive are used to update the agent model.During the evolutionary optimization process,the solutions in the convergence archive and diversity archive are alternately selected to update the training set,so that the training set has both good convergence and diversity solutions.