Surrogate-assisted Evolutionary Optimization Algorithm For Expensive Optimization Problems

Evolutionary Algorithms(EAs)is a heuristic stochastic optimization method that evolves a population of offspring best suited to the environment based on the idea of survival of the fittest.The population is randomly mutated to retain promising individuals based on the results of environmental selection,and the optimal individual is found after repeated iterations,i.e.,the optimal solution.Evolutionary algorithms are simple in principle,easy to implement,and converge quickly,so they have become one of the popular methods to deal with optimization problems.However,the computational size of the evaluation function for individual selection in some specific problems is too large,which often leads to expensive computation and extremely slow operation when evolutionary algorithms are applied to practical engineering fields.In order to broaden the application of evolutionary algorithms to expensive optimization problems,an agent-assisted evolutionary algorithms(SAEAs)has been proposed in recent years,which uses an agent model built from historical data to replace part of the evaluation of real functions,thus greatly reducing the evolutionary process computational cost and is able to find promising solutions in a limited time.To address the possible improvements of the existing SAEAs main framework and the use of agent models,the following work is done in this paper.(1)The impact of the selection of training samples in the agent model of SAEAs on the evolutionary effect of the algorithm is analyzed.When the agent model is selected,the training samples selected for two consecutive iterations have an important impact on the effect of model-guided evolution.Based on this,a collaborative agent-assisted modeling method is proposed in this paper.The method first classifies all samples,and then draws a certain number of samples from each classification,and these samples are used as training samples for the agent.The agent model thus trained can not only reduce the training cost of the agent in the later stage of the algorithm,but also reduce the risk of over-fitting problem of the agent to some extent.(2)A Collaborative Surrogate-assisted Social Learning Particle Swarm Optimization(CSSLPSO)algorithm is proposed,which implements a two-stage optimization search.In the global search phase,the algorithm uses all samples to train a collaborative agent model and combines it with a social particle swarm optimization algorithm to complete the search.Following that,it enters the local search phase and uses the front-end search technique to accelerate the convergence.In order to avoid the population falling into local optimum,this paper designs a precision variation strategy,which determines that the population enters the local optimum trap when the variance of fitness of the current population reaches a threshold by calculating the current population,and then enables the use of quadratic variation of distance to change the current population evolutionary state,so as to achieve a new exploration.(3)A comparison experiment between CSSLPSO and mainstream SAEAs algorithms was designed and completed.In terms of benchmark tests,the experimental results of 21 test function sets of CEC2005 and CEC2015 show that CSSLPSO performs well in different dimensions;in terms of application experiments,the effectiveness of CSSLPSO in practical engineering application problems is demonstrated by comparing this algorithm with other mainstream SAEAs in radar phase encoding optimization problems.