Surrogate Model-assisted Evolutionary Algorithms For Global Optimisation

Real-valued optimisation is widely applied in real world.As a population-based stochastic search method,evolutionary algorithms(EAs)have attracted considerable at-tentions in the field of real-valued optimisation.Nevertheless,EAs still face challenges in real-world applications.Since the selection mechanism in EAs,i.e.,survival of the fittest,is based on individuals' fitness,EAs usually need a lot of fitness evaluations to achieve a satisfactory solution.However,computationally expensive problems(CEPs)exist widely in the real world.For such problems,the cost of one fitness evaluation is very high,which restricts the application of EAs in relevant areas.Motivated by this,the goal of this dissertation is to research and design efficient EAs for real-valued CEPs.To achieve this goal,this dissertation employs computationally efficient models(usually called surrogate models or meta-models)to evaluate individuals instead of carrying out real fitness evaluations,which can enable EAs to achieve better solutions within limited computational resources.Specifically,the main research contents and contributions of this dissertation can be summarised into the following four aspects:1.Classification-assisted differential evolution(DE)Surrogate model-assisted EAs involve the interplay between learning and opti-misation.Generally speaking,the appropriate choice of surrogate model and the best way to use them depend on the underlying EA.Through the analysis of the search characteristics in DE,this dissertation formulates the selection process in DE as a classification problem.Based on this,this thesis proposes incorporating classification techniques into DE to reduce the need of real fitness evaluations.Different from regression models and ranking models that are usually employed in the field of surrogate model-assisted EAs,classification models can better match the essence of the selection mechanism of DE,and thus can better improve the performance of DE on solving CEPs within limited computational resources.2.Classification-and regression-assisted DE for CEPs There are different types of surrogate models.Different surrogate models aim at solving different problems,and thus can play different roles in EAs.Having this in mind,this thesis proposes incorporating both classification and regression tech-niques into DE for solving CEPs.In this proposed method,classification models can help DE avoid wasting real fitness evaluations on bad offsprings.Further-more,regression models can give good offsprings approximate fitness instead of doing real fitness evaluations.This can thus further reduce the number of real fitness evaluations needed by DE in every generation.3.Surrogate model-assisted self-adaptive DE The performance of DE highly depends on the choice of trial vector generation strategy and the setting of its control parameters.Self-adaptation mechanism plays an important role in the field of EAs.It can help EAs to automatically adjust evolutionary operators and settings of control parameters during the search process,and thus can improve EAs' performance.Based on the analysis of the disadvantages of existing self-adaptive DE variants in solving CEPs,this disser-tation proposes a surrogate model-assisted self-adaptation scheme for DE,which can suit the CEPs better.4.Evolutionary optimisation with hierarchical surrogates There are different modelling techniques.A modelling technique may have dif-ferent modelling performance on different problem landscapes.Therefore,the choice of modelling technique plays an important role in affecting the perfor-mance of surrogate model-assisted EAs.Generally speaking,one should select the appropriate modelling technique based on the properties of the underlying application problem.However,this might require some priori knowledge of the underlying problem that is not easy to be obtained beforehand,and thus is im-practical.Considering this,this dissertation proposes a novel scheme to adapt the surrogate modelling technique in the framework of memetic algorithm(MA),which differs from existing approaches in employing a hierarchical structure of surrogate models.In summary,this dissertation carries out work from two aspects:evolutionary al-gorithm and surrogate modelling technique.From the perspective of EA,consider-ing that surrogate models can play different roles in EAs,this dissertation proposes classification-assisted DE,classification-and regression-assisted DE,and surrogate model-assisted self-adaptation scheme for DE.From the perspective of surrogate modelling,considering that a modelling technique may model differently on different problem landscapes,this dissertation proposes an efficient self-adaptation scheme with a hierar-chical structure to automatically adjust the surrogate modelling technique in the frame-work of MA during the evolutionary process.The proposed classification-assisted DE,and classification-and regression-assisted DE algorithms are expected to be helpful for future work on combining surrogate model and EAs with a pairwise selection strat-egy.The proposed surrogate model-assisted self-adaptation framework can be com-bined with other DE variants with multiple search strategies.The proposed surrogate modelling adaptation scheme is expected to be helpful for future research on simulta-neous use of multiple surrogate models in EAs.