Analysis And Design Of Large-Scale Global Optimization Algorithms

Large-scale global optimization problems represent a complex and challenging class of optimization problems.To tackle such problems,numerous evolutionary algorithms and swarm intelligence algorithms have been developed in the field of large-scale global optimization.However,these algorithms face significant challenges in solving large-scale global optimization problems.This is mainly due to the size of the solution space growing exponentially with the increase of variables.It is difficult for algorithms to explore more solution space within the specified computational cost.Additionally,the number of local optimums also increases exponentially,leading to the methods being prone to fall into local optimum.Large-scale global optimization algorithms based on differential evolution are able to solve such optimization problems effectively.SHADE-ILS,as one of the competitive algorithms,may frequently fall into a specific local optimum and stagnate in solving largescale global optimization problems.In such cases,it is necessary to enhance the parameter adaptive capability as well as to improve the population diversity.Moreover,current largescale global optimization methods often require benchmark problems to verify performance.It is unclear whether these problems are representative or how representative the benchmark problems are.The research work in this thesis has two main aspects:(1)An adaptive differential evolution with extended historical memory and iterative local search is proposed to alleviate the tendency of SHADE-ILS to fall into local optima and stagnation.Firstly,an extended historical memory is proposed that increases the probability of generating successful individuals and improves parameter adaptive ability by combining three ways of generating control parameters.Secondly,a permanent record is proposed to enhance population diversity.Finally,they are applied to SHADE and combined with the iterative local search(ILS)for solving large-scale global optimization problems.Extensive numerical experiments demonstrate that our method can effectively improve parameter adaptivity and population diversity,as well as effectively mitigate the algorithm from falling into local optima and stagnation.The proposed algorithm effectively addresses large-scale optimization problems,particularly fully-nonseparable problems.(2)This thesis also proposes a representativeness metric to address the lack of "representativeness metrics" for current large-scale benchmark problems,and provides a specific representative analysis as an example for single-objective unconstrained optimization problems and large-scale benchmark problems.To explore representativeness,this thesis analyzes the benchmark problems and defines three types of representativenessmeasuring problems.Based on the third type representatives-measuring problems,we propose a Type-III representative-measuring method,which is able to explore the representativeness of all the available problems.It is experimentally verified that the proposed framework can measure the representativeness of the benchmark problem.