Improvement Of Two Many-objective Evolutionary Algorithms

High-dimensional multi-objective evolutionary algorithm is a valuable and difficult research topic in multi-objective evolutionary algorithm.In recent years,classical multi-objective algorithms have been improved for high-dimensional multi-objective optimization.Such as NSGA-? based on NSGA-II and algorithm based on MOEA/D.However,when these algorithms are used to optimize high-dimensional multi-objective problems,there are various defects in both convergence and distribution,and their performance is reduced to different degrees.On the basis of previous research,I have done some research and improvement on several typical high-dimensional multi-objective evolutionary algorithms.Specific as follows:(1)For the classical MOEA/D,the normal distributed crossover operator(NDX)and the differential evolution operator(DE/BEST/2)are combined together to replace the analog binary crossover operator(SBX).The convergence of the algorithm is explored from the perspective of update operator.The DE/Best/2 operator is used to recombine a large number of individuals in the subproblem neighborhood,so that the parent generation individuals have more diversity,so as to generate diverse individuals,expand the diversity of the population,and provide a larger search range.Search using the NDX operator.The numerical simulation results show that NDX can search for a wider range of solutions than SBX,and it is easy to jump out of the local optimum,and can obtain a Pareto solution set with better distribution and uniformity.(2)In the NSGA-? algorithm,the adaptive information feedback mechanism is introduced into the NSGA-?,and the adaptive information feedback NSGA-?algorithm is proposed.This model can be improved to solve the problem of insufficient convergence of NSGA-? algorithm when solving large-scale high-dimensional multi-objective optimization problems.The basic ideas of the adaptive information feedback model are as follows:firstly,the KTH individual in the parent generation is selected adaptively by using the gradient of the objective function;Then the parent population was processed by NSGA-? to obtain the intermediate individuals.Then the intermediate individual and the KTH individual are weighted average to get the next generation of individuals.The test results show that the adaptive information feedback NSGA-? can further improve the convergence of the algorithm when solving large-scale high-dimensional multi-objective optimization problems.The higher the target dimension m is,the more obvious the improvement of the convergence will be.(3)In order to improve the local optimality of Pareto solution of NSGA-III and improve the efficiency of the algorithm,this paper proposed to combine NSGA-? with a simulated annealing algorithm(SA)in neighborhood search,and developed an NSGA-III based on SA.In this paper,the concept of neighborhood search is introduced to improve the search ability of NSGA-? and the efficiency of the algorithm in terms of local optimization.By using Binh4 and Viennet 4 test functions,compared with NSGA-III,the efficiency of SA-NSGA-? is significantly improved,and on the basis of constant diversity,the obtained solution can be closer to the Pareto optimal solution.