Key Technology Research Of Many-objective Evolutionary Algorithms

Many-objective optimization problems(MaOPs)widely exist in the field of scientific research and engineering applications.At present,they have become hot research topic in the field of intelligent information processing.Although reference-point based and Pareto-dominance relation based many-objective evolutionary algorithms(MaOEAs)have become the two most promising techniques for many-objective optimization problems,both have the problem of the high computational complexity.Furthermore,the reference-point based MaOEAs have the difficulty in converging to the Pareto front in high-dimensional objective space and are sensitive to the shape of Pareto front.While the Pareto-dominance relation based MaOEAs have the difficulty in diversity and their parameters are hard to be adjusted before knowing the problem property.Particularly in recent years,more and more constrained MaOPs have arisen and the above algorithms fail to deal with them.Therefore,studying on more effective and practical MaOEAs has great theoretical significance and practical value.For the above problems,the key technologies of objective space transformation,convergence enhancement,diversity promotion and constraint handling in MaOEAs are deeply studied.Many improvement measures are put forward to increase the overall performance of the presented algorithms.The major work and contribution in this dissertation includes the following four aspects.(1)For the the high computational complexity of multi-objective evolutionary algorithms in solving many-objective optimization problems,the technology of objective space transformation is studied and an improved NSGA-III algorithm based on objective space decomposition is proposed.The proposed algorithm decomposes the objective space into several subspaces by the K-means clustering technique and each subproblem is optimized by a separate subpopulation.The numerical experiment results show that the proposed method could reduce complexity and ensure excellent effect.(2)For the problems that the convergence is insufficient and the algorithms are sensitive to the shape of Pareto front in MaOEAs based on reference-points,the technology of the convergence enhancement is studied and an improved NSGA-III algorithm based on adaptive penalty distance is proposed.In the proposed algorithm,an adaptive penalty distance(APD)function is presented to adaptively adjust the importance of convergence and diversity.Numerical simulation results verify the generality of the proposed algorithm.(3)For the problems that the diversity maintenance capability is unsatisfactory and parameters are hard to be adjusted in MaOEAs based on the Pareto-dominance relation,the technology of the diversity promotion is studied,and a many-objective evolutionary algorithms based on hyperplane projection(HPEA)and a many-objective evolutionary algorithms based on angle penalized distance(MaOEA-APD)are proposed,respectively.Where,HPEA falls into the system of the diversity evaluation based on distance information,which uses the hyperplane projection technique to improve the diversity.While MaOEA-APD falls into the system of the diversity evaluation based on angle information,which employs an angle penalized distance to remove the sensitivity parameter.The experimental results show that the two algorithms have more advantage than other algorithms in terms of the overall performance.(4)For the problem that many-objective evolutionary algorithms cannot efficiently combine the high dimensional characteristic and the constraint handling technique,the technology of constraint handling is studied and a reference point-based constrained dominance relation is presented,which regareds the feasible solutions and infeasible solutions as a whole and consider the convergence,the diversity and the feasibility simultaneously.Experimental results show that the proposed algorithm has better performance on convergence and distribution than the other three algorithms.