摘要
In order to exploit the enhancement of the multiobjective evolutionary algorithm based on decomposition(MOEA/D), we propose an improved algorithm with uniform design(UD), i.e. MOEA/D-UD. Three mechanisms in MOEA/D-UD are modified by introducing an experimental design method called UD. To fully employ the information contained in the domain of the multi-objective problem, we apply UD to initialize a uniformly scattered population. Then, motivated by the analysis of the relationship between weight vectors and optimal solutions of scalar subproblems in the study of MOEA/D with adaptive weight adjustment(MOEA/D-AWA), a new weight vector design method based on UD is introduced. To distinguish real sparse regions from pseudo sparse regions, i.e. discontinuous regions, of the complex Pareto front, the weight vector adjustment strategy in MOEA/D-UD adequately utilizes the information from neighbors of individuals. In the experimental study, we compare MOEA/D-UD with three outstanding algorithms, namely MOEA/D with the differential evolution operator(MOEA/D-DE), MOEA/D-AWA and the nondominated sorting genetic algorithm II(NSGA-II) on nineteen test instances. The experimental results show that MOEA/D-UD is capable of obtaining a well-converged and well diversified set of solutions within an acceptable execution time.
In order to exploit the enhancement of the multi- objective evolutionary algorithm based on decomposition (MOEA/D), we propose an improved algorithm with uniform de- sign (UD), i.e. MOEA/D-UD. Three mechanisms in MOEA/D-UD are modified by introducing an experimental design method called UD. To fully employ the information contained in the domain of the multi-objective problem, we apply UD to initialize a uniformly scattered population. Then, motivated by the analysis of the re- lationship between weight vectors and optimal solutions of scalar subproblems in the study of MOEND with adaptive weight ad- justment (MOEA/D-AWA), a new weight vector design method based on UD is introduced. To distinguish real sparse regions from pseudo sparse regions, i.e. discontinuous regions, of the complex Pareto front, the weight vector adjustment strategy in MOEMD-UD adequately utilizes the information from neighbors of individuals. In the experimental study, we compare MOEA/D-UD with three outstanding algorithms, namely MOEA/D with the dif- ferential evolution operator (MOEA/D-DE), MOEA/D-AWA and the nondominated sorting genetic algorithm II (NSGA-II) on nineteen test instances. The experimental results show that MOEA/D-UD is capable of obtaining a well-converged and well diversified set of solutions within an acceptable execution time.