The Research On Adaptive Multi-objective Evolutionary Algorithm Based On Decomposition

There are many multi-objective optimization problems in scientific research and engineering practice. Usually, these objectives conflict with each other and the solving of multi-objective optimization problem is to find the optimal solution under different weights in each objective. Traditional optimization algorithms usually decompose a multi-objective optimization problem into a set of single optimization sub-problems, and then solve the sub-problems one by one. However, this traditional method has the low efficiency, mainly because the algorithm only gets one solution from each run and each sub-problem is solved independently. In order to effectively solve multi-objective optimization problems, evolutionary algorithm, which is inspired by the natural evolution and “the survival of the fittest” theory, has been proposed and demonstrated great potentiality to solve multi-objective optimization problems. Multi-objective evolutionary algorithm based on decomposition(MOEA/D) is one of the outstanding algorithms, the core idea of which is to use a set of uniform weight vectors to decompose multi-objective optimization problems into a set of single-objective optimization sub-problems and then solve the sub-problems using the idea of evolution. All the sub-problems can exchange their superior information in MOEA/D. Therefore, the algorithm can obtain a set of solutions in each run to improve the efficiency significantly. However, similar with other multi-objective evolutionary algorithms, MOEA/D also suffers from some defects, such as the high sensitivity on mutation and crossover, slow convergence, and uneven distribution. Therefore, in order to further enhance its efficiency, more and more researchers pay attention to the study of MOEA/D in recent years.This paper starts from the analysis of the sensitivity of MOEA/D on mutation and crossover operators, and then improves the versatility, efficiency and robustness of MOEA/D. After the investigation of the original MOEA/D algorithm and some improved variants, two improved MOEA/D variants are accordingly presented. The main contributions of this paper are listed as follows.(1) An adaptive composite operator pools selection and parameter control strategy for MOEA/D is proposed and this algorithm is named MOEA/D-CDE. This algorithm designes a kind of composite operator pools and the corresponding adaptive selection strategy, and the adaptive parameter control for all the operator pools. MOEA/D-CDE performs significantly better on solving some complicated test problems, such as UF and WFG test problems.(2) An adaptive mixing operator selection strategy based on gene level is also designed for MOEA/D and this algorithm is named MOEA/D-GDE. This algorithm is used to evolve each dimensionality of individuals in the whole population and designs an adaptive operator control strategy based on gene level to evolve the dimensionality of individuals. By this way, the diversity and convergence of the population can be greatly promoted. This strategy is validated experimentally to further enhance the performance of MOEA/D in solving some complicated multi-objective optimization problems.