An effective offspring generation strategy for many-objective optimization driven by knee points under variable classification

Li-sen Wei,Er-chao Li

DOI: https://doi.org/10.1007/s10489-022-03307-8

IF: 5.3

2022-04-02

Applied Intelligence

Abstract:In many-objective optimization problems, the difficulty of optimization increases as the number of targets increases. When the number of objectives increases, the individuals become extremely sparse in the objective space and the performance of environmental selection strategies weaken. To solve this problem, we proposed an effective offspring generation strategy driven by knee points under variable classification, termed VKOS. In VKOS, there is no mating selection, and the excellent genes of the knee points are used to guide the generation of outstanding individuals with different attributes directionally. Specifically, first, the convergence variables and diversity variables of the problem are obtained by variable classification; then identify the knee points of the current population, and finally propose a parallel mutation method to mutate the current population to obtain offspring. In order to verify the versatility and effectiveness of the strategy, RPEA-VKOS, NSGAIII-VKOS, RPDNSGAII-VKOS, MaOEA/IBP-VKOS compared with the original algorithm on 16 widely used benchmark problems. In addition, taking RPEA as an example, The VKOS proposed in this paper compared with several single mutation operators DE, PLM, NUM and representative solution generation strategies DEMR, VCEM, MM on these benchmark problems. The extensive experiments on well-known benchmark problems ranging from 3 to 15 objectives show that VKOS improves the performance of MaOEAs and has the superior performance over three single mutation operators and typical offspring generation strategies when solving most of these test MaOPs.

computer science, artificial intelligence