Adaptive dual-space discriminator for large-scale multi-objective optimization

Numerous real-world applications involve large-scale multi-objective optimization problems (LSMOPs) with hundreds or even thousands of decision variables. Although multi-objective evolutionary algorithms (MOEAs) are effective for such tasks, relying solely on objective-space information often provides insufficient selection pressure in high-dimensional settings. To address this issue, an adaptive dual-space discriminator is proposed to coordinate selection strategies from both the objective and decision spaces through a state-aware adaptive operator-selection method with a recurrent selector based on direct population-improvement feedback, and a corresponding algorithm termed MOEA-RLD is developed by integrating dual-space selection, online strategy adaptation, a bounded population pool, and APD environmental selection. Comparative evaluations on the LSMOP benchmark suite against five representative large-scale multi-objective evolutionary algorithms show that MOEA-RLD achieves favorable average rankings across the tested settings. Further evaluations on three real-world application settings provide evidence of both the practical performance of the proposed method and its limitations.