Evolutionary multitask optimization shows great potential in large-scale multiobjective optimization problems (LSMOPs) by facilitating knowledge transfer across reduced spaces (constructed as auxiliary tasks). However, existing methods often rely on prior knowledge or fixed variable grouping strategies for constructing these auxiliary tasks, lacking adaptability to problem characteristics and population changes during optimization. Their knowledge transfer mechanisms also frequently lack sufficient guidance, potentially introducing misleading information that hinders the search of the original space (viewed as the main task). To address these challenges, this paper proposes a novel evolutionary algorithm. It first utilizes multiple variable importance evaluation methods to quantify the significance of each decision variable. Based on these evaluations, distinct subsets of critical variables are retained to construct multiple low-dimensional auxiliary tasks that adapt to the current stage of the evolutionary process. This approach concentrates computational resources on key variable groups, thereby accelerating convergence. Furthermore, a reference-solution-guided knowledge transfer strategy is introduced to enhance intertask collaboration. This strategy leverages global elite solutions to provide directional guidance for offspring generation, enabling effective transfer of beneficial knowledge from simpler auxiliary tasks to optimize the main one. Extensive experimental results demonstrate the superior efficiency and effectiveness of the proposed algorithm for solving LSMOPs.
Jiang et al. (Mon,) studied this question.