Adaptive Grouping-Based Offspring Generation for Sparse Large-Scale Multimodal Multi-objective Optimization

Sparse large-scale multimodal multi-objective optimization problems (SLMMOPs) are characterized by high dimensionality, sparse effective variables, and multiple disconnected Pareto-optimal regions. These properties create vast and deceptive search spaces, where only a small subset of variables contributes meaningfully to optimal solutions. Existing algorithms often emphasize diversity preservation, but struggle to identify and exploit sparse structures effectively, thus leading to inefficient search and premature convergence. To address this challenge, we propose AGOEA, an Adaptive Grouping-based Offspring Generation Evolutionary Algorithm. AGOEA streamlines the search process through two core innovations. First, it introduces a dual-granularity decision variable grouping strategy. A coarse-grained grouping leverages long-term population statistics to identify globally important variables and differentiate between Pareto sets, while a fine-grained grouping utilizes local feedback to refine variable selection within subpopulations. These grouping insights subsequently guide customized reproduction operators that perform dimension-specific variation to balance exploration and exploitation. Second, AGOEA adopts a multi-population evolution framework, enabling subpopulations to explore distinct sparse structures in parallel. To coordinate these subpopulations and balance computational effort between discovering new Pareto sets and refining sparse variable structures, a dynamic monitoring and resource allocation mechanism is employed. This mechanism adaptively identifies stagnation and redistributes resources to under-explored regions, thereby enhancing global coverage and convergence. Extensive experiments on benchmark SMMOPs with 500 to 1000 decision variables show that AGOEA consistently achieves superior convergence and diversity compared to six state-of-the-art algorithms.