.png){kind=icon}
.png){kind=icon}
Return
Large-Scale Sparse Multi-Objective Optimization via Customized Individual Learning
.png){kind=icon}
PRE
AI
DOI：10.1109/TETCI.2026.3670899.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Large-scale sparse multi-objective optimization problems (LSSMOPs) have attracted increasing attention due to their widespread presence in real-world applications, where optimal solutions typically exhibit inherent sparsity—most decision variables are zero. Traditional sparse multi-objective evolutionary algorithms (MOEAs) often focus on exploiting sparsity in the decision space and overlook the synergies and diversity in the population, impairing the search efficiency and directional guidance in high-dimensional sparse spaces. To address this limitation, this paper proposes a novel approach, termed Sparse Multi-objective Optimization via Customized Individual Learning (SMOCIL). SMOCIL introduces a Quad-Population Partition strategy that divides the population into four subgroups based on individual sparsity and solution quality, effectively capturing inter-individual differences. A Customized Individual Learning mechanism is then applied to refine both masks and decision variables in a subgroup-specific manner, enhancing the exploration and exploitation of promising sparse regions. Extensive experiments on benchmark problems and real-world LSSMOPs demonstrate that SMOCIL significantly outperforms seven state-of-the-art sparse MOEAs in terms of convergence speed and solution quality. Furthermore, ablation studies confirm the essential contribution of the individual learning strategies to the algorithm's overall effectiveness.
Keywords:
Evolutionary computation
sparse large-scale optimization
population partition
genetic algorithm
Journal
I
IF:
6.5
Papers:
1.4K
Citations:
4.5K