Matrix-Based Role Adaptive Particle Swarm Optimization for Large-Scale Optimization

초록

Evolutionary computation approaches often face two major challenges in large-scale global optimization (LSGO): 1) the difficulty of balancing exploration and exploitation 2) the heavy computational cost induced by high dimensionality. Therefore, this article proposes a novel matrix-based role adaptive particle swarm optimization (MRPSO) for LSGO. Designed to address the twin challenges of balancing exploration-exploitation and managing high computational cost, MRPSO introduces three key innovations within an efficient matrix-based evolutionary framework. First, an adaptive role allocation mechanism dynamically divides the swarm into two specialized groups: an explorer group tasked with maintaining population diversity, and an exploiter group dedicated to accelerating convergence. Second, a particle convergence indicator (PCI) facilitates communication between these groups, signaling when exploration should be intensified. Third, new update processes are defined as exploiters learn from superior explorers while explorers initiate self-reconstruction into unexplored regions when PCI indicates diminished optimization potential. Extensive experiments conducted on the CEC’2013 LSGO benchmark and real-world feature selection applications demonstrate that MRPSO outperforms eight state-of-the-art algorithms. Furthermore, the theoretical convergence of the proposed algorithm is analyzed, and the runtime analysis confirms its significantly superior computational efficiency, particularly in large-scale optimization scenarios.

키워드