Reinforcement learning-guided multi-objective particle swarm optimization with adaptive topology

Wu, Baotong, Chen, Zhixiang and Demir, Emrah ORCID: https://orcid.org/0000-0002-4726-2556 2026. Reinforcement learning-guided multi-objective particle swarm optimization with adaptive topology. Computers and Operations Research 195 , 107624. 10.1016/j.cor.2026.107624

PDF - Published Version Available under License Creative Commons Attribution. Download (3MB)

Abstract

Multi-objective optimization problems often involve a trade-off between convergence toward the Pareto front and diversity along the front. This study proposes a reinforcement learning-guided multi-objective particle swarm optimization algorithm, called RL-MOPSO, which uses adaptive topology selection to manage this trade-off during the search. The algorithm combines three network topologies: global best, dual-population ring, and Von-Neumann topology with an adaptive jump collaboration operator. Q-learning selects among these topologies dynamically using feedback from convergence and distribution metrics. Computational experiments compare RL-MOPSO with 11 algorithms, including six state-of-the-art methods and five topology-based MOPSO variants. The results show that RL-MOPSO improves both convergence and diversity. The ablation experiments further confirm the contributions of Q-learning and the proposed topology structures.

Item Type:	Article
Date Type:	Publication
Status:	Published
Schools:	Schools > Business (Including Economics)
Publisher:	Elsevier
ISSN:	0305-0548
Date of First Compliant Deposit:	3 August 2026
Date of Acceptance:	21 July 2026
Last Modified:	03 Aug 2026 11:01
URI:	https://orca.cardiff.ac.uk/id/eprint/188680

Actions (repository staff only)

Edit Item