Enhancing leadership-based metaheuristics using reinforcement learning: A case study in grey wolf optimizer

Metaheuristics are widely applied in optimization because of their flexibility and ability to address complex and high-dimensional problems. Nevertheless, they face persistent challenges, including susceptibility to local optima, limited parameter adaptability, and premature convergence. Leadership-based metaheuristics, in which leaders guide the search process, encounter additional difficulties such as limited exploration capacity, leader stagnation, and reduced diversity, often stemming from underutilization of data generated during the search. To overcome these limitations, this study proposes a reinforcement learning–based approach, RL-LGWO, which enhances the Grey Wolf Optimizer (GWO) by integrating multi-agent reinforcement learning. In RL-LGWO, agents share experiences to improve decision-making, and reinforcement learning is employed to decouple and adapt the leader update mechanism, thereby improving the exploration–exploitation balance and enabling leaders to dynamically escape local optima. The proposed method was evaluated against two GWO-enhancing algorithms, three RL-based GWO variants, PSO, WOA, and the original GWO across 23 well-known benchmark functions, in addition to the recent CEC2022 benchmark suite. Experimental results show that RL-LGWO achieved the best solutions on 17 of the 23 benchmark functions, with superior convergence speed and improved stability, while incurring only a minor runtime increase compared with the original GWO. Furthermore, on the CEC2022 suite, RL-LGWO outperformed competing algorithms on 10 of 12 test functions, underscoring its robustness and adaptability to recent and challenging benchmarks. Overall, the findings indicate that RL-LGWO delivers a substantive improvement over state-of-the-art alternatives and holds strong potential to advance leadership-based metaheuristics for a wide range of optimization problems.