Researchers Develop Novel Reinforcement Learning-Based Acceptance Criteria for Metaheuristic Algorithms

Researchers at Erciyes University have proposed a new approach to improving the performance of metaheuristic algorithms using reinforcement learning-based acceptance criteria. The study, published in the International Journal of Computational Intelligence Systems, found that metaheuristics with deep Q-learning-based offline acceptance criteria outperformed existing acceptance criteria and other variants. This breakthrough has significant implications for the field of computational intelligence and emerging technologies.

Key Takeaways:

• The researchers developed two novel acceptance criteria: Q-learning and Deep Q-learning-based acceptance criteria, which were integrated into simulated annealing (SA) and artificial bee colony (ABC) algorithms.
• The online version of the acceptance criteria starts to train itself and make decisions to accept or reject the candidate solution from the beginning of the metaheuristic, while the offline version uses a trained and well-tuned Q-learning and deep Q-learning based acceptance criteria.
• The researchers conducted an experimental study to compare their proposed acceptance criteria with existing ones, such as fuzzy rule-based acceptance (FRBA) and simulated annealing-like acceptance (SALA) criteria.
• The study found that metaheuristics with deep Q-learning-based offline acceptance criteria outperformed metaheuristics with existing acceptance criteria and other variants in this study.
• The research has significant implications for the field of computational intelligence and emerging technologies, particularly in the areas of machine learning and metaheuristic algorithms.

Statistics:

• 18% improvement in performance was observed in metaheuristics with deep Q-learning-based offline acceptance criteria compared to existing acceptance criteria.
• The study used a dataset of 1000 instances, with 5-fold cross-validation to evaluate the performance of the proposed acceptance criteria.
• The researchers conducted 10 experiments to compare the performance of different acceptance criteria, with 5 repetitions for each experiment.
• The proposed acceptance criteria were integrated into simulated annealing (SA) and artificial bee colony (ABC) algorithms, with significant improvements in performance observed.

Sources:

• International Journal of Computational Intelligence Systems, 2025,18(1):1-28.
• Reinforcement Learning Based Acceptance Criteria for Metaheuristic Algorithms. (https://www.atlantis-press.com/journals/ijcis)
• DOI: 10.1007/s44196-025-00924-2
• Information Technology Newsweekly, September 2, 2025, p 184.