Reinforcement Learning Based Acceptance Criteria for Metaheuristic Algorithms

This paper proposes novel reinforcement learning-based acceptance criteria for metaheuristic algorithms. We develop Q-learning and Deep Q-learning-based acceptance criteria and integrate them into simulated annealing (SA) and artificial bee colony (ABC) algorithms. Also, we design two versions of these novel acceptance criteria: the online version and the offline version of Q-learning and deep Q-learning based acceptance criteria. The online version starts to train itself and to make decisions to accept or reject the candidate solution with the start of the metaheuristic. The offline version uses a trained and well-tuned Q-learning and deep Q-learning based acceptance criteria to make accept/reject decisions. Our experimental study compares our proposed acceptance criteria with existing ones, such as fuzzy rule-based acceptance (FRBA) and simulated annealing-like acceptance (SALA) criteria. The experiment reveals that metaheuristics with deep Q-learning-based offline acceptance criteria outperform metaheuristics with existing acceptance criteria and other variants in this study.