DIGA: a reinforcement learning-driven hybrid metaheuristic for flexible job-shop scheduling

Purpose This study aims to develop a novel hybrid Genetic Algorithm (GA) for Flexible Job-shop Scheduling Problems (FJSP) to improve makespan. Design/methodology/approach DIGA, a Reinforcement Learning (RL)-driven algorithmic framework is proposed. This algorithmic framework consists of GA, Iterated Greedy Algorithm (IGA) and Double Q-Learning Algorithm (DQLA), which is an RL-based technique. IGA was used as the local search technique to improve the performance of GA. The DQLA tuned GA parameters – selection strategy, crossover rate and mutation rate – to their proper values so that it resulted in better makespan. Brandimarte benchmark instances were solved using this algorithm and statistical tests were carried out to assess its performance. Comparison with results from recent literature was also made. Findings Results show that DIGA is a statistically significant, stable, consistent and robust algorithm that can be used to solve FJSP. The tournament selection strategy was chosen by DQLA on vast majority of occasions, followed by Roulette wheel and Ranking strategies. The presence of IGA as local search technique has considerably reduced the number of crossover and mutation operations needed, helping faster convergence to the best makespan. Originality/value Tuning of all three GA parameters for FJSP was never reported in literature. Proposed DIGA is a novel algorithmic framework.