Tardy jobs minimization in a dynamic job-shop environment using double Q-Learning Algorithm

Job-shop scheduling problems are complex optimization challenges frequently encountered in manufacturing and production systems, especially in environments with high market demand and customized production modes. The goal of this study is to minimize delays in a dynamic job-shop environment using reinforcement learning. Specifically, we introduce a Double Q-Learning (Double-QL) approach to dynamically manage operations, accounting for varying processing times, unpredictable job arrivals, and machine constraints. A new state representation has been developed to accommodate workshops of all sizes, along with a novel reward function tailored to the agents’ specific contexts. Transfer learning is utilized by leveraging the Q-tables learned from previous environmental scenarios. The simulation was carried out with four machines and varying job limits. To assess the robustness of our approach, we compared our Double-QL algorithm with dispatching rules related to tardiness and evaluated its performance in multiple scenarios.