RL Optimizes Production Scheduling in Industrial Coating

Production scheduling in intricate manufacturing environments presents significant challenges, particularly when factors like sequence-dependent setup times, unpredictable disturbances, and strict due-date constraints must be managed simultaneously. While reinforcement learning (RL) has shown promise in academic settings, its industrial applicability has often been limited by reliance on oversimplified benchmark processes. This research bridges that gap by showcasing the effectiveness of RL-based scheduling within a realistic, industry-inspired coating process. This scenario incorporates practical complexities such as setup times that vary based on job sequence, machine breakdowns, and fluctuating machine utilization. The study utilizes the open-source Digital Model Playground (DMPG), a discrete event simulation framework, to model this complex environment and train RL agents. Two standard RL algorithms, Deep Q-Networks (DQN) and Proximal Policy Optimization (PPO), were benchmarked against traditional dispatching rules. The results indicate that RL-based scheduling achieves balanced improvements across key performance indicators, with PPO demonstrating the most robust performance. The core contribution is validating RL in a shareable, realistic industrial scenario and providing a reusable open-source framework for future research and practical application.