Reinforcement Learning for Dynamic Job Shop Scheduling: A Maskable PPO Approach

This thesis addresses the Dynamic Job Shop Scheduling Problem (DJSSP), a critical challenge in modern manufacturing characterized by unpredictable arrivals, strict deadlines, and routing flexibility. Traditional heuristics often lack adaptability in such dynamic environments. To overcome these limitations, a reinforcement learning framework based on Maskable Proximal Policy Optimization (PPO) is proposed. Key features include multi-discrete action spaces for parallel machine decisions, action masking for feasibility, and a multi-objective reward function balancing lateness reduction, on-time delivery, throughput, and operational stability. Domain randomization during training enhances generalization across varying conditions. A real case study validates the framework, comparing two PPO agents with classical heuristics (EDD, FIFO, LPT, SLACK) and a metaheuristic (Genetic Algorithm). Results show that RL agents consistently outperform classical heuristics, achieving lower lateness, higher on-time completion, and improved continuity. When compared to the Genetic Algorithm, the RL agents demonstrate overall comparable or superior performance across most scenarios, confirming their adaptability and robustness. The dynamic agent, in particular, exhibits stronger generalization under heterogeneous conditions, reinforcing the advantages of generalized training. These findings highlight reinforcement learning as a scalable and versatile solution for industrial scheduling.