PPO-based Dynamic Flexible Job-Shop Scheduling

With the development of artificial intelligence technologies, intelligent production management systems greatly enhance the competitiveness of an organization or supply chain when facing challenges such as flexibility, dynamic operations, multi-threading and complex rules. However, meta-heuristic algorithms often perform poorly when dealing with dynamic job shop scheduling problems. In this context, we employ an improved Proximal Policy Optimization (PPO) algorithm to solve complex dynamic flexible job shop scheduling problems and compare it with IMPROVED FCFS and Dynamic SPT. On this basis, we perform static global optimization for the same tasks, i.e., in the elimination of dynamic arrivals, all the task data are provided, and the global optimization is performed using a genetic algorithm to obtain the theoretically optimal scheduling plan. Although the PPO algorithm is prone to local optimality in large-scale task scheduling, its total completion time is close to that of the genetic algorithm, and it outperforms the previous two methods in terms of overdue tasks, showing great potential for application. The results show that improved FCFS and Dynamic SPT, as relatively primitive methods, do not perform well, although they still have some problem solving ability. Deep reinforcement learning algorithms show promising applications in dynamic and flexible job shop scheduling problems due to their ability to effectively handle complex events through learning.