Insight-driven Optimization to Improve Dynamic Scheduling for Flexible Job-shops

Planning and scheduling problems can be challenging to manage, and the related optimization problem is often too complex to solve in real-time. The typical approach to address this issue is to apply heuristic policies, which perform reasonably well. Machine learning algorithms can be used to replace heuristics, but this raises the issue of obtaining unexplainable solutions. The application of reinforcement learning with policy extraction technique can help produce explainable results to improve the dynamic system of the planning and scheduling problem. As a case study, we use the simulation model of a dynamic flexible job shop to learn a solution strategy for the associated scheduling problem with deep reinforcement learning. Based on this, we were able to extract a decision tree that is superior to classical dispatching heuristics for a wide variety of scenarios. It also provides valuable information about the criteria for decision-making. Here we offer a new approach to analyzing and solving these problems - leading to an improvement in dynamic systems.