Imitation Learning for Real-Time Job Shop Scheduling Using Graph-Based Representation

Scheduling of manufacturing systems in practice is challenging due to dynamic production environments, such as random job arrivals and machine breakdowns. Dispatching rules are often used because they can be easily applied even in such dynamic manufacturing environments. However, dispatching rules often fail to provide a satisfactory production schedule because they cannot consider overall system states when assigning jobs. Therefore, we develop a real-time scheduling method using imitation learning, especially behavior cloning, to solve job shop scheduling problems. We define a set of available actions, a target optimal policy, and a dynamic graph-based state representation method for imitation learning. The proposed method is size-agnostic, which then can be applied to unseen larger problems. The experimental results show that the proposed method performs better to minimize makespan than other dispatching rules in dynamic job shops.