Graph-based Reinforcement Learning for Flexible Job Shop Scheduling with Transportation Constraints

Recently, deep reinforcement learning (DRL) has been employed in flexible job-shop scheduling problems (FJSP) to minimize makespan within flexible manufacturing systems (FMS). In practice, numerous modern enterprises are incor-porating automated guided vehicles (AGV) into their FMS implementations. However, existing DRL-based FJSP solutions do not account for transportation constraints. To tackle this practical issue, we propose a novel graph-based DRL method, called Heterogeneous Job Scheduler (HJS), which interprets the environment status using the graph structure and then training the DRL model based on graph embeddings. Our findings indicate that the proposed approach surpasses conventional dispatching rules and existing DRL-based methods in terms of makespan, running time, and generalization performance.