Solving Size-Agnostic Job Shop Scheduling Problems Like GPT Speaks

Job shop scheduling problem (JSSP) presents a significant challenge in real-world manufacturing scheduling due to the presence of uncertainties and the large scale of production. Reinforcement Learning is an effective methodology for learning scheduling policy by interacting with a simulated job shop scheduling environment. However, the simulated environment is sometimes inaccurate or even unavailable, especially for scenarios with uncertainties in job arrivals and random machine breakdown. To eliminate the dependency on the simulated environment, this paper proposes the Decision-GPT-based Job Shop Scheduling Solver (DGSS). DGSS is trained by offline reinforcement learning where only offline and sub-optimal scheduling trajectories are needed. As a size-agnostic JSSP solver, DGSS combines the size generalization ability of the graph neural network and the simplicity and scalability of the Transformer architecture to model the evolution of the disjunctive graph of a JSSP instance under scheduled operations. Just like the speaking process of GPT, DGSS can generate the next approximate scheduling operation given manually set future rewards, just like the prompt used in GPT. Experiments on simulated JSSP instances show that the proposed DGSS can generate high-quality schedules and outperform the behavior policy and most traditional Priority dispatching rules.