The article introduces PetriRL, a new framework for resolving job shop scheduling problems (JSSP) by integrating Petri nets and reinforcement learning. The approach improves the explainability of neural networks and allows direct incorporation of raw data without preprocessing into disjunctive graphs. The Petri net controls automated components of the process, enabling the agent to focus on critical decision-making. The integration of event-based control and action masking in this approach produces competitive results on public test benchmarks. The method demonstrates robust generalizability across various instance sizes and uses the Petri net’s graph nature to dynamically add job operations during the inference phase without needing agent retraining, thereby enhancing flexibility.
Publication date: 2 Feb 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2402.00046
