Decomposition-based multi-objective reinforcement learning for dynamic disassembly job shop scheduling with urgency guidance

The dynamic disassembly job shop scheduling problem (DDJSSP) entails organizing multiple jobs with distinct requirements across machines, where job operations are subject to sequence constraints and urgency conditions. Existing deep reinforcement learning techniques for multi-objective job shop scheduling problems (JSSP) with manually weighted reward functions result in a single policy, limiting the ability to approximate multiple policies in the Pareto front. To minimize both makespan and total energy consumption in DDJSSP, we propose an urgency-driven decomposition-based multi-objective reinforcement learning (UD-MORL) approach. We simulate a dynamic scheduling environment reflecting real-world complexities by introducing uncertain processing times, random employee absences, and an urgency rate for orders. We then develop a decomposition approach to separate objectives by adjusting weights and iterating policies based on performance and information entropy metrics. Finally, we employ a mutual information mechanism to identify the weight combination exhibiting the strongest correlation with population points, thereby improving weight-fitting efficiency. Experimental results on public general-purpose JSSP datasets show UD-MORL outperforms existing multi-objective reinforcement learning algorithms in hypervolume and sparsity, achieving an average hypervolume improvement, sparsity reduction, and a win-rate of 55% across all benchmark instances.