LLM-Enhanced Dueling DQN for Multi-Factory Remanufacturing Optimization with Hybrid Disassembly Lines and Drone Delivery

Abstract

This paper addresses a profit-maximization problem in multi-factory remanufacturing that integrates hybrid disassembly lines (straight and U-shaped) with cross-factory drone delivery. The problem is formulated as a mixed-integer program that employs AND/OR graphs to represent disassembly sequencing and incorporates workstation opening and assignment, precedence and conflict constraints, and drone routing costs. To solve the resulting high-dimensional combinatorial problem, we propose the LLM-enhanced Dueling Deep Q-Network (LLM-DUEL), which extends the standard Dueling DQN by incorporating a large language model fine-tuned with low-rank adaptation. The fine-tuned LLM generates feasible disassembly sequences, compressing the reinforcement learning action space, while hierarchical action design and a profit-increment reward mechanism further accelerate policy learning. Experiments on multiple synthetic case sets demonstrate that LLM-DUEL achieves faster convergence, improved stability, and higher objective values compared with DQN, DUEL, and PPO, while closely approaching CPLEX optima on tractable instances. These results suggest that domain-adapted LLMs can substantially enhance reinforcement learning by improving feasibility and efficiency in complex remanufacturing scheduling problems.