User-friendly assembly line optimization: A hierarchical approach combining simulation-evaluated heuristics and reinforcement learning

Abstract Efficient assembly line optimization is essential for improving productivity and cost competitiveness in modern manufacturing systems. Despite its importance, many existing approaches neglect the limited number of resources and workers available in real production environments, leading to solutions that are difficult to deploy in practice. This study proposes a hierarchical framework that tightly integrates human intent interpretation, logical system configuration, and physical layout optimization. A large language model fine-tuned with low-rank adaptation translates user-defined objectives and constraints expressed in natural language into quantitative optimization goals. Logical configurations are optimized using metaheuristic methods, while physical layouts are refined using implicit quantile network (IQN)–based deep reinforcement learning. The IQN model captures the full return distribution through implicit quantile sampling, enabling more flexible policy learning under stochastic environments. Discrete-event simulation is used to evaluate key performance indicators and refine the system configuration. Simulation results demonstrate that the proposed framework generates feasible and robust assembly line designs with improved practical deployability.