Integrating simulation and deep reinforcement learning for multi-objective assembly line design

In smart manufacturing environments, dynamic reconfiguration of facility layouts is increasingly required to cope with fluctuating production demands and complex process flows. This study proposes an integrated optimization framework based on Quantile Regression Deep Q-Network, which simultaneously determines the optimal placement of components, generation and type/size of cells, and installation of conveyors through a multi-network structure. The proposed model collects key production flow indicators at the end of each episode via simulation and incorporates them into the reinforcement learning reward function. This enables not only spatial optimization but also dynamic optimization that reflects the operational efficiency of the process flow. The framework specifically targets the alleviation of bottlenecks between stations by minimizing flow interruptions and delays, thereby enhancing overall line productivity. By adopting an end-to-end learning approach that encompasses both the generation and arrangement of stations and cells, the model overcomes limitations inherent in conventional heuristic-based methods, such as fixed component structures and restricted search spaces.