Optimizing Dynamic Cellular Manufacturing System: A Deep Reinforcement Learning Approach to Profit Maximization and Inventory Management

This paper presents a novel dynamic cellular manufacturing system that incorporates order rejection, tardiness costs, and the costs of purchasing and holding raw materials. Orders arrive at different times, requiring real-time decisions on acceptance or rejection. Accepted orders necessitate raw materials, which must be procured at the optimal time. A mathematical model with two objectives—maximizing profit and minimizing the number of rejected orders—is proposed. To solve this problem, an iteration-based hierarchical solution method with three steps has been developed. First, machines are assigned to cells using a genetic algorithm. Next, a deep reinforcement learning (DRL) algorithm with a double network is employed to manage order acceptance, schedule operations for accepted orders, assign them to the most suitable machines, procure raw materials, and determine optimal safety stock levels. The inventory management within the DRL framework is further supported by an artificial neural network. Finally, a boxing match algorithm is introduced to optimize machine placement based on DRL outputs. A case study was conducted to evaluate the performance of the proposed method, and comparative results between real-world data and the algorithm’s results demonstrate the effectiveness of the proposed approach.