Reinforcement Learning-Based WIP Control for Balancing Productivity and Lead Time in Manufacturing Systems

In the era of Industry 4.0 and mass customization, manufacturing systems face the challenge of managing increased product variety and demand variability. Effective Work-In-Progress (WIP) control is crucial for maintaining optimal productivity and lead time in these dynamic environments. This paper proposes a novel adaptive WIP control approach for semi-heterarchical manufacturing systems using Deep Reinforcement Learning (DRL). The proposed approach leverages a Deep Q-Network (DQN) agent to learn optimal WIP control policies through interaction with a stochastic simulation environment. The DQN agent considers the current system state, processing time variability, and throughput targets to make real-time decisions and dynamically adjust WIP levels. The problem is formulated as a Markov Decision Process (MDP), and the DQN agent is trained using a custom simulation environment developed with the Simpy library. The experimental results validate the performance and adaptability of the proposed approach under different production scenarios and variability levels. The DQN-based WIP control approach demonstrates its ability to maintain the desired throughput while minimizing WIP levels, leading to improved overall performance of the manufacturing system. This research contributes to the advancement of intelligent manufacturing and provides a data-driven solution for adaptive WIP control in semi-heterarchical production systems.