Self-adapting WIP parameter setting using deep reinforcement learning

This study investigates the potential of dynamically adjusting WIP cap levels to maximize the throughput (TH) performance and minimize work in process (WIP), according to real-time system state arising from process variability associated with low volume and high-variety production systems. Using an innovative approach based on state-of-the-art deep reinforcement learning (proximal policy optimization algorithm), we attain WIP reductions of up to 50% and 30%, with practically no losses in throughput, against pure-push systems and the statistical throughput control method (STC), respectively. An exploratory study based on simulation experiments was performed to provide support to our research. The reinforcement learning agent's performance was shown to be robust to variability changes within the production systems.