A Reinforcement Learning Approach for Production Control in Manufacturing Systems

Abstract. The problem of production control in serial manufacturing lines that consist of a number of unreliable machines linked with intermediate buffers is addressed. We make use of Reinforcement Learning methodologies in order to derive efficient control policies. Our aim is to derive control policies that are more state-dependent and therefore more efficient than well-known pull type control policies such as Kanban. Manufacturing systems of this type are studied under average measures such as average WorkInProcess inventories etc. and thus, a learning algorithm from the currently developing field of Average Reward Reinforcement Learning was applied. The Reinforcement Learning control policy was compared to three existing efficient pull type control policies, namely Kanban, Base Stock and CONWIP on the basis of simulated data and found to outperform them. The simulation experiments involved a single-product system with two machines that allows backordering. Numerical results are presented along with a qualitative interpretation of our findings. The paper concludes with directions for future research. 1