A deep reinforcement learning approach for the throughput control of a flow-shop production system

This paper proposes a new method for controlling a flow shop in terms of throughput and Work In Process (WIP). In order to achieve a throughput target, a Deep Q-Network (DQN) is used to define the constant WIP quantity in the system. The main contribution of this paper is the novel approach used to formulate the state, action space, and reward function. An extensive preexperimental campaign is conducted to determine the best network structure and appropriate hyperparameter values. Finally, the system's performance is compared to the known results of an analytical model from the literature (Practical Worst Case, PWC).