Smart Master Production Scheduling by Deep Reinforcement Learning: An Exploratory Analysis

By providing a consolidated view of demand, production capacity and inventory status, the master production schedule (MPS) enables collaborative networks to be optimised, planned and coordinated, which leads to more robust strategic decision making. In this context, flexibility, agility and automation are key elements to consider, and make machine-learning methods strong candidates for problem modelling. This paper researches some modelling alternatives to address the MPS problem subject to the eventuality of unsatisfied demand due to production capacity and inventory constraints by the deep reinforcement learning (DRL) method. The modelling of observation space, action space and reward function, and the choice of the DRL algorithm, among some of those currently at the forefront of the technique, are analysed and discussed. The sensitivity to the problem dimension, as a function that depends mostly on the number of considered product references, is also approached. As a major contribution, this study constitutes a valuable step prior to final modelling and subsequent validation by clearing the way for a wide range of possible implementations.