Scaling Reinforcement Learning

Online Job Shop Scheduling in Flexible Manufacturing Systems to Various Products

Abstract

This article presents results of a reactive scheduling solution for flexible manufacturing systems. The number of product variants that can be produced in flexible manufacturing systems can be extremely large and results in a huge state space. Therefore, in this work, the authors focus on the generalization aspect of our solution to scale it to a variety of products. For this purpose, a state design with an advanced job specification encoding was developed and implemented. A multi-agent version of Deep Q-Networks is used to control the navigation of products through the plant, including transportation decisions and dispatching to machines. Three agents were trained on 600 randomly generated job specifications and proved their ability to generalize to jobs not seen during training. Furthermore, it is demonstrated that the trained agents are able to deal with unseen job specification
formats. With this design, a self-learning reactive scheduling system is achieved that is flexible and generalized to all kinds of product variants within a defined flexible manufacturing system.