Using Decision Trees and Reinforcement Learning for the Dynamic Adjustment of Composite Sequencing Rules in a Flexible Manufacturing System

Integrating machine learning methods into the scheduling process to adjust priority rules dynamically can improve the performance of manufacturing systems. In this paper, three methods for adjusting the k-values of the ATCS sequencing rule are analyzed: neural networks, decision trees and reinforcement learning. They are evaluated in a static and a dynamic scenario. The required dataset was synthetically generated using a discrete event simulation of a flow shop environment, where product mix and system utilization were varied systematically. Across all scenarios, it is shown that all three methods can improve the performance. On par, RL and NN can reduce the mean tardiness by up to 15% and compensate for unplanned product mix changes