Gaussian Processes for Dispatching Rule Selection in Production Scheduling: Comparison of Learning Techniques

Abstract

Decentralized scheduling with dispatching rules is applied in many fields of logistics and production, especially in semiconductor manufacturing, which is characterized by high complexity and dynamics. Many dispatching rules have been found, which perform well on different scenarios, however no rule has been found, which outperforms other rules across various objectives. To tackle this drawback, approaches, which select dispatching rules depending on the current system conditions, have been proposed. Most of these use learning techniques to switch between rules regarding the current system status. Since the study of Rasmussen has shown that Gaussian processes as a machine learning technique have outperformed other techniques like neural networks under certain conditions, we propose to use them for the selection of dispatching rules in dynamic scenarios. Our analysis has shown that Gaussian processes perform very well in this field of application. Additionally, we showed that the prediction quality Gaussian processes provide could be used successfully.