Online Machine Learning for the Estimation of Process Times in Dynamic Scheduling

This paper presents a study on applying online machine-learning models to estimate the processing times of different production tasks for dynamic scheduling problems. Specifically, various machine-learning approaches and their impact on the schedule quality are evaluated. A discrete event simulation of a production process was created, with different functions determining processing time. This production simulation was optimized with the OERAPGA optimization algorithm. Based on the experiments, the performance of different machine learning models was assessed. Moreover, the speed and prediction quality of these models and their resulting optimization quality were evaluated. Results showed that the speed of evaluation plays a more significant role in optimization quality than prediction accuracy, as the optimizer seems to be focused on optimizing macroscopic aspects of the production schedule.