Publication
A Dynamic Multi-objective Scheduling Approach for Gradient-Based Reinforcement Learning
Katharina Hengel; Achim Wagner; Martin Ruskowski
In: 18th IFAC Symposium on Information Control Problems in Manufacturing INCOM 2024. IFAC Symposium on Information Control Problems in Manufacturing (INCOM-2024), 18th IFAC Symposium on Information Control Problems in Manufacturing, August 28-30, Vienna, Austria, Elsevier, 2024.
Abstract
In manufacturing dynamic scheduling is a complex task which is influenced by various factors, including several possibly contrary optimization objectives. While multi-objective optimization approaches are not novel anymore, prevailing deep learning solutions lack in the ability to dynamically adjust the preferences of the different objectives. For this reason, we developed a collaborate human-AI scheduling algorithm where Reinforcement Learning (RL) is utilized for the optimization of single objectives, while the human is left in control to flexibly mix and weight the influence of each objective into the final schedule. The approach performs on par with prevailing RL multi-objective approaches. However, it surpasses the later in its ability to dynamically adjust the preferences of the objectives without retraining.