Technology Readiness Levels of Reinforcement Learning methods for simulation-based production scheduling

Seipolt, ArneBuschermöhle, RalfHöfinghoff, MaximilianKorn, Goy-HinrichSchumacher, MarcelKlein, MaikeKrupka, DanielWinter, CorneliaWohlgemuth, Volker2023-11-292023-11-292023978-3-88579-731-9https://dl.gi.de/handle/20.500.12116/43068Digital Twins (DT) are nowadays widely used and provide a benefit for the companies using it. One service of the DT is the simulation of a production process. This enables an optimization of the production process by simulation optimization, for example with Reinforcement Learning (RL). To support researchers and practitioners in deciding which algorithm is suitable for an implementation under real-life conditions, a literature research is performed, and a Machine Learning Technology Readiness Level is assigned to the different RL-Algorithms. It can be shown that recent research focuses mainly on model free value based and evolutionary algorithms, and both are suitable for an implementation in a real-world scenario. Both algorithms can outperform widely applied dispatching rules. Nevertheless, it should be evaluated why other algorithms are not in the focus of recent research and how the algorithms perform in comparison to each other.enTechnology ReadinessReinforcement LearningSimulationProduction SchedulingTechnology Readiness Levels of Reinforcement Learning methods for simulation-based production schedulingText/Conference Paper10.18420/inf2023_1441617-5468