Umsetzung einer querführenden Fahrassistenzfunktion mithilfe von Reinforcement Learning

Autoren

L. Witt, D. Münning, H. Oschelies - Volkswagen AG, S. Schmidt - Otto-von-Guericke-Universität Magdeburg

Zusammenfassung

The demand and performance of current driver assistance systems is constantly in-creasing, whereby their application is an extensive task. In order to enable an optimal customer experience, a vehicle-specific application is necessary, which takes into account both differences due to component tolerances and changes over the lifetime of a vehicle. Adaptive control methods represent an option that can be implemented with the help of machine learning due to the increasing availability of data.

In this paper, an approach for an adaptive control for a lateral driver assistance system using machine learning is presented. A setup is being developed that enables online training in the vehicle using reinforcement learning. A comparison of different exploration metrics shows that adaptation to a defined desired behavior is possible.