Autonome Optimierung des Verhaltens von Fahrzeugsteuerungen auf der Basis von Verstärkungslernen

Driver and Driver Assistance Systems move more and more into the focus of
research projects. Corresponding research areas consist of two main parts: the
understanding of the environment as well as the generation of steering
commands. Learning capabilities gain more and more importance and is the
main aspect of this research. In detail: the term “learning capability” represents
the optimization of driving behaviour, i.e. the optimized situation-specific
selection of actions.

The current research implements first-time a system based on Reinforcement
Learning (RL) – in contrast to many other research work using modeling or
neural nets. In the light of Reinforcement Learning, situations are being classified
and possible actions are being identified for each situation. Rewards following
such actions are used for cumulated ratings which in turn converge over time. In
the end, the cumulated ratings indicate as to how much an action was
appropriate for the determined situation.

After a mathematical analysis of Reinforcement Learning methods in general,
test-series are being analyzed within the framework of a real implementation and
convergence behaviour as well as driving capabilities are being achieved.