Reinforcement Learning is a promising tool to automate controller tuning. However significant
extensions are required for real-world applications to enable fast and robust learning. This
work proposes several additions to the state of the art and proves their capability in a series
of real world experiments.
