Name of the participant: Sven Gronauer
Description of the IT research project: Reinforcement Learning is a powerful tool for learning complex control strategies from empirical data. Exploration and optimization processes are crucial for successful learning. However, any exploration without further restrictions in a real environment can lead to damage of the system and its immediate environment. Therefore, reinforcement learning algorithms are preferably trained in simulations. After learning an optimal behavior strategy, it can be transferred from the simulation to a physical system. However, a transfer often turns out to be a challenge because of inaccuracies in the model or tolerances in the real system. Such a discrepancy between the simulator and the physical world can lead to a failure of the transfer.
The microproject “Sirrel” (Safe and Robust Reinforcement Learning) pursues the goal of making behavioral strategies of reinforcement learning algorithms safer and more robust. The following research questions can be formulated for this purpose.
- How can safety-relevant aspects be embedded in a reinforcement learning problem?
- How can robustness in electromechanical systems be guaranteed?
- Which methods can be used to reliably transfer control strategies from a simulation to the physical world?
Software Campus partner: TU Munich, IAV GmbH
Implementation period: 01.01.2020 – 30.06.2021
