Robust and General Reinforcement Learning via AutoML

Reinforcement Learning (RL) enables learning through interaction with the environment. Therefore, it is a component of AI systems for sequential decision-making, e.g., in robotics, in natural sciences such as physics or medicine, or in large language models.

However, RL is not only powerful but also difficult to apply: current methods tend to be unstable, have limited generalization capabilities, and their success depends heavily on design choices. In recent years, automated RL (AutoRL) methods have gained traction to achieve better performance, robustness, and training efficiency through systematic and data-driven approaches. Our goal is to improve both the efficiency and robustness of AutoRL to enable widespread use of these tools for RL practitioners and researchers alike. AutoRL will help democratize RL training, opening up new opportunities across all areas of RL research and applications.