How can AI systems better transfer their learned behavior to new, previously unknown situations? Dr. André Biedenkapp will be working on this question at the Karlsruhe Institute of Technology (KIT). For his research on improving the generalizability of reinforcement learning methods, he has successfully acquired an Emmy Noether Group grant from the German Research Foundation (DFG) and will receive around 1.2 million euros over the next three years. Following a successful interim evaluation after the first three years, the DFG has promised around 920,000 euros for a further three years.
"Emmy Noether funding enables outstanding young researchers to achieve scientific independence at an early stage," says Professor Stefan Hinz, Vice Provost for Young Researchers at KIT. "With his work on the further development of adaptive AI systems, Dr. André Biedenkapp is an example of excellent young research at KIT."
Reinforcement learning (RL) is an artificial intelligence (AI) learning paradigm in which an AI agent learns by trial and error how it should behave in a given environment. Feedback in the form of rewards helps the system to repeat desirable behavior and avoid unfavorable behavior. This method is particularly effective for problems where decisions have to be made sequentially - for example in robotics, logistics or resource management.
However, a central problem with classic RL approaches is that the strategies learned are often strongly tailored to the respective training environment. Even minor changes can result in the AI agent no longer knowing how to behave appropriately. "Today, RL agents work excellently under the conditions for which they have been trained - but quickly reach their limits when these change," says Dr. André Biedenkapp. He is working at the University of Freiburg until August 2026. From September 2026, he will start at the Institute of Anthropomatics and Robotics at KIT with the newly acquired DFG Emmy Noether Group "From Mediocre to Masterful Generalists: The Power of Context in RL".
More context for more robust learning methods
This is where the Emmy Noether Group comes in: The aim is to expand training procedures for RL in such a way that artificial intelligences become more robust and adaptable. To this end, Biedenkapp and his team take into account additional information about the respective environment or world in which the agent operates. In this way, the AI should learn which behavior is particularly suitable in which situation - and also be able to transfer this knowledge to similar, unknown situations.
In the long term, this approach could be a decisive step towards using RL more in real-life applications. Until now, many RL-based AI systems have relied on high-precision simulators that reproduce real environments as accurately as possible. However, such simulators are complex, expensive and hardly feasible for complex scenarios. "If RL-based systems generalize better, it becomes less crucial to perfectly simulate every possible situation. This would significantly expand the possible applications of this technology," says Biedenkapp.
About the Emmy Noether Program
The Emmy Noether Program gives outstandingly qualified scientists in the early stages of their careers the opportunity to qualify for a university professorship by independently leading a group over a period of six years.
