Logic-based multi-agent reinforcement learning

Brian Logan

Associate Professor at Utrecht University

Many activities that are easy for humans, such as walking together with other humans, are hard to program as a set of rules for Artificial Intelligence (AI) robots. This is why machine learning is so popular in AI: correct behaviour can be learned by trial and error. However, learning can be a slow process and the results may not be completely correct or even safe. The project will investigate techniques to: (a) speed up learning, and (b) guarantee that the learnt behaviour satisfies pre-defined properties. The research will contribute to Trustworthy AI by ensuring that the policies learned as a result of multi-agent reinforcement learning conform to a declarative logical specification.
The funding is sought for a month-long visit by researchers from the TAILOR network mem- ber lab at Utrecht University (Netherlands): Natasha Alechina, Mehdi Dastani, Brian Logan and Giovanni Varricchione, to the TAILOR partner lab at La Sapienza (Italy).