IPAB Workshop - 30/06/2022

Speaker: Michael Herrmann

Talk title: Reflexive Reinforcement Learning

Abstract: Inverse reinforcement learning (RL) aims at inferring the motivation for observed behaviour and enables behavioural learning that is based on an understanding of the target behaviour in terms of the RL paradigm, instead of a mere copying of the behaviour. Reflexive RL arises from an integration of forward RL and inverse RL. It not only emphasises their duality, but intends also to introduce a simple form of self-reflection into RL. The idea of reflexive RL will be illustrated by several simulation examples related to tasks such as learning to learn, learning to explore, and learning to teach. In addition, we will discuss an interpretable approach to inverse RL and benefits of reflexive RL for explainable and intrinsically motivated RL.