Title: Explainability in Reinforcement Learning
Abstract: Behaviour acquired through reinforcement learning (RL) is trivially explainable by reward maximisation. This talk aims at a more comprehensive picture of RL that also includes aspects such as optimality, interpretability, expedience and safety. These considerations will be based on recent projects, where we have studied theoretically and in simulations, (i) the efficiency of RL based on relevant information, (ii) the structure of biasses that can facilitate RL as a form of intrinsic motivation or consolidate acquired behaviour, (iii) the interpretability of representations formed in RL, and (iv) the benefits of cooperation in RL. Generally, we have adopted a learning-to-learn approach that modulates RL in order to enhance versatility and acceptability of this learning paradigm.