Explaining the decision-making of a reinforcement learning agent

Résumé: The omnipresence of AI models in our everyday lives has led to the development of the domain of eXplainable AI (XAI) in order to address the need for explanations of model outputs. This need has recently been highlighted by various institutions, such as the European Commission’s AI act, which stipulates a transparency requirement based on the risk associated with the model’s field of application. This presentation focuses on a XAI sub-domain, called eXplainable Reinforcement Learning (XRL), which aims to explain the decisions of an agent that has learned by Reinforcement Learning. The first part briefly presents the different trends of XRL methods according to the explanation target and the explanation means. The second part presents my contribution to this domain.