Post-modelling Explainability for Deep Reinforcement Learning

Artificial Intelligence (AI) has developed tremendously in recent years, notably thanks to the advances in neural networks. However, the "black box" character of the latter has slowed down the diffusion of Deep Learning (DL) in the industry. Indeed, despite the growing efficiency of neural networks, they still do not have the confidence of industrials. This is why explainability is a rapidly expanding research sector. Delfox is working on Deep Reinforcement Learning (DRL) for important industrial actors working in particular with critical systems. Explainability applied to Reinforcement Learning (RL) is therefore a key issue for Delfox and thus it is the focus of this internship. Explainability is still a young field of research and there is no industrial application of such a technology known to date. Hence the challenge of Delfox is to make that happen, they also need to show that their AIs are reliable. This report presents an exhaustive bibliography, a taxonomy of the eXplainable Artificial Intelligence (XAI) methods applicable to RL and the methods from XRL. From this bibliography, three methods merged, they have been studied and applied on a project. This report presents the methods called Feature Relevance (FR), Observation Clustering (OC) and Probe Sensing (PS). They allow to generate complementary explanations of the decisions and behavior of an Artificial Intelligence (AI) of RL.