Towards an Automatic Ensemble Methodology for Explainable Reinforcement Learning

Nowadays, the performances of Deep Reinforcement Learning algorithms have surpassed human capabilities at the cost of losing transparency. For this reason, the attention has been moved to methodologies related to Interpretable and Explainable Reinforcement Learning. A fundamental aspect in generating reliable explanations for Reinforcement Learning agents consists of the identification of important states. Therefore, a multitude of metrics have been developed to characterize specific situations and exigencies. In this paper, we propose a methodology for the recognition of the states helpful for getting insights into the choices of the agent through the analysis of the importance measures of stored trajectories. Moreover, we couple this mechanism with a pipeline for the creation of explanations through Bayesian Networks and Recurrent Neural Networks. In this way, we are able to extend the predicted information provided to human users by including a risk-awareness report. Finally, this advanced approach is evaluated in the Taxi grid world environment to underline the efficiency in identifying fundamental states and the generation of extensive explanations.