Mean-variance Based Risk-sensitive Reinforcement Learning with Interpretable Attention

Abstract

Risk-sensitive reinforcement learning (RL) has been studied to address the risk and uncertainty in autonomous systems. While a comprehensive understanding for the behaviors of RL agents plays an important role, interpretability was rarely discussed in the context of risk-sensitivity RL. In this paper, we present an interpretable visualization scheme with attention mechanism in which a saliency map represents the relative influence degree of an input state on the decision-making of mean-variance based risk-sensitive RL. Through 2D navigation experiments, we demonstrate that our scheme provides the interpretability with regard to risk-sensitive levels.