Explain Your Move: Understanding Agent Actions Using Specific and\n Relevant Feature Attribution

Abstract

As deep reinforcement learning (RL) is applied to more tasks, there is a need\nto visualize and understand the behavior of learned agents. Saliency maps\nexplain agent behavior by highlighting the features of the input state that are\nmost relevant for the agent in taking an action. Existing perturbation-based\napproaches to compute saliency often highlight regions of the input that are\nnot relevant to the action taken by the agent. Our proposed approach, SARFA\n(Specific and Relevant Feature Attribution), generates more focused saliency\nmaps by balancing two aspects (specificity and relevance) that capture\ndifferent desiderata of saliency. The first captures the impact of perturbation\non the relative expected reward of the action to be explained. The second\ndownweighs irrelevant features that alter the relative expected rewards of\nactions other than the action to be explained. We compare SARFA with existing\napproaches on agents trained to play board games (Chess and Go) and Atari games\n(Breakout, Pong and Space Invaders). We show through illustrative examples\n(Chess, Atari, Go), human studies (Chess), and automated evaluation methods\n(Chess) that SARFA generates saliency maps that are more interpretable for\nhumans than existing approaches. For the code release and demo videos, see\nhttps://nikaashpuri.github.io/sarfa-saliency/.\n