State Influence Calculations for Deep Q-Networks

Abstract

Machine learning has become a popular tool for decision making. As these algorithms continue to make more of our decisions, the question arises of whether they are sound and just. Seeking the answer to this question is a difficult task due to the black box nature of these algorithms, and although researchers are making headway on interpreting currently employed algorithms, many algorithms in machine learning remain uninterpretable. Reinforcement learning is a branch of machine learning that has seen great improvements recently, yet there is a severe lack of interpretability methods for it. In this work we propose an interpretability approach for deep Q-networks by evaluating the influence state experiences have on the network.