Enhancing Explainability of Deep Reinforcement Learning Through Selective Layer-Wise Relevance Propagation

摘要

Modern deep reinforcement learning agents are capable of achieving super-human performance in tasks like playing Atari games, solely based on visual input. However, due to their use of neural networks the trained models are lacking transparency which makes their inner workings incomprehensible for humans. A promising approach to gain insights into the opaque reasoning process of neural networks is the layer-wise relevance propagation (LRP) concept. This visualization technique creates saliency maps that highlight the areas in the input which were relevant for the agents' decision-making process. Since such saliency maps cover every possible cause for a prediction, they are often accentuating very diverse parts of the input. This makes the results difficult to understand for people without a machine-learning background. In this work, we introduce an adjustment to the bRP concept that utilizes only the most relevant neurons of each convolutional layer and thus generates more selective saliency maps. We test our approach with a dueling Deep Q-Network (DQN) agent, which we trained on three different Atari games of varying complexity. Since the dueling DQN approach considerably alters the neural network architecture of the original DQN algorithm, it requires its own LRP variant which will be presented in this paper.