A robustness analysis of Deep Q Networks

摘要

Deep Q Networks are a type of deep reinforcement learning algorithm that have been shown to be particularly adept at learning a variety of tasks with minimal priors. Specifically, DQN agents have been shown to learn a variety of Atari 2600 video games using only raw images of the game screen and the game score. To leverage DQNs in real world robotics applications, we must first understand how robust these networks are to the perceptual noise common to all robotics domains. In this pa- per, we present an analysis of the robustness of Deep Q Networks to various types of perceptual noise (changing brightness, Gaussian blur, salt and pepper, distractors). We present a benchmark example that involves playing the game Breakout though a webcam and screen environment, like humans do. We present a simple training approach to improve the performance maintained when transferring a DQN agent trained in simulation to the real world (36% vs. 1% maintained performance - see Table 1). We also evaluate DQN agents trained under a variety of simulation environments to report for the first time how DQNs cope with perceptual noise, common to real world robotic applications.