Abstract by McKell Woodland
Unsupervised Creation of Robust Environment Representations for Reinforcement Learning
Reinforcement learning (RL) methods traditionally have suffered from large sample complexities, which reduces their widespread effectiveness. This is particularly true for the beginning stages of deep RL methods, where the methods must learn what the objects in an environment are in addition to what actions to take. We aim to improve the sample complexity of deep RL methods by disentangling these two tasks and by completing the first task with with unsupervised learning methods. In particular, we propose a method that embeds learned environmental representations into a deep Q-network (DQN), so that the DQN does not have to learn the environment itself. Specifically, we provide a Rainbow DQN with robust environment representations of Atari video games created through variational ladder autoencoders.