Domain adaptation for reinforcement learning on the Atari

Abstract

Deep Reinforcement learning is a powerful machine learning paradigm that has had significant success across a wide range of control problems. This success often requires long training times to achieve. Observing that many problems share similarities, it is likely that much of the training done could be redundant if knowledge could be efficiently and appropriately shared across tasks. In this paper we demonstrate a novel adversarial domain adaptation approach to transfer state knowledge between domains and tasks on the Atari game suite. We show how this approach can successfully transfer across very different visual domains of the Atari platform. We focus on semantically related games that involve returning a ball with the user controlled agent. Our experiments demonstrate that our method reduces the number of samples required to successfully train an agent to play an Atari game.