Beating Atari with Natural Language Guided Reinforcement Learning
This addresses the challenge of sample efficiency and exploration in reinforcement learning for complex environments, though it is incremental as it builds on prior multimodal and instruction-following work.
The authors tackled the problem of learning to play Atari games by using natural language instructions to guide reinforcement learning, resulting in an agent that significantly outperforms existing methods like DQN and A3C on the challenging Montezuma's Revenge environment.
We introduce the first deep reinforcement learning agent that learns to beat Atari games with the aid of natural language instructions. The agent uses a multimodal embedding between environment observations and natural language to self-monitor progress through a list of English instructions, granting itself reward for completing instructions in addition to increasing the game score. Our agent significantly outperforms Deep Q-Networks (DQNs), Asynchronous Advantage Actor-Critic (A3C) agents, and the best agents posted to OpenAI Gym on what is often considered the hardest Atari 2600 environment: Montezuma's Revenge.