Exploration by Random Network Distillation

TL;DR

This paper presents Random Network Distillation (RND), a simple yet effective exploration bonus for deep reinforcement learning that significantly improves performance on challenging Atari games like Montezuma's Revenge, achieving human-level results without demonstrations.

Contribution

Introduces RND, a novel exploration bonus based on prediction error of a fixed random network, enhancing exploration in deep RL with minimal computational overhead.

Findings

Achieves state-of-the-art results on Montezuma's Revenge

First method to outperform average human performance without demonstrations

Enables occasional level completion in a notoriously difficult game

Abstract

We introduce an exploration bonus for deep reinforcement learning methods that is easy to implement and adds minimal overhead to the computation performed. The bonus is the error of a neural network predicting features of the observations given by a fixed randomly initialized neural network. We also introduce a method to flexibly combine intrinsic and extrinsic rewards. We find that the random network distillation (RND) bonus combined with this increased flexibility enables significant progress on several hard exploration Atari games. In particular we establish state of the art performance on Montezuma's Revenge, a game famously difficult for deep reinforcement learning methods. To the best of our knowledge, this is the first method that achieves better than average human performance on this game without using demonstrations or having access to the underlying state of the game, and occasionally completes the first level.