Noisy Networks for Exploration

Mohammad Gheshlaghi Azar
Bilal Piot
Jacob Menick
Ian Osband
Alexander Graves
Vlad Mnih
Remi Munos
Demis Hassabis
Olivier Pietquin
Charles Blundell
Shane Legg
Proceedings of the International Conference on Representation Learning (ICLR 2018), Vancouver (Canada)
Google Scholar

Abstract

We introduce NoisyNet, a deep reinforcement learning agent with parametric noise added to its weights, and show that the induced stochasticity of the agent's policy can be used to aid efficient exploration.
The parameters of the noise are learned with gradient descent along with the remaining network weights. NoisyNet is straightforward to implement and adds little computational overhead.
We find that replacing the conventional exploration heuristics for A3C, DQN and dueling agents (entropy reward and epsilon-greedy respectively) with NoisyNet yields substantially higher scores for a wide range of Atari games, in some cases advancing the agent from sub to super-human performance.

Research Areas