Noisy Networks for Exploration
About
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.
Meire Fortunato, Mohammad Gheshlaghi Azar, Bilal Piot, Jacob Menick, Ian Osband, Alex Graves, Vlad Mnih, Remi Munos, Demis Hassabis, Olivier Pietquin, Charles Blundell, Shane Legg• 2017
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Reinforcement Learning | Atari 2600 Montezuma's Revenge ALE (test) | Score3.70e+3 | 24 | |
| Reinforcement Learning | Atari 57 | Atlantis9.24e+5 | 21 | |
| Reinforcement Learning | Atari 2600 57 games | Median Human-Normalized Score118 | 20 | |
| Reinforcement Learning | Atari 2600 Gravitar ALE (test) | Score2.21e+3 | 19 | |
| Reinforcement Learning | Atari-57 (test) | Median Human Norm Return172 | 15 | |
| Reinforcement Learning | Solaris | Final Mean Performance1.24e+4 | 5 | |
| Reinforcement Learning | PrivateEye | Final Mean Performance1.58e+4 | 5 | |
| Reinforcement Learning | Venture | Final Mean Performance1.81e+3 | 5 | |
| Reinforcement Learning | Pitfall! | Final Mean Performance0.00e+0 | 5 |
Showing 9 of 9 rows