Agent57: Outperforming the Atari Human Benchmark

Adrià Puigdomènech Badia, Bilal Piot, Steven Kapturowski, Pablo Sprechmann, Alex Vitvitskyi, Zhaohan Daniel Guo, Charles Blundell
Proceedings of the 37th International Conference on Machine Learning, PMLR 119:507-517, 2020.

Abstract

Atari games have been a long-standing benchmark in the reinforcement learning (RL) community for the past decade. This benchmark was proposed to test general competency of RL algorithms. Previous work has achieved good average performance by doing outstandingly well on many games of the set, but very poorly in several of the most challenging games. We propose Agent57, the first deep RL agent that outperforms the standard human benchmark on all 57 Atari games. To achieve this result, we train a neural network which parameterizes a family of policies ranging from very exploratory to purely exploitative. We propose an adaptive mechanism to choose which policy to prioritize throughout the training process. Additionally, we utilize a novel parameterization of the architecture that allows for more consistent and stable learning.

Cite this Paper


BibTeX
@InProceedings{pmlr-v119-badia20a, title = {Agent57: Outperforming the {A}tari Human Benchmark}, author = {Badia, Adri{\`a} Puigdom{\`e}nech and Piot, Bilal and Kapturowski, Steven and Sprechmann, Pablo and Vitvitskyi, Alex and Guo, Zhaohan Daniel and Blundell, Charles}, booktitle = {Proceedings of the 37th International Conference on Machine Learning}, pages = {507--517}, year = {2020}, editor = {Hal Daumé III and Aarti Singh}, volume = {119}, series = {Proceedings of Machine Learning Research}, month = {13--18 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v119/badia20a/badia20a.pdf}, url = { http://proceedings.mlr.press/v119/badia20a.html }, abstract = {Atari games have been a long-standing benchmark in the reinforcement learning (RL) community for the past decade. This benchmark was proposed to test general competency of RL algorithms. Previous work has achieved good average performance by doing outstandingly well on many games of the set, but very poorly in several of the most challenging games. We propose Agent57, the first deep RL agent that outperforms the standard human benchmark on all 57 Atari games. To achieve this result, we train a neural network which parameterizes a family of policies ranging from very exploratory to purely exploitative. We propose an adaptive mechanism to choose which policy to prioritize throughout the training process. Additionally, we utilize a novel parameterization of the architecture that allows for more consistent and stable learning.} }
Endnote
%0 Conference Paper %T Agent57: Outperforming the Atari Human Benchmark %A Adrià Puigdomènech Badia %A Bilal Piot %A Steven Kapturowski %A Pablo Sprechmann %A Alex Vitvitskyi %A Zhaohan Daniel Guo %A Charles Blundell %B Proceedings of the 37th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2020 %E Hal Daumé III %E Aarti Singh %F pmlr-v119-badia20a %I PMLR %P 507--517 %U http://proceedings.mlr.press/v119/badia20a.html %V 119 %X Atari games have been a long-standing benchmark in the reinforcement learning (RL) community for the past decade. This benchmark was proposed to test general competency of RL algorithms. Previous work has achieved good average performance by doing outstandingly well on many games of the set, but very poorly in several of the most challenging games. We propose Agent57, the first deep RL agent that outperforms the standard human benchmark on all 57 Atari games. To achieve this result, we train a neural network which parameterizes a family of policies ranging from very exploratory to purely exploitative. We propose an adaptive mechanism to choose which policy to prioritize throughout the training process. Additionally, we utilize a novel parameterization of the architecture that allows for more consistent and stable learning.
APA
Badia, A.P., Piot, B., Kapturowski, S., Sprechmann, P., Vitvitskyi, A., Guo, Z.D. & Blundell, C.. (2020). Agent57: Outperforming the Atari Human Benchmark. Proceedings of the 37th International Conference on Machine Learning, in Proceedings of Machine Learning Research 119:507-517 Available from http://proceedings.mlr.press/v119/badia20a.html .

Related Material