We propose a conceptually simple and lightweight framework for deep reinforcement learning that uses asynchronous gradient descent for optimization of deep neural network controllers. We present asynchronous variants of four standard reinforcement learning algorithms and show that parallel actor-learners have a stabilizing effect on training allowing all four methods to successfully train neural network controllers. The best performing method, an asynchronous variant of actor-critic, surpasses the current state-of-the-art on the Atari domain while training for half the time on a single multi-core CPU instead of a GPU. Furthermore, we show that asynchronous actor-critic succeeds on a wide variety of continuous motor control problems as well as on a new task of navigating random 3D mazes using a visual input.
%0 Journal Article
%1 mnih2016a3c
%A Mnih, Volodymyr
%A Badia, Adrià Puigdomènech
%A Mirza, Mehdi
%A Graves, Alex
%A Lillicrap, Timothy P.
%A Harley, Tim
%A Silver, David
%A Kavukcuoglu, Koray
%D 2016
%J CoRR
%K DRLAlgoComparison a3c actor_critic reinforcement_learning
%T Asynchronous Methods for Deep Reinforcement Learning.
%U http://dblp.uni-trier.de/db/journals/corr/corr1602.html#MnihBMGLHSK16
%V abs/1602.01783
%X We propose a conceptually simple and lightweight framework for deep reinforcement learning that uses asynchronous gradient descent for optimization of deep neural network controllers. We present asynchronous variants of four standard reinforcement learning algorithms and show that parallel actor-learners have a stabilizing effect on training allowing all four methods to successfully train neural network controllers. The best performing method, an asynchronous variant of actor-critic, surpasses the current state-of-the-art on the Atari domain while training for half the time on a single multi-core CPU instead of a GPU. Furthermore, we show that asynchronous actor-critic succeeds on a wide variety of continuous motor control problems as well as on a new task of navigating random 3D mazes using a visual input.
@article{mnih2016a3c,
abstract = {We propose a conceptually simple and lightweight framework for deep reinforcement learning that uses asynchronous gradient descent for optimization of deep neural network controllers. We present asynchronous variants of four standard reinforcement learning algorithms and show that parallel actor-learners have a stabilizing effect on training allowing all four methods to successfully train neural network controllers. The best performing method, an asynchronous variant of actor-critic, surpasses the current state-of-the-art on the Atari domain while training for half the time on a single multi-core CPU instead of a GPU. Furthermore, we show that asynchronous actor-critic succeeds on a wide variety of continuous motor control problems as well as on a new task of navigating random 3D mazes using a visual input. },
added-at = {2019-12-16T18:24:08.000+0100},
author = {Mnih, Volodymyr and Badia, Adrià Puigdomènech and Mirza, Mehdi and Graves, Alex and Lillicrap, Timothy P. and Harley, Tim and Silver, David and Kavukcuoglu, Koray},
biburl = {https://www.bibsonomy.org/bibtex/230e02198c1496f07bc211290fe11aa0b/lanteunis},
ee = {http://arxiv.org/abs/1602.01783},
interhash = {02e623113f85237b4ec7daf03736c6cc},
intrahash = {30e02198c1496f07bc211290fe11aa0b},
journal = {CoRR},
keywords = {DRLAlgoComparison a3c actor_critic reinforcement_learning},
timestamp = {2019-12-16T21:10:33.000+0100},
title = {Asynchronous Methods for Deep Reinforcement Learning.},
url = {http://dblp.uni-trier.de/db/journals/corr/corr1602.html#MnihBMGLHSK16},
volume = {abs/1602.01783},
year = 2016
}