Distributed Deep Reinforcement Learning: Learn how to play Atari games in 21 minutes (2018-01-09T00:00:00.000000Z)

TL;DR

This study shows that using the Adam optimization algorithm with a batch size of up to 2048 is a viable choice for carrying out large scale machine learning computations.

Abstract

We present a study in Distributed Deep Reinforcement Learning (DDRL) focused on scalability of a state-of-the-art Deep Reinforcement Learning algorithm known as Batch Asynchronous Advantage ActorCritic (BA3C). We show that using the Adam optimization algorithm with a batch size of up to 2048 is a viable choice for carrying out large scale machine learning computations. This, combined with careful reexamination of the optimizer's hyperparameters, using synchronous training on the node level (while keeping the local, single node part of the algorithm asynchronous) and minimizing the memory footprint of the model, allowed us to achieve linear scaling for up to 64 CPU nodes. This corresponds to a training time of 21 minutes on 768 CPU cores, as opposed to 10 hours when using a single node with 24 cores achieved by a baseline single-node implementation.

Authors

H. Michalewski

4 papers

Igor Adamski

1 papers

R. Adamski

1 papers

TL;DR

Abstract

Authors

References40 items

Accelerated Methods for Deep Reinforcement Learning

Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm

Mastering the game of Go without human knowledge

Emergent Complexity via Multi-Agent Competition

ImageNet Training in Minutes

100-epoch ImageNet Training with AlexNet in 24 Minutes

Scaling SGD Batch Size to 32K for ImageNet Training

Proximal Policy Optimization Algorithms

Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour

TernGrad: Ternary Gradients to Reduce Communication in Distributed Deep Learning

Atari Games and Intel Processors

Sparse Communication for Distributed Gradient Descent

Evolution Strategies as a Scalable Alternative to Reinforcement Learning

GA3C: GPU-based A3C for Deep Reinforcement Learning

Practical Efficiency of Asynchronous Stochastic Gradient Descent

Reinforcement Learning through Asynchronous Advantage Actor-Critic on a GPU

Sample Efficient Actor-Critic with Experience Replay

QSGD: Randomized Quantization for Communication-Optimal Stochastic Gradient Descent

Distributed Training of Deep Neural Networks: Theoretical and Practical Limits of Parallel Scalability

OpenAI Gym

Revisiting Distributed Synchronous SGD

Asynchronous Methods for Deep Reinforcement Learning

Mastering the game of Go with deep neural networks and tree search

Massively Parallel Methods for Deep Reinforcement Learning

Asynchronous parallel stochastic gradient descent: a numeric core for scalable distributed machine learning algorithms

Human-level control through deep reinforcement learning

Trust Region Policy Optimization

Adam: A Method for Stochastic Optimization

1-bit stochastic gradient descent and its application to data-parallel distributed training of speech DNNs

Playing Atari with Deep Reinforcement Learning

On the importance of initialization and momentum in deep learning

Large Scale Distributed Deep Networks

The Arcade Learning Environment: An Evaluation Platform for General Agents

On optimization methods for deep learning

Adaptive Subgradient Methods for Online Learning and Stochastic Optimization

Deep Learning

Tensorpack

Scalable distributed DNN training using commodity GPU cloud computing

TensorFlow: Large-scale machine learning on heterogeneous systems

Lecture 6.5—RmsProp: Divide the gradient by a running average of its recent magnitude

Field of Study

Venue Information

Name

Type

URL

Alternate Names