Skip to content
This repository has been archived by the owner on Apr 24, 2024. It is now read-only.
/ srl Public archive

SRL: Scaling Distributed Reinforcement Learning to Over Ten Thousand Cores

Notifications You must be signed in to change notification settings

openpsi-projects/srl

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Note

This repository has been moved to a new address https://github.com/openpsi-project/srl !!! This repository is no longer maintained. Please check our new updates in the new repository!

SRL (ReaLly Scalable RL): Scaling Distributed Reinforcement Learning to Over Ten Thousand Cores

SRL is an efficient, scalable and extensible distributed Reinforcement Learning system. SRL supports running several state-of-the-art RL algorithms on some common environments with one simple configuration file, and also exposes general APIs for users to develop their self-defined environments, policies and algorithms. SRL even allows users to implement new system components to support their algorithm designs, if current system architecture is not sufficient.

Currently, our scheduler with slurm is not released. We are planning to implement a ray version launcher for users to easily deploy SRL on a large scale!

Algorithms and Environments

In this repository, one algorithm (Proximal Policy Optimization) and five environments (Gym Atari, Google football, Gym MuJoCo, Hide and Seek, SMAC) are implemented as examples. In the future, more environment and algorithm supports will be added to build an RL library with SRL.

Installation

Before installation, make sure you have python>=3.8 and torch>=1.10.0, gym installed. Wandb is also supported, please install wandb package if you intend to use it for logging. You should also install environments you intend to run. For more information, check links about supported envrionment in previous section. (Note that Google football environment requires a older version of gym==0.21.0)

Contents in this repository could be installed as a python package. To install, you should clone this repository and install the package by:

git clone https://github.com/openpsi-projects/srl.git

cd srl && pip install -e .

Running an Experiment

After installing SRL and atari environment, to run a simple experiment we provide as an example:

srl-local run -e atari-mini -f test

This command line will start a run of simple PPO training on environment atari, defined by:

Documentation

For more user guides:

For more information about SRL:

Full paper

Full paper: SRL: Scaling Distributed Reinforcement Learning to Over Ten thousand cores available in arxiv! Link: https://arxiv.org/abs/2306.16688

About

SRL: Scaling Distributed Reinforcement Learning to Over Ten Thousand Cores

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages