Skip to content

Code accompanying the paper "TiZero: Mastering Multi-Agent Football with Curriculum Learning and Self-Play" (AAMAS 2023) 足球游戏智能体

License

Notifications You must be signed in to change notification settings

OpenRL-Lab/TiZero

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

41 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

License PyPI PyPI - Python Version Documentation Status

Introduction

Reinforcement learning agent for Google Research Football.

Code accompanying the paper "TiZero: Mastering Multi-Agent Football with Curriculum Learning and Self-Play" (AAMAS 2023). [paper] [videos].

Installation

  • Follow the instructions in gfootball to set up the environment.
  • pip install gfootball openrl "openrl[selfplay]"
  • pip install tizero (or clone this repo and pip install -e .).
  • test the installation by python3 -m gfootball.play_game --action_set=full.

Evaluate JiDi submissions locally

You can evaluate your agent locally using tizero:

tizero eval --left_agent submission_dir1 --right_agent submission_dir2 --total_game 10

For example, you can evaluate tizero with random agent as below:

tizero eval --left_agent submission/tizero --right_agent submission/random_agent --total_game 10

For evaluations for JiDi submissions on other games, please refer to the Arena of OpenRL and this example for the snake game.

Show a saved dump file

  • show detailed infomation of a match via: tizero show dump_file
  • show keypoints of a mactch via: tizero keypoint dump_file

You can download an example dump file from here.

Then execute: tizero show daily_6484285.dump or tizero keypoint daily_6484285.dump. Then you will see a GUI as below:

Convert dump file to video

After the installation, you can use tizero to convert a dump file to a video file. The usage is tizero dump2video <dump_file> <output_dir> --episode_length <the length> --render_type <2d/3d>.

You can download an example dump file from here. And then execute tizero dump2video daily_6484285.dump ./ in your terminal. By default, the episode length is 3000 and the render type is 2d. Wait a minute, you will get a video file named daily_6484285.avi in your current directory.

Submit TiZero to JIDI(及第评测平台)

JIDI is a public evaluation platform for RL agents. You can submit your agent of GRF at: http://www.jidiai.cn/env_detail?envid=34.

We provide several agents under ./submission/ directory, which can be submitted to JIDI directly:

  • ./submission/tizero: the final model of TiZero for JIDI submission, which ranked 1st on October 28th, 2022.
  • ./submission/random_agent: the random agent for JIDI submission.

Cite

Please cite our paper if you use our codes or our weights in your own work:

@inproceedings{lin2023tizero,
  title={TiZero: Mastering Multi-Agent Football with Curriculum Learning and Self-Play},
  author={Lin, Fanqi and Huang, Shiyu and Pearce, Tim and Chen, Wenze and Tu, Wei-Wei},
  booktitle={Proceedings of the 2023 International Conference on Autonomous Agents and Multiagent Systems},
  pages={67--76},
  year={2023}
}