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
• pip install tizero (or clone this repo and pip install -e .).
• test the installation by python3 -m gfootball.play_game --action_set=full.

Convert dump file to video

After the installation, you can use tizero to convert a dump file to a video file. You can download an example dump file from here. And then execute tizero --tool dump2video --input daily_6484285.dump --output ./ in your terminal. 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 two agents under ./submission/ directory:

• ./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:

@article{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},
  journal={arXiv preprint arXiv:2302.07515},
  year={2023}
}