Multi-agent Coordination via Flow Matching




Paper

Overview

This codebase provides an official implementation of MAC-Flow, which is an algorithm for offline MARL.

Installation

# Environment (OG-MARL: https://github.com/instadeepai/og-marl)
conda create -n recipe python=3.8
conda activate recipe
python -m pip install --upgrade pip
pip install -r requirements/datasets.txt

Install the SMAC, MPE, MA-MuJoCo environment you plan to use:

# SMAC v1
bash install_environments/smacv1.sh
pip install -r install_environments/requirements/smacv1.txt

# SMAC v2
bash install_environments/smacv2.sh
pip install -r install_environments/requirements/smacv2.txt

# MPE
pip install -r install_environments/requirements/pettingzoo.txt

# MA-MuJoCo
bash install_environments/mujoco200.sh
pip install -r install_environments/requirements/mamujoco200.txt

If you need baseline TensorFlow dependencies:

pip install -r requirements/baselines.txt

Quick start

# TD3BC
python smac_main.py --env smac_v1 --source og_marl --scenario 2s3z --dataset Good --seed 0 --agent.alpha 3.0