Codebase for Order Matters: Agent-by-agent Policy Optimization. Please note that this repo is currently undergoing reconstruction and may potentially contain bugs. As a result, its performance may not be consistent with that which was reported in the paper.

CTDE MARL algorithms for StarCraft II Multi-agent Challenge (SMAC), PettingZoo, Multi-agent Particle Environment, Multi-agent MUJOCO, and Google Research Football (comming later).

This repo is heavily based on https://github.com/marlbenchmark/on-policy.

Implemented algorithms include:

• MAPPO
• CoPPO (without advantage-mix / credit assignment)
• HAPPO
• A2PO

Parameter sharing is recommended for SMAC, while separated parameters are recommended for other environments.

Monotonic Bounds

Cooperative Break Through

Serve, Pass and Shoot

# create conda environment
conda create -n co-marl python==3.8
conda activate co-marl
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch -y
pip install psycopg2-binary setproctitle absl-py pysc2 gym tensorboardX
# sudo apt install build-essential -y
pip install Cython
pip install sacred aim
# install on-policy package
pip install -e .

or

bash install.sh

bash install_sc2.sh

bash install_pettingzoo.sh

bash install_mujoco.sh

bash install_grf.sh

Here we use training MMM2 (a hard task in SMAC) as an example:

sh run_scripts/SMAC/MMM2.sh