Skip to content

Latest commit

 

History

History
125 lines (87 loc) · 5.77 KB

README.md

File metadata and controls

125 lines (87 loc) · 5.77 KB
Raw

MAPPO

Chao Yu*, Akash Velu*, Eugene Vinitsky, Jiaxuan Gao, Yu Wang, Alexandre Bayen, and Yi Wu.

This repository implements MAPPO, a multi-agent variant of PPO. The implementation in this repositorory is used in the paper "The Surprising Effectiveness of PPO in Cooperative Multi-Agent Games" (https://arxiv.org/abs/2103.01955). This repository is heavily based on https://github.com/ikostrikov/pytorch-a2c-ppo-acktr-gail. We also make the off-policy repo public, please feel free to try that. off-policy link

All hyperparameters and training curves are reported in appendix, we would strongly suggest to double check the important factors before runing the code, such as the rollout threads, episode length, ppo epoch, mini-batches, clip term and so on. Besides, we have updated the newest results on google football testbed and suggestions about the episode length and parameter-sharing in appendix, welcome to check that.

We have recently noticed that a lot of papers do not reproduce the mappo results correctly, probably due to the rough hyper-parameters description. We have updated training scripts for each map or scenario in /train/train_xxx_scripts/*.sh. Feel free to try that.

Environments supported:

1. Usage

WARNING: by default all experiments assume a shared policy by all agents i.e. there is one neural network shared by all agents

All core code is located within the onpolicy folder. The algorithms/ subfolder contains algorithm-specific code for MAPPO.

  • The envs/ subfolder contains environment wrapper implementations for the MPEs, SMAC, and Hanabi.

  • Code to perform training rollouts and policy updates are contained within the runner/ folder - there is a runner for each environment.

  • Executable scripts for training with default hyperparameters can be found in the scripts/ folder. The files are named in the following manner: train_algo_environment.sh. Within each file, the map name (in the case of SMAC and the MPEs) can be altered.

  • Python training scripts for each environment can be found in the scripts/train/ folder.

  • The config.py file contains relevant hyperparameter and env settings. Most hyperparameters are defaulted to the ones used in the paper; however, please refer to the appendix for a full list of hyperparameters used.

2. Installation

Here we give an example installation on CUDA == 10.1. For non-GPU & other CUDA version installation, please refer to the PyTorch website. We remark that this repo. does not depend on a specific CUDA version, feel free to use any CUDA version suitable on your own computer.

# create conda environment
conda create -n marl python==3.6.1
conda activate marl
pip install torch==1.5.1+cu101 torchvision==0.6.1+cu101 -f https://download.pytorch.org/whl/torch_stable.html
# install on-policy package
cd on-policy
pip install -e .

Even though we provide requirement.txt, it may have redundancy. We recommend that the user try to install other required packages by running the code and finding which required package hasn't installed yet.

2.1 StarCraftII 4.10

unzip SC2.4.10.zip
# password is iagreetotheeula
echo "export SC2PATH=~/StarCraftII/" > ~/.bashrc

2.2 Hanabi

Environment code for Hanabi is developed from the open-source environment code, but has been slightly modified to fit the algorithms used here.
To install, execute the following:

pip install cffi
cd envs/hanabi
mkdir build & cd build
cmake ..
make -j

Here are all hanabi models.

2.3 MPE

# install this package first
pip install seaborn

There are 3 Cooperative scenarios in MPE:

  • simple_spread
  • simple_speaker_listener, which is 'Comm' scenario in paper
  • simple_reference

2.4 GRF

Please see the football repository to install the football environment.

3.Train

Here we use train_mpe.sh as an example:

cd onpolicy/scripts
chmod +x ./train_mpe.sh
./train_mpe.sh

Local results are stored in subfold scripts/results. Note that we use Weights & Bias as the default visualization platform; to use Weights & Bias, please register and login to the platform first. More instructions for using Weights&Bias can be found in the official documentation. Adding the --use_wandb in command line or in the .sh file will use Tensorboard instead of Weights & Biases.

We additionally provide ./eval_hanabi_forward.sh for evaluating the hanabi score over 100k trials.

4. Publication

If you find this repository useful, please cite our paper:

@misc{yu2021surprising,
      title={The Surprising Effectiveness of PPO in Cooperative Multi-Agent Games}, 
      author={Chao Yu and Akash Velu and Eugene Vinitsky and Jiaxuan Gao and Yu Wang and Alexandre Bayen and Yi Wu},
      year={2021},
      eprint={2103.01955},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}