This is the implementation of the paper "Multi-Agent Incentive Communication via Decentralized Teammate Modeling". This repo is currently maintained by the LAMDA-RL group.
Note: the experiments of MAIC is conducted in SC2.4.6.2.69232, which is same as the SMAC run data release (https://github.com/oxwhirl/smac/releases/tag/v1). The results are not always comparable with results run in SC2.4.10.
Set up StarCraft II and SMAC:
bash install_sc2.shThis will download SC2.4.6.2.69232 into the 3rdparty folder and copy the maps necessary to run over. You may also need to set the environment variable for SC2:
export SC2PATH=[Your SC2 folder like /abc/xyz/3rdparty/StarCraftII]Install Python environment with conda:
conda create -n pymarl python=3.7 -y
conda activate pymarl
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch -y
pip install sacred numpy scipy matplotlib seaborn pyyaml pygame pytest probscale imageio snakeviz tensorboard-logger tensorboard tensorboardx
# install qplex smac
pip install qplex_smac/or install with requirements.txt of pip:
pip install -r requirements.txt
# install qplex smac
pip install qplex_smac/As MAIC contains implementation with QPLEX mixing network (https://github.com/wjh720/QPLEX), please install the QPLEX version of SMAC in qplex_smac folder. You can also run algorithms without QPLEX mixing network integration with default SMAC installed (https://github.com/oxwhirl/smac).
python3 src/main.py --config=[Algorithm name] --env-config=[Env name] with env_args.map_name=[Map name if choosing SC2 env]The config files act as defaults for an algorithm or environment.
They are all located in src/config.
--config refers to the config files in src/config/algs includeing MAIC and other baselines.
--env-config refers to the config files in src/config/envs including sc2, foraging as the LB-Foraging environment (https://github.com/semitable/lb-foraging), join1 as the hallway environment (https://github.com/TonghanWang/NDQ).
All results will be stored in the results folder.
For example, run MAIC with QMIX mixing network (default SC2 evaluation in the paper) :
python src/main.py --config=maic --env-config=sc2 with env_args.map_name=MMM2 seed=42
Run MAIC with QPLEX mixing network:
python src/main.py --config=maic_qplex --env-config=sc2 with env_args.map_name=MMM2 seed=42