Deep Coordination Graphs (Böhmer et al., 2020)

This GIT contains the implementations of the ICML 2020 paper "Deep Coordination Graphs" by Wendelin Böhmer, Vitaly Kurin and Shimon Whiteson. The code is based on the python/pytorch framework PyMARL from the Starcraft Multi-Agent Challenge. For questions and comments please contact Wendelin Böhmer.

Installation instructions

Follow the first 2 steps from PyMARL :

1. Build the Dockerfile using

cd docker 
bash build.sh 

2. Set up StarCraft II and SMAC:

bash install_sc2.sh

This will download SC2 into the 3rdparty folder and copy the maps necessary to run over.

Finally, copy the StarCraft II maps used in the DCG paper:

cp -f maps/* ./3rdparty/StarCraftII/Maps/Melee 

Using an existing PyMARL copy

Both algorithms and environments can be used with existing versions of PyMARL as well.

The Dockerfile must install the torch_scatter library, which can lead to conflicts with the installed torch version and the Ubuntu version of the base image. See docker/Dockerfile.

The environments can be imported by copying:

src/config/envs/rel_overgen.yaml
src/config/envs/ghost_hunt.yaml
src/envs/stag_hunt.py

and by appending src/envs/__init__.py with:

from .stag_hunt import StagHunt  
REGISTRY["stag_hunt"] = partial(env_fn, env=StagHunt)

The algorithms can be imported by copying:

src/config/algs/cg.yaml
src/config/algs/dcg.yaml
src/config/algs/dcg_noshare.yaml
src/config/algs/lrq.yaml
src/controllers/cg_controller.py
src/controllers/dcg_controller.py
src/controllers/dcg_noshare_controller.py
src/controllers/low_rank_controller.py
src/learners/dcg_learner.py
src/modules/agents/rnn_feature_agent.py

by appending src/controllers/__init__.py with:

from .dcg_controller import DeepCoordinationGraphMAC  
REGISTRY["dcg_mac"] = DeepCoordinationGraphMAC  
  
from .dcg_noshare_controller import DCGnoshareMAC  
REGISTRY["dcg_noshare_mac"] = DCGnoshareMAC  
  
from .cg_controller import SimpleCoordionationGraphMAC  
REGISTRY["cg_mac"] = SimpleCoordionationGraphMAC  
  
from .low_rank_controller import LowRankMAC  
REGISTRY["low_rank_q"] = LowRankMAC

by appending src/learners/__init__.py with:

from .dcg_learner import DCGLearner  
REGISTRY["dcg_learner"] = DCGLearner

and finally by appending src/modules/agents/__init__.py with:

from .rnn_feature_agent import RNNFeatureAgent  
REGISTRY["rnn_feat"] = RNNFeatureAgent

Replicate the experiments

As in the PyMARL framework, all experiments are run like this:

python3 src/main.py --config=$ALG --env-config=$ENV with $PARAMS  

The experiments should be run in a Docker container, to avoid installing complicated dependencies:

bash run.sh $GPU python3 src/main.py --config=$ALG --env-config=$ENV with $PARAMS  

The sacred logs containing the results will be stored as json files in the results folder.

Parameters for different algorithms

All DCG variants and CG can have the following topologies:

$LIST must be a list of tuples of node indices (starting with 0), for example, $LIST="[(0,1),(1,2),(2,3),(3,0)]" for a cycle of 4 agents.

Parameters for individual plots

Citing DCG

If you use DCG in your research, or any other implementation provided here that is not included in PyMARL, please cite the DCG paper:

W. Böhmer, V. Kurin and S. Whiteson. Deep Coordination Graphs, to appear at the International Conference on Machine Learning (ICML), 2020. URL: https://arxiv.org/abs/1910.00091

In BibTeX format:

@InProceedings{boehmer2020dcg,  
    title = {Deep Coordination Graphs}, 
    author = {Wendelin B\"ohmer and Vitaly Kurin and Shimon Whiteson}, 
    booktitle = {International Conference on Machine Learning}, 
    url = {https://arxiv.org/abs/1910.00091}, 
    year = {2020},
}  

As DCG uses PyMARL, you should also cite the SMAC paper:

M. Samvelyan, T. Rashid, C. Schroeder de Witt, G. Farquhar, N. Nardelli, T.G.J. Rudner, C.-M. Hung, P.H.S. Torr, J. Foerster, S. Whiteson. The StarCraft Multi-Agent Challenge, CoRR abs/1902.04043, 2019.

In BibTeX format:

@article{samvelyan19smac,
  title = {{The} {StarCraft} {Multi}-{Agent} {Challenge}},
  author = {Mikayel Samvelyan and Tabish Rashid and Christian Schroeder de Witt and Gregory Farquhar and Nantas Nardelli and Tim G. J. Rudner and Chia-Man Hung and Philiph H. S. Torr and Jakob Foerster and Shimon Whiteson},
  journal = {CoRR},
  volume = {abs/1902.04043},
  year = {2019},
}

License

Code licensed under the Apache License v2.0