Code of WQMIX used in the paper: Learning Fair Policies in Decentralized Cooperative Multi-Agent Reinforcement Learning (https://arxiv.org/abs/2012.09421)

Please use the pymarl_wrapper for the environments used in our paper. Please see the instructions below for SMAC installation and WQMIX setup.

Weighted QMIX: Expanding Monotonic Value Function Factorisation (NeurIPS 2020)

Based on PyMARL (https://github.com/oxwhirl/pymarl/). Please refer to that repo for more documentation.

This repo contains the cleaned-up code that was used in "Weighted QMIX: Expanding Monotonic Value Function Factorisation" (https://arxiv.org/abs/2006.10800).

Included in this repo

In particular implementations for:

• OW-QMIX
• CW-QMIX
• Versions of DDPG & SAC used in the paper

We thank the authors of "QPLEX: Duplex Dueling Multi-Agent Q-Learning" (https://arxiv.org/abs/2008.01062) for their implementation of QPLEX (https://github.com/wjh720/QPLEX/), whose implementation we used. The exact implementation we used is included in this repo.

Note that in the repository the naming of certain hyper-parameters and concepts is a little different to the paper:

• α in the paper is w in the code
• Optimistic Weighting (OW) is referred to as hysteretic_qmix

For all SMAC experiments we used SC2.4.6.2.69232 (not SC2.4.10). The underlying dynamics are sufficiently different that you cannot compare runs across the 2 versions!

The install_sc2.sh script will install SC2.4.6.2.69232.

Running experiments

The config files (src/config/algs/*.yaml) contain default hyper-parameters for the respective algorithms. These were changed when running the experiments for the paper (epsilon_anneal_time = 1000000 for the robustness to exploration experiments, and w=0.1 for the predator prey punishment experiments for instance). Please see the Appendix of the paper for the exact hyper-parameters used.

Set central_mixer=atten to get the modified mixing network architecture that was used for the final experiment on corridor in the paper.

As an example, to run the OW-QMIX on 3s5z with epsilon annealed over 1mil timesteps using docker:

bash run.sh $GPU python3 src/main.py --config=ow_qmix --env-config=sc2 with env_args.map_name=3s5z w=0.5 epsilon_anneal_time=1000000

Citing

Bibtex:

@inproceedings{rashid2020weighted,
  title={Weighted QMIX: Expanding Monotonic Value Function Factorisation},
  author={Rashid, Tabish and Farquhar, Gregory and Peng, Bei and Whiteson, Shimon},
  booktitle={Advances in Neural Information Processing Systems},
  year={2020}
}

License

Code licensed under the Apache License v2.0