Sparse Topology-Aware Pairwise Scoring for Large-Scale Multi-Agent Reinforcement Learning

Official repo of Sparse Topology-Aware Pairwise Scoring for Large-Scale Multi-Agent Reinforcement Learning , ICML 2026.

Requirements

Create environment with conda:

conda create -n SOPS python=3.8
conda activate SOPS
pip install -r requirements.txt

MAgent environment:

pip install magent==0.1.14
pip install pettingzoo==1.12.0
cp env/battle_v3_view7.py PATH_TO_YOUR_PETTINGZOO_ENV/pettingzoo/magent/
cp env/adversarial_pursuit_view8_v3.py PATH_TO_YOUR_PETTINGZOO_ENV/pettingzoo/magent/

IMP environment:

pip install git+https://github.com/moratodpg/imp_marl.git

Usage

nohup env CUDA_VISIBLE_DEVICES=0 python src/main.py --config=[Algorithm] --env-config=[Experiment] > train_$(date +%Y%m%d_%H%M%S).log 2>&1 &

[Algorithm] and [Experiment] refer to YAML config names under src/config/algs/ and src/config/envs/, respectively.

For Example:

nohup env CUDA_VISIBLE_DEVICES=0 python src/main.py --config=SOPS --env-config=MAgent_Battle_25 > train_$(date +%Y%m%d_%H%M%S).log 2>&1 &

Citing

@inproceedings{deng2026sops,
  title={Sparse Topology-Aware Pairwise Scoring for Large-Scale Multi-Agent Reinforcement Learning},
  author={Deng, Zhibo and Liang, Feng and Zhang, Yong and Zhang, Xiaoxi and Hu, Xiping},
  booktitle={Proceedings of the 43rd International Conference on Machine Learning},
  year={2026}
}