Robust Multi-agent Coordination via Evolutionary Generation of Auxiliary Adversarial Attackers

This work is accepted as Oral at the Association for the Advancement of Artificial Intelligence (AAAI) 2023.

Environment Installation

Build the environment by running:

pip install -r requirements.txt

conda env create -f environment.yaml

Running

Some example scripts have been provided. The code allows various parameters in the model names to enable different settings.

By running:

python3 src/main.py --config=qmix_attack --env-config=sc2 with env_args.map_name=2s3z

diverse adversarial attackers against trained ego-system under SMAC 2s3z map can be generated.

By running:

python3 src/main.py --config=qmix_robust --env-config=sc2 with env_args.map_name=2s3z attack_num=8
python3 src/main.py --config=vdn_robust --env-config=sc2 with env_args.map_name=2s3z attack_num=8 ego_agent_path="./ego_models/vdn_2s3z"

robust ego-system on map 2s3z will be generated.

We provide checkpoints of pre-trained agents (vanilla QMIX), advesarial attackers and robust trained agents (ROMANCE) on map 2s3z, 3s_vs_3z for quick evaluation by running:

bash run_evaluate.sh 2s3z
bash run_evaluate.sh 3s_vs_3z

Visualization of Attack and Defense

Comparison with Baselines

License

All the source code that has been taken from the EPyMARL repository was licensed (and remains so) under the Apache License v2.0 (included in LICENSE file). Any new code is also licensed under the Apache License v2.0.

Citation

@inproceedings{romance,
  title={Robust Multi-agent Coordination via Evolutionary Generation of Auxiliary Adversarial Attackers},
  author={Yuan, Lei and Zhang, Ziqian and Xue, Ke and  Yin, Hao and   Chen,  Feng and  Guan, Cong and  Li, Lihe and  Qian, Chao and  Yu, Yang},
  booktitle={AAAI},
  year={2023}
}