Noisy Zero-Shot Coordination (NZSC)

This repository implements the SelfPlay, Noisy Zero-Shot Coordination (NZSC), and meta-learning methods for training agents to effectively coordinate in such settings.

Available Environments

• OS-NLG (One Shot Noisy Lever Game)
• I-NLG (Iterated Lever Game)
• CEE (Coordinated Exploration Environment)
• SSE (SyncSight Environment)

Environment Setup

conda create -n NZSC python=3.10
conda activate NZSC
pip install -r requirements.txt

Training Procedures

Training via SelfPlay

python SelfPlay_{env_name}.py

Make sure to adjust the environment parameters listed in the main() function as needed. Additional environment parameters are available in the config files along with training hyperparameters.

Training via NZSC

python NZSC_{env_name}.py

Note: Make sure to utilize the population of seeds that were trained in SelfPlay.

Training via MetaNZSC

python Meta_NZSC_{env_name}.py

Acknowledgements

We would like to thank the authors of JaxMARL for their MultiAgent JAX implementations that inspired the creation of our environments.