E-MAPP

This is the official repository for the paper: E-MAPP: Efficient Multi-Agent Reinforcement Learning with Parallel Program Guidance.[Project page]

Setup the env

pip install -r requirements.txt

Dataset

Examplary programs are provided in the program directory. You can also add your own programs following the grammar. You can generate different overcooked scenes as well as the guiding programs with scene_generation.py Exemplary program and scene description can be found at levels/mamt.

Training

There are sevel stages in training the whole e-mapp framework.

1. train the single agent single subtask policy bash scripts/train_sast.sh
2. train the perception module bash scripts/train_perception.sh
3. train the rechability function bash scripts/train_reach.sh
4. train cooperative policies bash scripts/train_cooperation.sh
5. train the feasibility function bash scripts/train_feas.sh
6. evaluate the complete e-mapp bash scripts/evaluate.sh

References

The code is built upon

• [Stable Baselines3]
• [Overcooked Env]

Citation

If you find this code helpful, please consider citing:

@inproceedings{changmapp,
  title={E-MAPP: Efficient Multi-Agent Reinforcement Learning with Parallel Program Guidance},
  author={Chang, Can and Mu, Ni and Wu, Jiajun and Pan, Ling and Xu, Huazhe},
  booktitle={Advances in Neural Information Processing Systems}
}