UPDeT

Official Implementation of UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers (ICLR 2021 spotlight)

The framework is inherited from PyMARL. UPDeT is written in pytorch and uses SMAC as its environment.

Installation instructions

Installing dependencies:

pip install -r requirements.txt

Download SC2 into the 3rdparty/ folder and copy the maps necessary to run over.

bash install_sc2.sh

Run an experiment

Before training your own transformer-based multi-agent model, there are a list of things to note.

• Currently, this repository supports marine-based battle scenarios. e.g. 3m, 8m, 5m_vs_6m.
• If you are interested in training a different unit type, carefully modify the Transformer Parameters block at src/config/default.yaml and revise the _build_input_transformer function in basic_controller.python.
• Before running the experiment, check the agent type in Agent Parameters block at src/config/default.yaml.
• This repository contains two new transformer-based agents from the UPDeT paper including
• Standard UPDeT
• Aggregation Transformer

Training script

python3 src/main.py --config=vdn --env-config=sc2 with env_args.map_name=5m_vs_6m

All results will be stored in the Results/ folder.

Performance

Single battle scenario

Surpass the GRU baseline on hard 5m_vs_6m with: - QMIX: QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning - VDN: Value-Decomposition Networks For Cooperative Multi-Agent Learning - QTRAN: QTRAN: Learning to Factorize with Transformation for Cooperative Multi-Agent Reinforcement Learning

Multiple battle scenarios

Zero-shot generalize to different tasks:

• Result on 7m-5m-3m transfer learning.

Note: Only UPDeT can be deployed to other scenarios without changing the model's architecture.

More details please refer to UPDeT paper.

Bibtex

@article{hu2021updet,
  title={UPDeT: Universal Multi-agent Reinforcement Learning via Policy Decoupling with Transformers},
  author={Hu, Siyi and Zhu, Fengda and Chang, Xiaojun and Liang, Xiaodan},
  journal={arXiv preprint arXiv:2101.08001},
  year={2021}
}

License

The MIT License