Learning Implicit Credit Assignment for Cooperative Multi-Agent Reinforcement Learning (LICA)

This repository hosts the official PyTorch implementation of "Learning Implicit Credit Assignment for Cooperative Multi-Agent Reinforcement Learning". The implementation is based on the PyMARL framework and SMAC.

[PDF]

Setup

Set up the working environment:

pip3 install -r requirements.txt

Set up the StarCraftII game core

bash install_sc2.sh

Training

To train LICA on the 5m_vs_6m scenario,

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

Change the map_name accordingly for other scenarios (e.g. mmm2). Remember to update hyperparameters accordingly (e.g. regularization coefficient). All results will be saved in the results folder.

Evaluation

TensorBoard

The training tensorboards are saved in the results/tb_logs directory, containing useful info such as policy entropy trajectory and test battle win rate during training.

Saving models

Same as PyMARL, set save_model to True in src/config/default.yaml will save the learnt model during training (saved in result/ directory be default). The frequency for saving models can be adjusted using the parameter save_model_interval.

Loading models

Saved models can be loaded by adjusting the checkpoint_path parameter in src/config/default.yaml. For instance, to load model under path result/model/[timesteps]/agent.th, set checkpoint_path to result/model/[timesteps].

Saving Starcraft II Replay

The learned model loaded from checkpoint_path can be evaluated by setting evaluate to True in src/config/default.yaml. To save the Starcraft II replays, please make sure configure save_replay to True, and use the episode_runner.

Check out PyMARL documentation for more information.

Visualizations

The demo videos of two of the best battles performed by LICA agents (red) are available at visualizations/. The GIF previews are shown below:

5m_vs_6m	MMM2

Citation

Please cite this work if you find it useful:

@article{zhou2020learning,
  title={Learning Implicit Credit Assignment for Cooperative Multi-Agent Reinforcement Learning}, 
  author={Meng Zhou and Ziyu Liu and Pengwei Sui and Yixuan Li and Yuk Ying Chung},
  journal={arXiv preprint arXiv:2007.02529},
  year={2020}
}

Contact

Please email mzho7212 AT gmail.com and kenziyuliu AT outlook.com for further questions.

mzho7212/LICA