tengshiquan/doudizhu-C

C++/python fight the lord with pybind11 (强化学习AI斗地主)

Dou Di Zhu with Combinational Q-Learning

Step by step training tutorial

1. Clone the repo

git clone https://github.com/qq456cvb/doudizhu-C.git

2. Change work directory to root

cd doudizhu-C

3. Create env from environment.yml

conda env create -f environment.yml

4. Activate env

conda activate doudizhu

5. Build C++ files

mkdir build
cd build
cmake ..
make

6. Have fun training!

cd TensorPack/MA_Hierarchical_Q
python main.py

Evaluation against other baselines

1. Download pretrained model from https://jbox.sjtu.edu.cn/l/L04d4A, then put it into pretrained_model
2. Build Monte-Carlo baseline and move the lib into root

git clone https://github.com/qq456cvb/doudizhu-baseline.git
cd doudizhu-baseline/doudizhu
mkdir build
cd build
cmake ..
make
mv mct.cpython-36m-x86_64-linux-gnu.so [doudizhu-C ROOT]

3. Run evaluation scripts in scripts

cd scripts
python experiments.py

Directory Structure

• TensorPack contain different RL algorithms to train agents
• experiments contain scripts to evaluate agents' performance against other baselines
• simulator contain scripts to evaluate agents' performance against online gaming platform called "QQ Dou Di Zhu" (we provide it for academic use only, use it at your own risk!)

Miscellaneous

• We provide a Monte-Carlo-Tree-Search algorithm in https://github.com/qq456cvb/doudizhu-baseline
• We provide a configured Dou Di Zhu mini-server in https://github.com/qq456cvb/doudizhu-tornado for you to play interactively. NOTE you should build the server and load pretrained model by yourself! Tutorial coming soon!
• If you meet any problems, open an issue.

References

See our paper https://arxiv.org/pdf/1901.08925.pdf. If you find this algorithm useful or use part of its code in your projects, please consider cite

@article{DBLP:journals/corr/abs-1901-08925,
    author    = {Yang You and
                Liangwei Li and
                Baisong Guo and
                Weiming Wang and
                Cewu Lu},
    title     = {Combinational Q-Learning for Dou Di Zhu},
    journal   = {CoRR},
    volume    = {abs/1901.08925},
    year      = {2019},
    url       = {http://arxiv.org/abs/1901.08925},
    archivePrefix = {arXiv},
    eprint    = {1901.08925},
    timestamp = {Sat, 02 Feb 2019 16:56:00 +0100},
    biburl    = {https://dblp.org/rec/bib/journals/corr/abs-1901-08925},
    bibsource = {dblp computer science bibliography, https://dblp.org}
}