jielite / pysc2-a3c

Neural Grafting Network

DEMO

https://www.youtube.com/watch?v=u-AF5hc8l-o

Environment Setup

Our experiments are based on StarCraftII 3.16, PySC2, Pytorch, and handcrafted mini-games. We highly recommend you install the package through conda.

STEP1: Install StarCraftII 3.16

Please refer to the GitHub page of PySC2 to download and install StarCraftII 3.16.

NOTICE: If you are using AMD CPUs, it may not be compatible with our repo. When I used AMD 5950X to install the StarCraftII 3.16 and our handcrafted mini-games, it didn't work.

STEP2: Create a conda environment and install dependent packages

• Create a conda environment for Python3.6

conda create --name pysc2_torch18_py36 python=3.6
conda activate pysc2_torch18_py36

• Install Pytorch 1.8.1 LTS (Here, we take Ubuntu as an example)

conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch-lts -c nvidia

• Install PySC2 1.2

pip install pysc2==1.2
pip install pygame==1.9.6 ## Please refer to the troubleshooting as follows for more information

TROUBLE SHOOTING: pygame.error: Unable to make GL context current

STEP3: Install the handcrafted mini-games

Please copy the *.SC2Map files from Maps/mini_games/ to StarCraftII/Maps/mini_games/, and then go to the `pysc2

# TODO

STEP4: Verify the installation

Just play with it!

python -m pysc2.bin.play --map 

Handcrafted Mini-Games

# TODO

Train Reinforcement Learning Agent

# TODO

Results

# TODO

Reference

1. https://github.com/wing3s/pysc2-rl-mini
2. A3G: https://github.com/dgriff777/rl_a3c_pytorch

pysc2-a3c's Issues

Extract the setting of num of channels in CNN

as title

Extract task and non spatial actions

use if-else block to choose the action space