A "build to learn" implementation of the Alpha Zero algorithm written in Python that uses LightGBM (Gradient Boosted Decision Trees) in place of a Deep Neural Network for value/policy functions.

A few environments (i.e., games) are implemented: Quoridor, Connect Four, and Tic-Tac-Toe.

python3.7 play.py <environment> <species-generation> <species-generation> <time-to-move>

python3.7 play.py connect_four gbdt-1 human-1 5.0

python3.7 train_bot.py <environment> <species> <num batches>

python3.7 train_bot.py connect_four gbdt 10

If you haven't already, install pyenv/pyenv-virtualenv (see Install pyenv/pyenv-virtualenv below)

Clone repo:

git clone [email protected]:cgreer/alpha-zero-boosted.git

Create a virtual environment for the project:

cd alpha-zero-boosted
pyenv install 3.7.7
pyenv virtualenv 3.7.7 alpha_boosted_env
pyenv local alpha_boosted_env

Install packages:

pip install -r requirements.txt

MacOS Issue: Because some wheels don't appear to be built properly, you may need to first install libomp and then retry installing packages:

brew install libomp

# Then try installing again
pip install -r requirements.txt

Install the plugin:

brew install pyenv

Note: These instructions are copied here for convenience. Check pyenv-virtualenv to ensure they are up to date.

Install the plugin:

brew install pyenv-virtualenv

Add the following two lines to your profile file (~/.zprofile if using zsh, ~/.bash_profile if bash):

eval "$(pyenv init -)"
eval "$(pyenv virtualenv-init -)"

Restart terminal (so profile commands above execute).