BlackBird

BlackBird is a self-learning algorithm designed for board games, based on the AlphaZero implementation. Future versions of this project will include extensions from state-of-the-art research into reinforcement learning optimization.

Code-level documentation can be found here.

Requirements and Suggested Tools

• This project is not compatible with Python 2.x; please use 3.5+.
• It is strongly advised that you use the GPU version of tensorflow, should you have a compatible Nvidia GPU. If you have not yet installed this, you can follow the directions here.

Getting Started

Depending on the level of interaction you want with BlackBird, you can either run the default setup, or modify the parameters and the game that BlackBird learns.

• The learning script main.py expects that there is a file called parameters.yaml, which describes the setup for the algorithm to learn from. By default, this file does not exist, but you should copy the parameters_template.yaml file into parameters.yaml, and modify it to your liking.
• Because this project builds its documentation automatically with Sphinx, note that there are Sphinx-related packages in the requirements.txt file.
• The packages that are required to run BlackBird may be unrelated to your workflow elsewhere, so it is recommended that you follow these steps to create a virtual environment in which to install them.
  1. pip install virtualenv
  2. python -m virtualenv blackbird
  3. source blackbird/bin/activate

Quickstart

To get started quickly, run the following commands in a terminal shell.

$ git clone https://github.com/ZackWolf614/BlackBird
$ cd BlackBird
$ pip install -r requirements.txt
$ cp parameters_template.yaml parameters.yaml
$ python main.py

Tuning Parameters

The parameters.yaml file contains the setup for BlackBird's neural network architecture, and how it learns from its self-generated games. The major parameter headings are

• mcts: Settings determining how the Monte Carlo Tree Search algorithm performs.
• network: Settings determining the structure of the neural network, and learning parameters.
• selfplay: Settings determining how many games BlackBird should play against itself - for game generation and for performance testing.