/alpha-zero-general

A clean implementation based on AlphaZero for any game in any framework + tutorial + Othello/Gobang/TicTacToe/Connect4/PickGame

Primary LanguagePythonMIT LicenseMIT

PickGame implementation for Alpha Zero General

An implementation of a simple game provided to check extendability of the framework. Neural network architecture was copy-pasted from the game of Othello, so possibly it can be simplified.

Coach.py contains the core training loop and MCTS.py performs the Monte Carlo Tree Search. The parameters for the self-play can be specified in main.py. Additional neural network parameters are in pickgame/keras/NNet.py (cuda flag, batch size, epochs, learning rate etc.).

To start training a model for PickGame:

python main.py

To play againt the model-based player:

python pit.py

You can choose players of the game (human / random / model) by switching flags in pit.py

Environment Requirement

Python3, NumPy, Keras Library, CUDA Toolkit (optional)

Experiments

I trained a Keras model for 4x4 PickGame (45 iterations, 100 episodes, 10 epochs per iteration and 50 MCTS simulations per turn). This took about 40 hours on an i5-8250U with MX150 CUDA. The pretrained model (Keras) can be found in pretrained_models/pickgame/keras/. You can play a game against it using pit.py.

Contributors and Credits

The implementation is based on this project.

A concise description of this algorithm can be found here.