AlphaZero

This is a reimplementation of the AlphaZero algorithm as described by Silver et al. (2017).

We use PyTorch and Numpy as the primary backend.

The algorithm is applied to a custom game of repeated multi-dimensional TicTacToe, but could be easily adapted to play any other two-player perfect information game.

Main Features

1. PyTorch reimplementation of MCTS and policy training from scratch
2. Custom game environment following the Gym API
3. Parallel training workflow using Python's multiprocessing module