We used a scaled down alpha-zero approach implemented in python from suragnair/alpha-zero-general to train a chess bot on a simple chess variant we call halfchess. This layout of this variant is depicted here, though we also enforce an additional restriction that pieces can only move forward or laterally, not backwards. This greatly simplifies the training process (making it somewhat feasible for a non-distributed alpha zero approach implemented entirely in python) by reducing the game length and total action space.

Joseph Green and James Lee

• alpha-zero-general - the main framework we adapted to train a halfchess network
• python-chess - the underlying chess engine we adapted to fit our halfchess logic
• PySimpleGUI - we use this for a GUI to play the network with