AlphaZero learns Xiangqi, the hard way

What it does

CCZero is an AlphaZero-style reinforcement-learning pipeline for Chinese chess (Xiangqi). It runs self-play games, trains neural nets on the results, and pits new models against the current champion in an endless loop. There is also a built-in pygame GUI so you can lose to the machine in style.

The interesting bit

The project is deliberately a distributed volunteer effort — the README openly admits you need “a huge amount of computations” and asks strangers to run self and opt workers to feed a central model pool. It is essentially SETI@home, but for training a board-game engine.

Key highlights

• Four worker types: self (self-play), opt (training), eval (model vs. model), and sl (supervised warm-start from internet game databases)
• Supports both UCI engine mode and a built-in GUI with swappable piece/board themes (WOOD, QIANHONG, SKELETON, etc.)
• Distributed mode uploads play data and downloads the latest model automatically
• TensorBoard logging and ELO tracking via https://cczero.org
• Windows executables available for non-developers

Caveats

• Stuck on a very specific legacy stack: Python 3.6.3, TensorFlow-GPU 1.3.0, Keras 2.0.8
• You must manually source a PingFang.ttc font file before the GUI will run; the author removed it for size
• The README is thorough but reads like a wiki dump; finding the right --type flag for distributed training takes patience

Verdict

Worth a look if you want to study AlphaZero mechanics on a smaller, more tractable game than Go, or if you have spare GPU cycles and a fondness for Xiangqi. Everyone else should probably use a modern Leela Chess Zero derivative instead.