lightvector/KataGo

What it solves

KataGo is a high-performance open-source Go bot designed to be one of the strongest available. It solves the problem of providing a superhuman-level AI for the game of Go that can be trained from scratch using self-play, without requiring massive external datasets, and can be used for both high-level play and detailed game analysis.

How it works

KataGo uses an AlphaZero-like process for training, utilizing neural networks (including recent support for transformer models) and Monte-Carlo Graph Search (an extension of MCTS). It is designed for efficiency, allowing it to reach high amateur dan strength on a 19x19 board using only a few GPUs for a few days, or superhuman strength on a single consumer GPU over several months. It supports multiple hardware backends including OpenCL, CUDA, TensorRT, ROCm, ONNX Runtime, and Eigen (CPU).

Who it’s for

• Go players who want a strong opponent or a tool for game analysis.
• Developers whoต้องการ a GTP-compatible engine for Go software.
• AI researchers interested in self-play training and Monte-Carlo Graph Search.

Highlights

• Score and Territory Estimation: Unlike bots that only provide win rates, KataGo estimates territory and score, making it useful for analyzing amateur games.
• Broad Support: Supports board sizes from 7x7 to 19x19 and a wide variety of rulesets.
• Hardware Flexibility: Offers six different backends to ensure compatibility across NVIDIA, AMD, Intel GPUs, and CPUs.
• Efficient Training: Capable of training from scratch to superhuman levels using relatively modest hardware.
• Human-style Play: Can imitate human play at various ranks using a specific human SL model.