KataGo in Practice

KataGo is the strongest open-source Go engine ever built -- and the most instructive case study in modern AI engineering. This book takes you inside every layer of the system, from raw board positions to production-grade inference, giving you the deep technical understanding that no tutorial or README can provide.

Whether you are a machine learning engineer curious about how AlphaZero-style self-play actually works at scale, a Go enthusiast who wants to understand what your analysis engine is really doing, or a software architect studying how research prototypes become production systems, this book was written for you.

What you will learn:

• How KataGo's neural network architecture evolved from simple residual towers to the global-pooling and nested-bottleneck designs that outperform DeepMind's original AlphaGo Zero
• The complete Monte Carlo Tree Search (MCTS) pipeline -- from UCB selection and virtual losses to the PUCT formula and how KataGo's search differs from vanilla AlphaZero
• KataGo's self-play training loop: game generation, position sampling, data augmentation, and the curriculum strategies that let it reach superhuman strength on consumer hardware
• Ownership, territory, and score estimation heads -- the auxiliary predictions that make KataGo uniquely useful for analysis and teaching
• The GTP protocol, KataGo's analysis engine, and how to integrate the engine into your own applications via JSON queries
• Production deployment patterns: TensorRT and OpenCL backends, batched inference, multi-GPU scaling, and performance tuning for real-world workloads
• How to extend and modify KataGo -- custom training runs, network surgery, rule variants, and contributing to the open-source project