Background Knowledge Overview

Before diving into KataGo practice, understanding the development history and core technology of Go AI is very important. This chapter will guide you through the technological evolution from AlphaGo to modern Go AI.

Why Learn Background Knowledge?

The development of Go AI is one of the most exciting breakthroughs in artificial intelligence. The 2016 match where AlphaGo defeated Lee Sedol was not only a milestone in Go history but also marked the tremendous success of combining deep learning with reinforcement learning.

Understanding this background knowledge helps you:

Make better technical decisions: Understand the pros and cons of various approaches, choose the right solution for your project
Debug more effectively: Understanding underlying principles makes diagnosing problems easier
Keep up with latest developments: Mastering fundamentals makes it easier to understand new papers and technologies
Contribute to open source projects: Participating in KataGo and similar project development requires deep understanding of design philosophy

Chapter Contents

AlphaGo Paper Analysis

In-depth analysis of DeepMind's classic paper, including:

Historical significance and impact of AlphaGo
Policy Network and Value Network design
Monte Carlo Tree Search (MCTS) principles and implementation
Self-play training method innovation
Evolution from AlphaGo to AlphaGo Zero to AlphaZero

KataGo Paper Analysis

Understanding the technical innovations of the current strongest open-source Go AI:

KataGo improvements over AlphaGo
More efficient training methods and resource utilization
Technical implementation for supporting multiple Go rules
Design for simultaneously predicting win rate and score
Why KataGo achieves stronger play with fewer resources

Other Go AI Introduction

Comprehensive understanding of the Go AI ecosystem:

Commercial AI: Zen, Fine Art (Tencent), Golaxy
Open source AI: Leela Zero, ELF OpenGo, SAI
Comparison of technical features and use cases for each AI

Technical Development Timeline

Time	Event	Significance
October 2015	AlphaGo defeats Fan Hui	First AI defeat of professional player
March 2016	AlphaGo defeats Lee Sedol	World-shaking human-machine match
May 2017	AlphaGo defeats Ke Jie	Confirms AI surpasses top human level
October 2017	AlphaGo Zero published	Pure self-play, no human games needed
December 2017	AlphaZero published	Generalized design, conquers Go, chess, and shogi
2018	Leela Zero reaches superhuman level	Open source community victory
2019	KataGo published	More efficient training methods
2020-present	KataGo continues improving	Becomes strongest open-source Go AI

Core Concepts Preview

Before reading detailed chapters, here's a brief introduction to several core concepts:

Role of Neural Networks in Go

Board state → Neural Network → { Policy (move probabilities), Value (win rate estimate) }

The neural network receives current board state as input and outputs two types of information:

Policy: Probability of playing at each position, guides search direction
Value: Win rate estimate of current position, used to evaluate positions

Monte Carlo Tree Search (MCTS)

MCTS is a search algorithm that combines with neural networks to determine the best move:

Selection: From root node, select the most promising path
Expansion: At leaf node, expand new possible moves
Evaluation: Use neural network to evaluate position value
Backpropagation: Pass evaluation results back to update nodes along the path

Self-play

AI plays against itself to generate training data:

Initial model → Self-play → Collect games → Train new model → Stronger model → Repeat

This cycle allows AI to continuously improve itself without depending on human games.

Why Learn Background Knowledge?​

Chapter Contents​

AlphaGo Paper Analysis​

KataGo Paper Analysis​

Other Go AI Introduction​

Technical Development Timeline​

Core Concepts Preview​

Role of Neural Networks in Go​

Monte Carlo Tree Search (MCTS)​

Self-play​

Recommended Reading Order​