Other Go AI Introduction
Besides AlphaGo and KataGo, there are many important projects in the Go AI field. This article introduces major commercial and open-source AIs to help you understand the entire ecosystem.
Commercial Go AI
Zen
Developer: Yoji Ojima / Japan First Release: 2009 License: Commercial
Zen was one of the strongest Go programs before AlphaGo, reaching professional level in the traditional MCTS era.
Development History
| Time | Version | Milestone |
|---|---|---|
| 2009 | Zen 1.0 | First release |
| 2011 | Zen 4 | Reached amateur 6-dan level |
| 2012 | Zen 5 | Defeated Takemiya Masaki 9-dan with 4 stones |
| 2016 | Zen 7 | Adopted deep learning technology |
| 2017+ | Deep Zen Go | Combined AlphaGo architecture |
Technical Features
- Hybrid architecture: Combines traditional heuristics with deep learning
- Commercial optimization: Optimized for consumer hardware
- High stability: Verified through years of commercial use
- Multi-platform: Runs on Windows, macOS
Product Forms
- Tengen (Tencho no Go): Desktop software, ~10,000 JPY
- Online play: Active on KGS under Zen19 account
Fine Art (Jue Yi)
Developer: Tencent AI Lab / China First Release: 2016 License: Not publicly available
Fine Art is Tencent's Go AI with significant influence in China's Go community.
Development History
| Time | Event |
|---|---|
| November 2016 | First appearance, playing on Fox Go network |
| March 2017 | UEC Cup Computer Go Tournament champion |
| 2017 | Adopted by Chinese National Go Team as training tool |
| 2018 | World AI Go Championship winner |
| Continuing | Continues as national team training aid |
Technical Features
- Large-scale training: Uses Tencent cloud computing resources
- Collaboration with top human players: Receives extensive professional guidance
- Rich practical experience: Accumulated many games on Fox Go
- Teaching function integration: Provides review analysis features
Influence
Fine Art's impact on Chinese professional Go is profound:
- Became national team standard training tool
- Changed how professionals prepare for matches
- Promoted widespread AI-assisted training
Golaxy (Xing Zhen)
Developer: DeepMind China / Tsinghua University team First Release: 2018 License: Commercial
Golaxy was designed with the goal of being the "most human-like AI," with playing style closer to human players.
Technical Features
- Human-like style: Deliberately trained to play more like humans
- Adjustable difficulty: Can simulate opponents of different ranks
- Teaching-oriented: Designed with teaching applications in mind
- Handicap game specialist: Special optimization for handicap games
Product Applications
- Yike Go: Integrated into Yike App
- Teaching platform: Used for online Go teaching
- Rank testing: Provides standardized rank assessment
Other Commercial AI
| Name | Developer | Features |
|---|---|---|
| CGI | NCTU (Taiwan) | Academic research oriented |
| Dolbaram | NHN (Korea) | Integrated into Korean Go platforms |
| AQ | Japan AQ Team | Open source then went commercial |
Open Source Go AI
Leela Zero
Developer: Gian-Carlo Pascutto / Belgium First Release: 2017 License: GPL-3.0 GitHub: https://github.com/leela-zero/leela-zero
Leela Zero was the first successful open-source project to replicate AlphaGo Zero, trained through distributed community effort.
Development History
Technical Features
- Faithful reproduction: Strictly implements AlphaGo Zero paper
- Distributed training: Global volunteers contribute GPU computing
- Fully transparent: All training data and models are public
- Standard GTP: Compatible with all GTP Go software
Training Statistics
| Item | Value |
|---|---|
| Total self-play games | ~18 million |
| Training iterations | ~270 |
| Contributors | Thousands |
| Training duration | ~1.5 years |
Usage
# Installation
brew install leela-zero # macOS
# Run
leelaz --gtp --weights best-network.gz
# GTP commands
genmove black
play white D4
Current Status
Although Leela Zero is no longer actively training:
- Code remains an excellent resource for learning AlphaGo Zero
- Trained models are still usable
- Community still maintains basic functionality
ELF OpenGo
Developer: Facebook AI Research (FAIR) First Release: 2018 License: BSD GitHub: https://github.com/pytorch/ELF
ELF OpenGo is Facebook's Go AI demonstrating large-scale distributed training capabilities.
Technical Features
- ELF framework: Based on Facebook's ELF (Extensive, Lightweight, and Flexible) game research platform
- Large-scale training: Uses 2000 GPUs for training
- PyTorch implementation: Uses Facebook's own deep learning framework
- Research oriented: Main purpose is research, not practical use
Performance
- Reached top level on KGS
- Stable win rate against professional 9-dan
- Paper published at top conference
Current Status
- Project no longer actively maintained
- Code and models still available for download
- Main value is academic reference
SAI (Sensible Artificial Intelligence)
Developer: SAI Team / Europe First Release: 2019 License: MIT GitHub: https://github.com/sai-dev/sai
SAI is an improved version based on Leela Zero, focusing on experimental features.
Technical Features
- Improved training methods: Experiments with various training optimizations
- More rule support: Supports more Go rules than Leela Zero
- Experimental features: Tests new network architectures and training tricks
Current Status
- Still has small community maintaining it
- Mainly used for experiments and learning
PhoenixGo
Developer: Tencent WeChat Team First Release: 2018 License: BSD-3 GitHub: https://github.com/Tencent/PhoenixGo
PhoenixGo is Tencent's open-sourced Go AI, winning the 2018 World AI Go Championship.
Technical Features
- Commercial grade quality: From Tencent internal project
- TensorFlow implementation: Uses mainstream framework
- Multi-platform support: Linux, Windows, macOS
- Distributed support: Can run in multi-machine multi-GPU environment
Usage
# Build
bazel build //src:mcts_main
# Run
./mcts_main --gtp --config_path=config.conf
MiniGo
Developer: Google Brain First Release: 2018 License: Apache-2.0 GitHub: https://github.com/tensorflow/minigo
MiniGo is Google's open-sourced educational Go AI, designed to help more people understand AlphaGo principles.
Technical Features
- Education oriented: Clear, readable code
- TensorFlow implementation: Official Google example
- Complete documentation: Has detailed technical explanations
- Colab support: Can run directly in Google Colab
Use Cases
- Learning AlphaGo Zero architecture
- Understanding reinforcement learning in games
- As starting point for your own projects
Comparison of AI Features
Strength Comparison (Approximate)
| AI | Strength Level | Notes |
|---|---|---|
| KataGo | Top superhuman | Continuously training |
| Fine Art | Top superhuman | Not public |
| Leela Zero | Superhuman | Training stopped |
| ELF OpenGo | Superhuman | Training stopped |
| PhoenixGo | Near superhuman | Training stopped |
| Zen | Professional level | Commercial product |
| Golaxy | Professional level | Adjustable difficulty |
Feature Comparison
| Feature | KataGo | Leela Zero | PhoenixGo | Zen |
|---|---|---|---|---|
| Open source | Yes | Yes | Yes | No |
| Score prediction | Yes | No | No | Partial |
| Multi-rule support | Yes | No | No | No |
| Analysis API | Yes | No | No | No |
| CPU mode | Yes | Yes | Yes | Yes |
| Active updates | Yes | No | No | Partial |
Use Case Recommendations
| Need | Recommended | Reason |
|---|---|---|
| General play/analysis | KataGo | Strongest with most features |
| Learning AlphaGo | Leela Zero / MiniGo | Clear code |
| Commercial use | Zen / Self-trained KataGo | Clear licensing |
| Teaching aid | KataGo / Golaxy | Rich analysis features |
| Research experiments | KataGo / SAI | Can modify training |
Future Development Trends
Technical Trends
-
More efficient training methods
- As demonstrated by KataGo's efficiency improvements
- Fewer resources to achieve higher strength
-
Better interpretability
- Explain why AI plays certain moves
- Help humans understand AI's thinking
-
Combining with human styles
- Train AI to play like specific players
- For teaching and research
-
Cross-game generalization
- As demonstrated by AlphaZero
- Single framework for multiple games
Application Trends
-
Democratization
- More Go enthusiasts using AI analysis
- Even phones can run it
-
Professionalization
- Professional players deeply rely on AI training
- AI assistance becomes standardized
-
Commercialization
- More AI-assisted Go products
- Teaching, analysis, training partner services
Summary
The Go AI ecosystem is rich and diverse:
- Want strongest strength and most features: Choose KataGo
- Want to learn AI principles: Study Leela Zero or MiniGo code
- Commercial use needs: Evaluate Zen or train your own model
- Special needs: Choose or combine based on specific requirements
Next, let's move to practical content and learn how to install and use KataGo!