Infinite AI Tsumego Miner

"We do not build the puzzles; we mine them from the chaos of high-temperature logic."

The Infinite AI Tsumego Miner is an autonomous "mining rig" that uses adversarial AI self-play to generate high-quality Go (Weiqi/Baduk) puzzles.

Instead of randomly placing stones, it orchestrates an Arena where randomly selected superhuman Neural Networks play against each other with semi-randomized strength settings.

By varying temperature (creativity), visits (search depth), and playing style for each match, the system creates an infinite variety of unique game states. When one agent blunders—turning a winning game into a losing one—the system freezes time, analyzes the state with a stricter Referee engine, and extracts a Tsumego puzzle where the goal is to punish that specific mistake.

🌟 Key Features

• The Arena: Automates matchmaking between 11+ distinct neural network "personalities" (e.g., The Chameleon vs The Titan).
• Personality Profiles: Every match features randomized Temperature (0.8-1.5) and Visits (20-300) settings, ensuring no two games are ever alike.
• The Delta: Puzzles are detected by a drop in winrate (e.g., 99% → 10%) as judged by the unrelated High Referee network.
• Rich Metadata: Puzzles include SGF SGFs, policy distributions, wrong-move refutations, and detailed difficulty metrics.
• Headless Autonomy: Designed to run indefinitely, mining thousands of puzzles per day on a consumer GPU.

🏛️ System Architecture

graph TD
    %% Nodes
    Orch[Match Orchestrator]
    Bench[NetworkBench]
    Ref["The High Referee<br/>(KataGo b28)"]
    
    subgraph "The Arena (Randomized Self-Play)"
        P1["Player Black<br/>(Random Personality)"]
        P2["Player White<br/>(Random Personality)"]
    end

    Dec{"The Delta<br/>(Blunder?)"}
    Output[Extract Puzzle]

    %% Styles
    style Ref fill:#f9f,stroke:#333,stroke-width:2px,color:black
    style Output fill:#bbf,stroke:#333,stroke-width:2px,color:black
    style Dec fill:#ff9,stroke:#333,stroke-width:2px,color:black

    %% Flow
    Orch -->|Initialize| Bench
    Orch -->|Summon| Ref
    Bench -->|Generate Personalities| P1 & P2
    
    P1 <-->|Game Loop| P2
    
    P1 -.->|Propose Move| Ref
    Ref -->|Analyze State| Dec
    
    Dec -->|No| P2
    Dec -->|YES| Output
    Output -->|SGF/JSON| Orch

Loading

🛠️ Prerequisites

• OS: Windows 10/11 (Primary) or Linux.
• Hardware: NVIDIA GPU (RTX 3060 or better recommended). Developed on RTX 4070.
• Software: Python 3.11+.
• Dependencies: numpy

📂 Installation

1. Clone the Repository

   git clone https://github.com/MachineKomi/Infinite_AI_Tsumego_Miner.git
   cd Infinite_AI_Tsumego_Miner

2. Install Python Dependencies

   pip install -r requirements.txt

3. Asset Setup (Required) This miner requires specific KataGo binaries and model files.

   • KataGo: Put the Windows OpenCL binary in assets/katago/.
   • Models: Download the required neural networks and place them in assets/models/.

   📄 See assets/models/README.md for the exact list of required files and download links.

🚀 Usage

To start the mining session:

python src/miner.py

CLI Options

Command	Description
python src/miner.py --dry-run	Test connectivity to engines without playing games.
python src/miner.py --debug	Enable verbose logging for debugging.
python src/miner.py --log-file mining.log	output logs to a file.

Joseki Mining

To mine Joseki trees:

python src/joseki_miner.py

📚 Documentation

For deep dives into the architecture and philosophy:

• AGENTS.MD: The "Source of Truth" for AI agents, including the Model Registry and Personality Profiles.
• assets/models/README.md: Detailed list of supported networks and their roles.

⚖️ License

This project is licensed under the GNU Affero General Public License v3 (AGPLv3). See LICENSE for details.

---

Copyright (C) 2025