Chess Training Dashboard & AI Play

A high-performance, refactored PyQt6 GUI for monitoring chess training and playing against the AI.It will randomly plateau at like 100 ish episodes

Prerequisites

• Linux (Tested on Ubuntu/Debian based systems)
• Python 3.10+
• Stockfish (Chess engine for evaluation and playback)

Installation

1. Install System Dependencies:

   sudo apt update
   sudo apt install stockfish libxcb-cursor0

   Note: libxcb-cursor0 is often needed for PyQt6 on some Linux distributions.

2. Navigate to the Project: Ensure you are in the project directory:

   cd /home/mathias/2tb/Projects

3. Create a Virtual Environment (Recommended):

   python3 -m venv venv
   source venv/bin/activate

4. Install Python Dependencies:

   pip install -r requirements.txt

Usage

1. Launch the Dashboard

To start the main application (Training Monitor + Play Zone):

python dashboard_gui.py

2. Features

• Left Panel (Monitoring):

  • Hardware Vitals: Real-time CPU, RAM, GPU, and VRAM tracking.
  • Training Stats: Monitor samples collected and processing speed.
  • Start/Stop Training: Toggle the background training process (train_az.py).
  • Theme: Instant Dark/Light mode switching.

• Right Panel (Play Zone):

  • Interactive Board: Smooth, responsive piece movement.
  • Play vs AI: Challenge the latest MonkeyChessAi Mark2 checkpoint (800 MCTS simulations).
  • Live Evaluation: Integrated Stockfish evaluation bar.

• Bottom Panel (Analytics):

  • Persistent Metrics: Interactive, multi-color graphs tracking loss and sample stats across sessions.
  • Log Throttling: High-performance terminal output designed to prevent UI lag during intense training phases.

Automated Elo Benchmarking

The suite includes a surgical, high-speed Elo estimation system:

• Automatic Trigger: Benchmarking occurs every 250 training episodes.
• Manual Evaluation: Toggle estimation with "ESTIMATE ELO" and "RESUME" buttons.
• Ultra-Fast Matchmaking: Plays 5 parallel games against Stockfish on core 12 for near-instant strength checks.
• Visualization: Live progress and persistent Elo growth curves on the main dashboard.

Advanced Training Architecture

1. Multi-PV Margin Selection

Workers (mcts_worker.py) use an advanced policy:

• Stockfish calculates top 3 moves (Multi-PV = 3).
• Moves within a 50cp margin are selected with weighted probability, ensuring high-quality diversity.

2. Opening Book Support

• Uses super_gm_book.pgn to guide the first 8-12 moves (16-24 plies).
• Prevents line memorization and ensures diverse training data.

3. Performance Optimizations

• Geometric Augmentation: Horizontally flips board positions to double dataset size for free.
• Zero-Copy Shared Memory: 300k+ positions passed between processes with zero overhead.
• Efficient Inference: board_to_tensor optimized for maximum throughput.

Troubleshooting

• Stockfish not found: The app looks for Stockfish at /usr/games/stockfish. Ensure it's in your system PATH.
• PyQt6 xcb-cursor error: Run sudo apt install libxcb-cursor0 if the GUI fails to launch.

---

Optimized for 32GB RAM / 4GB VRAM environments.

Advanced Training Features

1. Multi-PV Margin Selection

The training workers (mcts_worker.py) use an advanced move selection strategy:

• Stockfish calculates the top 3 moves (Multi-PV = 3).
• The script identifies all moves within 50 centipawns (0.5 points) of the best move.
• A move is picked randomly from this "Grandmaster approved" set, providing variety without sacrificing quality.

2. Opening Book Support

To prevent the AI from playing the same opening lines repeatedly:

• Uses a super_gm_book.pgn containing high-level grandmaster games.
• Workers follow random game lines for the first 8-12 moves (16-24 plies).
• This ensures the model trains on thousands of diverse, high-quality opening positions.

Troubleshooting

• Stockfish not found: The app looks for Stockfish at /usr/games/stockfish or in your system PATH. If installed elsewhere, ensure it's in your PATH or update engine_connector.py.
• PyQt6 xcb-cursor error: If the GUI doesn't open, install the missing library: sudo apt install libxcb-cursor0.

Elite Features & Robustness (Latest Update)

1. Geometric Data Augmentation (2x Speed)

The training pipeline now automatically flips every board position horizontally, effectively doubling the dataset size without any extra computational cost. This ensures the bot learns patterns symmetrically for both White and Black.

2. Chaos Mode (Human Robustness)

To prevent the AI from overfitting to "perfect" play, we inject controlled chaos into the training data:

• Blunder Injection (5%): The bot is forced to play random moves occasionally to learn how to recover from mistakes.
• Sub-optimal Paths (20%): The bot takes non-optimal but valid paths to see a wider variety of positions.

3. Value Jitter & Softmax Policy

• Value Jitter: Small random noise is added to evaluation targets to prevent memorization.
• Softmax Selection: Moves are selected probabilistically based on their strength (Temperature=50), rather than a hard cutoff.

4. Advanced MCTS Search

• Dirichlet Noise: Adds creativity to the opening phase.
• First Play Urgency (FPU): Optimizes search exploration for unvisited nodes.