Implementing NNUE

[Captain3BoOd]

Implementing NNUE Support in GibnaMish

Description

As we know, NNUE (Efficiently Updatable Neural Networks) has significantly improved modern chess engines by enhancing evaluation accuracy and increasing playing strength. Our goal is to implement NNUE in this engine and train it properly to achieve competitive performance.

Tasks

• Implement NNUE architecture within the engine.
• Define appropriate input features.
• Create and train an initial neural network.
• Optimize inference performance for speed and efficiency.
• Improve training using high-quality game data.
• Compare performance against traditional evaluation and fine-tune parameters.

Benefits

• Improved evaluation accuracy within the engine.
• Enhanced playing strength, making it more competitive against modern NNUE engines.
• Ability to train custom networks based on different playing styles.

Recommended Resources

• Stockfish NNUE repositories for reference: Stockfish
• NNUE training guides: Guide

Additional Notes

Any suggestions regarding network architecture, training data, or further optimizations are welcome!