Introducing Luna: My Rust-based NNUE Chess Engine

[Spunc595]

​Hello everyone,
​I am thrilled to have been accepted into this community! I would like to take this opportunity to introduce my chess engine, Luna.
​I have been developing Luna from scratch using Rust, focusing on creating a robust and efficient architecture. One of the core components I am particularly proud of is the NNUE evaluation; I have trained the network personally using a custom pipeline written in Python.
​The engine currently supports the standard UCI protocol, and I have implemented features like a bitwise-masked transposition table to enhance search performance. As I am still actively working on optimizations and refining the search algorithms, I would truly appreciate any feedback, constructive criticism, or advice from the experienced developers here.
​You can find the source code and the current status of the project on my GitHub repository:
https://github.com/Spunc595/Luna-Chess-Engine
​I look forward to learning from this community and contributing where I can. Thank you all for your time!
​Best regards,
​Daniele

[Sylwy]

​Hello everyone,
​I am thrilled to have been accepted into this community! I would like to take this opportunity to introduce my chess engine, Luna.
​I have been developing Luna from scratch using Rust, focusing on creating a robust and efficient architecture. One of the core components I am particularly proud of is the NNUE evaluation; I have trained the network personally using a custom pipeline written in Python.
​The engine currently supports the standard UCI protocol, and I have implemented features like a bitwise-masked transposition table to enhance search performance. As I am still actively working on optimizations and refining the search algorithms, I would truly appreciate any feedback, constructive criticism, or advice from the experienced developers here.
​You can find the source code and the current status of the project on my GitHub repository:
https://github.com/Spunc595/Luna-Chess-Engine
​I look forward to learning from this community and contributing where I can. Thank you all for your time!
​Best regards,
​Daniele

viewtopic.php?start=820&t=81223&sid=40b ... 715beeaf45

1.- let static_eval = crate::evaluation::evaluate(board); ...or,

let static_eval = if let Some(n) = nnue {
n.evaluate(board)
} else {
crate::evaluation::evaluate(board)
};

2.- let stand_pat = crate::evaluation::evaluate(board); ...or,

let stand_pat = if let Some(n) = nnue {
n.evaluate(board)
} else {
crate::evaluation::evaluate(board)
};

???

[Spunc595]

Hi.

I am actually using the second approach. Since the NNUE is passed as an Option, I implemented a fallback to the classical evaluation function in both the main search (for static_eval) and the quiescence search (for stand_pat) in case the network file is missing or fails to load.

This allows me to test and compare the search behavior with and without the neural network weights.

[Graham Banks]

There is already a chess engine named Luna.

[Spunc595]

Thanks for pointing that out! Yes, I’m aware that "Luna" is a beautiful and popular name, and there might be older or other projects out there sharing it.

To avoid any confusion, my project is officially named Luna Chess Engine (LCE). It’s a completely new, independent engine written from scratch in Rust, featuring a custom NNUE evaluation. I'll make sure to use the full name or the acronym LCE in the future to keep things distinct!

[Sylwy]

There is already a chess engine named Luna.

This engine is named (exactly) Luna Chess Engine (LCE) (v1.0.0). The other is Luna (2.0.0)-Thomas Mergener chess engine.

[Spunc595]

Exactly, thanks Sylwy for clearing that up!

Luna Chess Engine (LCE) is a completely independent, modern project built from scratch in Rust with a custom NNUE evaluation. There is no connection with the older engine by Thomas Mergener. Thanks everyone for the support!

[Graham Banks]

Could you please give more information regarding your net?

[tomitank]

Could you please give more information regarding your net?

// --- NNUE ARCHITECTURE CONFIGURATION ---
// 768 input features (12 piece types * 64 squares)[cite: 5].
const INPUT_SIZE: usize = 768;
// Size of the first hidden layer (Feature Transformer)[cite: 5].
const L1_SIZE: usize = 256;
// Size of the second hidden layer[cite: 5].
const L2_SIZE: usize = 32;

short: 768x256x32x1