https://cp4space.hatsya.com/2021/01/08/the-neural-network-of-t | The lazy guy channel

https://cp4space.hatsya.com/2021/01/08/the-neural-network-of-the-stockfish-chess-engine/

I’ve stumbled on a description of (extremely simple) NNUE neural network architecture used in current strongest chess engine – Stockfish 13. This is interesting because in 2017 Stockfish was beaten by Google’s AlphaZero and the time it seemed that chess is “solved” and deep neural nets are the future of board games. But that was not an actual end of story.

Since then, AlphaZero paper has been implemented in open source Leela Zero chess engine and Stockfish continued to steadily increase. So, these two are competing with various success for the first place in top chess engine championship for the last 7 seasons. And it really is very interesting which approach is better – “classic” brute-force like approach which evaluates tens of millions of positions per move using shallow network or “modern” Monte Carlo tree search which evaluates tens of thousands of positions using very deep network. It is not clear at all.

As a side note – both Stockfish and Leela chess are opensource and feature very high-quality C++ code – both readable and very performant (a rare combination). Very recommended reading if you write C++ for living.

Useful links.
[1] Last top chess engine championship results: https://en.wikipedia.org/wiki/Top_Chess_Engine_Championship#Tournament_results_(TCEC)
[2] Stockfish source: https://github.com/official-stockfish/Stockfish
[3] lc0 source: https://github.com/LeelaChessZero/lc0