The neural network of the stockfish chess engine

• ChessCounter

  7 11 0.0 C++

  Estimate the number of legal chess positions

You don't need to care about possible games, but game states, what's the best result and the move to achieve it in each. Another user computed an upper bound at 8.7E+45 positions: https://github.com/lechmazur/ChessCounter . I pointed out that our planet has about 1E+50 atoms. And we can further compress the database by perhaps two orders of magnitude if we don't store symmetric positions or positions close to mate. A Kardashev 2 civilization could play perfect chess.

• fishtest

  23 255 9.1 Python

  The Stockfish testing framework

  

The NUE architecture was only put in place around August. If you see https://github.com/glinscott/fishtest/wiki/Regression-Tests you'll notice that gave a very significant (100+ ELO) boost.

There is a current tournament going on at tcec-chess.com/ which stockfish has been leading so far, but I see Leela has just caught up in the head to head.

Of course Leela also keeps evolving.

• InfluxDB

  www.influxdata.com featured

  Power Real-Time Data Analytics at Scale. Get real-time insights from all types of time series data with InfluxDB. Ingest, query, and analyze billions of data points in real-time with unbounded cardinality.

  

• Ceres

  4 143 9.5 C#

  Ceres - an MCTS chess engine for research and recreation

That being said it is not clear which approach is better e.g. a very smart but relatively slow evaluation function (Leela) or SFs approach to use a relatively dumb eval function with a more sophisticated search. It is pretty clear that Leelas search can be improved (Check https://github.com/dje-dev/Ceres for example)

• nnue-pytorch

  14 284 6.3 C++

  Stockfish NNUE (Chess evaluation) trainer in Pytorch

  

If anyone wants to experiment with training these nets, it's a great way to get exposed to a nice mix of chess and machine learning.

There are two trainers currently, the original one, which runs on CPU: https://github.com/nodchip/Stockfish, and a pytorch one which runs on GPU: https://github.com/glinscott/nnue-pytorch.

The SF Discord is where all of the discussion/development is happening: https://discord.gg/KGfhSJd.

Right now there is a lot of experimentation to try adjusting the network architecture. The current leading approach is a much larger net which takes in attack information per square (eg. is this piece attacked by more pieces than it's defended by?). That network is a little slower, but the additional information seems to be enough to be stronger than the current architecture.

Btw, the original Shogi developers really did something amazing. The nodchip trainer is all custom code, and trains extremely strong nets. There are all sorts of subtle tricks embedded in there as well that led to stronger nets. Not to mention, getting the quantization (float32 -> int16/int8) working gracefully is a huge challenge.

Suggest a related project