chess4j 5.0

[jswaff]

Hello all,

I'm happy to release version 5.0 of my Java chess program chess4j. It can be downloaded from the chess4j github page: https://github.com/jswaff/chess4j . It should be around 150 ELO stronger than 4.0. (I don't really test the strength of chess4j directly, but it is algorithmically identical to Prophet.)

The zip and tar.gz release artifacts contain two java jar files each. chess4j-5.0.jar is the "plain" jar file. chess4j-5.0-linux.jar is platform dependent. It contains an embedded static library of Prophet code for faster execution on Linux machines.

The most notable thing about this release is the addition of an automated tuner. The initial implementation was "Texel style." While that did work well, it was also very slow. That inspired me to dig deeper and learn more about logistic regression and gradient descent. I actually stopped development long enough to take an online class- the popular "Machine Learning" course by Andrew Ng. Now it's a fairly quick process to add new eval terms and recalibrate. For training data, I used the standard Zuri chess set as well as the ones published by Andrew Grant. (The ones by Andrew worked better for me, so thank you Andrew.) Tuner implementation: https://github.com/jswaff/chess4j/tree/ ... ss4j/tuner

I plan to continue to improve the HCE for a while, particularly with pawns and endgames, before diving into neural networks.

--
James