Tucano 10.00 release - Neural network evaluation

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Rebel, chrisw

sedicla
Posts: 178
Joined: Sat Jan 08, 2011 12:51 am
Location: USA
Full name: Alcides Schulz

Tucano 10.00 release - Neural network evaluation

Post by sedicla »

Hi,

Tucano 10 NN has been released. It has a couple of changes but the main change is the introduction of neural network evaluation.
I used tucano hand crafted evaluation to generate about 1.5B positions, and used the nodchip trainer to generate the nn weights.
For NN probing I'm using Daniel Shawn NNUE probe library.

The improvement is about 200 elo over version 9.00. It is recommended to use AVX2 whenever possible. I also provided the "_OLD" version that does not make use of "intrinsics", meaning slower code.

It is available for download from my github page, but you can download the release 10.00 which contains binaries for different architectures.
--> https://github.com/alcides-schulz/Tucan ... /tag/10.00

Note that are binaries for windows and Linux (Centos). The network weights file is tucano_nn01.bin. Tucano will try to load this file by default if it is the same folder as the binary file, otherwise you can specify the UCI parameter EvalFile. More information in the README file.

Please let me know any issues.

For future plans, I'm planning to continue the work on NN, Maybe can create a probing code or create my own trainer. I'm already doing some research on this.

Alcides.
bastiball
Posts: 5254
Joined: Tue Oct 20, 2020 4:18 am
Full name: Basti Dangca

Re: Tucano 10.00 release - Neural network evaluation

Post by bastiball »

Congrats for the release! Can't wait to tests it :D
Basti Dangca
CCRL testing group
Frank Quisinsky
Posts: 6808
Joined: Wed Nov 18, 2009 7:16 pm
Location: Gutweiler, Germany
Full name: Frank Quisinsky

Re: Tucano 10.00 release - Neural network evaluation

Post by Frank Quisinsky »

Hi Alcides

good news, the next strong NN engine is available!
Congrats!

Best
Frank