I'm pleased to announce a new release of Myrddin, version 0.91. This is the first version to use an "NNUE-like" eval, with all credit and a great amount of thanks to David Carteau and his Cerebrum library. Only a few small adjustments needed to be made to make his inference code compatible with Myrddin's design.
I expect this version to be around 2900 CCRL, an improvement of about 250-275 elo. My code and release repository is here: https://github.com/JVMerlino/Myrddin
Again, many thanks to David, and to the CC community at large, which I have enjoyed being a part of for 30 years.
JVMerlino wrote: ↑Sun Oct 20, 2024 5:52 pm
I'm pleased to announce a new release of Myrddin, version 0.91. This is the first version to use an "NNUE-like" eval, with all credit and a great amount of thanks to David Carteau and his Cerebrum library. Only a few small adjustments needed to be made to make his inference code compatible with Myrddin's design.
I expect this version to be around 2900 CCRL, an improvement of about 250-275 elo. My code and release repository is here: https://github.com/JVMerlino/Myrddin
Again, many thanks to David, and to the CC community at large, which I have enjoyed being a part of for 30 years.
jm
Hi John,
Congratulations ! After adapting Orion to your piece scheme, I did a quick test (400 games at 40/1) and your NN performed (slightly) better than the one embedded in Orion v1.0 (50.7%-49.3%) : well done !!
Kind regards,
David
Download the Orion chess engine --- Train your NNUE with the Cerebrum library --- Contribute to the Nostradamus experiment !
JVMerlino wrote: ↑Sun Oct 20, 2024 5:52 pm
I'm pleased to announce a new release of Myrddin, version 0.91. This is the first version to use an "NNUE-like" eval, with all credit and a great amount of thanks to David Carteau and his Cerebrum library. Only a few small adjustments needed to be made to make his inference code compatible with Myrddin's design.
I expect this version to be around 2900 CCRL, an improvement of about 250-275 elo. My code and release repository is here: https://github.com/JVMerlino/Myrddin
Again, many thanks to David, and to the CC community at large, which I have enjoyed being a part of for 30 years.
jm
Hi John,
Congratulations ! After adapting Orion to your piece scheme, I did a quick test (400 games at 40/1) and your NN performed (slightly) better than the one embedded in Orion v1.0 (50.7%-49.3%) : well done !!
Kind regards,
David
Ah, interesting. It's a mix of (in percentage order), high-end Lichess games, CCRL, Myrddin gauntlet, and Myrddin self-play. As you suggested, adding games that ended in checkmate (via Lichess) definitely improved play, as CCRL games are often adjudicated long before checkmate.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
