Discussion of anything and everything relating to chess playing software and machines.
Moderators: hgm , Rebel , chrisw
Gabor Szots
Posts: 1365Joined: Sat Jul 21, 2018 7:43 amLocation: Szentendre, HungaryFull name: Gabor Szots
Post by Gabor Szots Fri Aug 06, 2021 6:19 pm
I am a bit disappointed to see that the net you use is an old SF net: nn-a9f9e868162a.nnue,
https://tests.stockfishchess.org/nns?page=4 Gabor Szots
kranium
Posts: 2129Joined: Thu May 29, 2008 10:43 am
Post by kranium Fri Aug 06, 2021 6:56 pm
Congratulations!
I see it's been several years since your last release...and 0.7 with many many improvements!
https://www.oscargavira.es/?sec=Coiled_Novedades
must have been a lot of work...and open source as well!
best of luck!
Bmark
Posts: 28Joined: Wed Aug 14, 2013 5:24 amLocation: Spain
Post by Bmark Fri Aug 06, 2021 8:27 pm
Hi Gabor Szots
Coiled-NNUE is ready to load other neural networks. Less modern 40MB. It is limited by nnue-probe.
Bmark
Posts: 28Joined: Wed Aug 14, 2013 5:24 amLocation: Spain
Post by Bmark Fri Aug 06, 2021 8:34 pm
Hi kranium
Three years.
I want everything to be perfect and it is difficult for me to launch new versions.
Yes, there are improvements
connor_mcmonigle
Posts: 541Joined: Sun Sep 06, 2020 4:40 amFull name: Connor McMonigle
Post by connor_mcmonigle Fri Aug 06, 2021 8:38 pm
Bmark wrote: ↑ Fri Aug 06, 2021 8:27 pm
Hi Gabor Szots
Coiled-NNUE is ready to load other neural networks. Less modern 40MB. It is limited by nnue-probe.
Most, maybe all, CCRL testers seem uninterested in testing engines without original weights produced via training on data originating from the given engine's own unique search/evaluation function. In using a Stockfish network with CFish derived probing code, your evaluation function ends up functionally identical to SF 12's evaluation function. In my understanding, this is far too much similarity in the eyes of most testers.
Bmark
Posts: 28Joined: Wed Aug 14, 2013 5:24 amLocation: Spain
Post by Bmark Fri Aug 06, 2021 9:16 pm
connor_mcmonigle wrote: ↑ Fri Aug 06, 2021 8:38 pm Bmark wrote: ↑ Fri Aug 06, 2021 8:27 pm
Hi Gabor Szots
Coiled-NNUE is ready to load other neural networks. Less modern 40MB. It is limited by nnue-probe.
Most, maybe all, CCRL testers seem uninterested in testing engines without original weights produced via training on data originating from the given engine's own unique search/evaluation function. In using a Stockfish network with CFish derived probing code, your evaluation function ends up functionally identical to SF 12's evaluation function. In my understanding, this is far too much similarity in the eyes of most testers.
When using a neural network, the evaluation is simulate and lower.
What's new are: Support Chess960. Neural Network (NNUE). Opening book chess960. Open Source (Github).
Loss: Internal evaluation.
Now many engines will appear using nnue, as it makes them stronger.
Chessqueen
Posts: 5595Joined: Wed Sep 05, 2018 2:16 amLocation: MovingFull name: Jorge Picado
Post by Chessqueen Fri Aug 06, 2021 9:35 pm
Bmark wrote: ↑ Fri Aug 06, 2021 9:16 pm connor_mcmonigle wrote: ↑ Fri Aug 06, 2021 8:38 pm Bmark wrote: ↑ Fri Aug 06, 2021 8:27 pm
Hi Gabor Szots
Coiled-NNUE is ready to load other neural networks. Less modern 40MB. It is limited by nnue-probe.
Most, maybe all, CCRL testers seem uninterested in testing engines without original weights produced via training on data originating from the given engine's own unique search/evaluation function. In using a Stockfish network with CFish derived probing code, your evaluation function ends up functionally identical to SF 12's evaluation function. In my understanding, this is far too much similarity in the eyes of most testers.
When using a neural network, the evaluation is simulate and lower.
What's new are: Support Chess960. Neural Network (NNUE). Opening book chess960. Open Source (Github).
Loss: Internal evaluation.
Now many engines will appear using nnue, as it makes them stronger.
What is the approximate Elo of Coiled = Enroscado https://www.alamy.com/search.html?pn=1& ... ilters%3D0
Bmark
Posts: 28Joined: Wed Aug 14, 2013 5:24 amLocation: Spain
Post by Bmark Fri Aug 06, 2021 9:45 pm
Chessqueen wrote: ↑ Fri Aug 06, 2021 9:35 pm Bmark wrote: ↑ Fri Aug 06, 2021 9:16 pm connor_mcmonigle wrote: ↑ Fri Aug 06, 2021 8:38 pm Bmark wrote: ↑ Fri Aug 06, 2021 8:27 pm
Hi Gabor Szots
Coiled-NNUE is ready to load other neural networks. Less modern 40MB. It is limited by nnue-probe.
Most, maybe all, CCRL testers seem uninterested in testing engines without original weights produced via training on data originating from the given engine's own unique search/evaluation function. In using a Stockfish network with CFish derived probing code, your evaluation function ends up functionally identical to SF 12's evaluation function. In my understanding, this is far too much similarity in the eyes of most testers.
When using a neural network, the evaluation is simulate and lower.
What's new are: Support Chess960. Neural Network (NNUE). Opening book chess960. Open Source (Github).
Loss: Internal evaluation.
Now many engines will appear using nnue, as it makes them stronger.
What is the approximate Elo of Coiled = Enroscado https://www.alamy.com/search.html?pn=1& ... ilters%3D0
Estimated ELO: ± 3050.
Sylwy
Posts: 4473Joined: Fri Apr 21, 2006 4:19 pmLocation: IASI - the historical capital of MOLDOVAFull name: SilvianR
Post by Sylwy Sat Aug 07, 2021 8:36 pm
[pgn]
https://www.alamy.com/search.html?pn=1& ... ilters%3D0