LizardFish 0.1 NNUE now available
Moderators: hgm, Rebel, chrisw
-
- Posts: 1631
- Joined: Tue Aug 21, 2018 7:52 pm
- Full name: Dietrich Kappe
LizardFish 0.1 NNUE now available
After the big snafu from earlier, here LizardFish 0.1, trained on only 6 million games. Somewhere between the last Texel and Komodo 13 in strength. https://www.patreon.com/posts/39496596
Fat Titz by Stockfish, the engine with the bodaciously big net. Remember: size matters. If you want to learn more about this engine just google for "Fat Titz".
-
- Posts: 365
- Joined: Mon May 14, 2007 8:20 pm
- Full name: Boban Stanojević
Re: LizardFish 0.1 NNUE now available
Thanks, Dietrich. I hope this time it will be OK.
Meanwhile, to answer a question you asked on Discord, a lot of engines are stronger than Komodo (and SF) at depth 8: probably Hiarcs, Gandalf 7, even Pro Deo. A net trained on Benjamin depth 8 could be very intriguing.
Meanwhile, to answer a question you asked on Discord, a lot of engines are stronger than Komodo (and SF) at depth 8: probably Hiarcs, Gandalf 7, even Pro Deo. A net trained on Benjamin depth 8 could be very intriguing.
-
- Posts: 5960
- Joined: Sun Jan 10, 2010 6:15 am
- Location: Maryland USA
Re: LizardFish 0.1 NNUE now available
Probably this is just because Komodo reduces much more heavily than most other engines (excluding Stockfish). If you just set Use LMR in Komodo to false, it should improve a huge amount at depth 8. But if you want stronger play for the same amount of time, just using depth 9 instead is probably better. I don't know whether more depth or turning off LMR would be better for training a net.matejst wrote: ↑Mon Jul 20, 2020 1:31 am Thanks, Dietrich. I hope this time it will be OK.
Meanwhile, to answer a question you asked on Discord, a lot of engines are stronger than Komodo (and SF) at depth 8: probably Hiarcs, Gandalf 7, even Pro Deo. A net trained on Benjamin depth 8 could be very intriguing.
Komodo rules!
-
- Posts: 6995
- Joined: Thu Aug 18, 2011 12:04 pm
Re: LizardFish 0.1 NNUE now available
Very interesting, the Lc0 training method is quite different than the SF and Komodo *.bin approach, see the link.dkappe wrote: ↑Mon Jul 20, 2020 12:11 am After the big snafu from earlier, here LizardFish 0.1, trained on only 6 million games. Somewhere between the last Texel and Komodo 13 in strength. https://www.patreon.com/posts/39496596
http://rebel13.nl/html/nn.html
90% of coding is debugging, the other 10% is writing bugs.
-
- Posts: 6995
- Joined: Thu Aug 18, 2011 12:04 pm
Re: LizardFish 0.1 NNUE now available
yes, Yes, YES.......matejst wrote: ↑Mon Jul 20, 2020 1:31 am Thanks, Dietrich. I hope this time it will be OK.
Meanwhile, to answer a question you asked on Discord, a lot of engines are stronger than Komodo (and SF) at depth 8: probably Hiarcs, Gandalf 7, even Pro Deo. A net trained on Benjamin depth 8 could be very intriguing.
90% of coding is debugging, the other 10% is writing bugs.
-
- Posts: 365
- Joined: Mon May 14, 2007 8:20 pm
- Full name: Boban Stanojević
Re: LizardFish 0.1 NNUE now available
Larry, Ed Shroeder made multiple tests with different depths (you can find them on his site) and, yes, it is the most probable reason. But it seems also that the evaluation of these engines is better adapted to lower depths.lkaufman wrote: ↑Mon Jul 20, 2020 3:37 amProbably this is just because Komodo reduces much more heavily than most other engines (excluding Stockfish). If you just set Use LMR in Komodo to false, it should improve a huge amount at depth 8. But if you want stronger play for the same amount of time, just using depth 9 instead is probably better. I don't know whether more depth or turning off LMR would be better for training a net.matejst wrote: ↑Mon Jul 20, 2020 1:31 am Thanks, Dietrich. I hope this time it will be OK.
Meanwhile, to answer a question you asked on Discord, a lot of engines are stronger than Komodo (and SF) at depth 8: probably Hiarcs, Gandalf 7, even Pro Deo. A net trained on Benjamin depth 8 could be very intriguing.
-
- Posts: 365
- Joined: Mon May 14, 2007 8:20 pm
- Full name: Boban Stanojević
Re: LizardFish 0.1 NNUE now available
So far, although far from conclusive, it seems closer to the upper limit. It was better than RubiChess 1.7.3 and Komodo 11 in short matches at 2'2".dkappe wrote: ↑Mon Jul 20, 2020 12:11 am After the big snafu from earlier, here LizardFish 0.1, trained on only 6 million games. Somewhere between the last Texel and Komodo 13 in strength. https://www.patreon.com/posts/39496596
-
- Posts: 1631
- Joined: Tue Aug 21, 2018 7:52 pm
- Full name: Dietrich Kappe
Re: LizardFish 0.1 NNUE now available
You have a Benjamin linux executable?Rebel wrote: ↑Mon Jul 20, 2020 9:37 amyes, Yes, YES.......matejst wrote: ↑Mon Jul 20, 2020 1:31 am Thanks, Dietrich. I hope this time it will be OK.
Meanwhile, to answer a question you asked on Discord, a lot of engines are stronger than Komodo (and SF) at depth 8: probably Hiarcs, Gandalf 7, even Pro Deo. A net trained on Benjamin depth 8 could be very intriguing.
Fat Titz by Stockfish, the engine with the bodaciously big net. Remember: size matters. If you want to learn more about this engine just google for "Fat Titz".
-
- Posts: 6995
- Joined: Thu Aug 18, 2011 12:04 pm
Re: LizardFish 0.1 NNUE now available
Unfortunately not.dkappe wrote: ↑Mon Jul 20, 2020 3:32 pmYou have a Benjamin linux executable?Rebel wrote: ↑Mon Jul 20, 2020 9:37 amyes, Yes, YES.......matejst wrote: ↑Mon Jul 20, 2020 1:31 am Thanks, Dietrich. I hope this time it will be OK.
Meanwhile, to answer a question you asked on Discord, a lot of engines are stronger than Komodo (and SF) at depth 8: probably Hiarcs, Gandalf 7, even Pro Deo. A net trained on Benjamin depth 8 could be very intriguing.
90% of coding is debugging, the other 10% is writing bugs.
-
- Posts: 1631
- Joined: Tue Aug 21, 2018 7:52 pm
- Full name: Dietrich Kappe
Re: LizardFish 0.1 NNUE now available
I might be able to run a Windows one in WINE. Right now I’m generating some data for Toga II 4.01. It’s rattling and wheezing it’s way to d8 with lots of extra nodes.
Fat Titz by Stockfish, the engine with the bodaciously big net. Remember: size matters. If you want to learn more about this engine just google for "Fat Titz".