Testrun of Ethereal 12.62 finished - impressive progress!
https://www.sp-cc.de
(Perhaps you have to clear your browsercache or reload the website)
SPCC: Testrun of Ethereal 12.62 finished
Moderators: hgm, Rebel, chrisw
-
- Posts: 2439
- Joined: Sat Sep 03, 2011 7:25 am
- Location: Berlin, Germany
- Full name: Stefan Pohl
-
- Posts: 671
- Joined: Sun Jan 26, 2020 10:38 pm
- Location: Turkey
- Full name: Mehmet Karaman
Re: SPCC: Testrun of Ethereal 12.62 finished
Yes the development of Ethereal is impressive.
Maybe we will see an Ethereal NNUE engine in the near future.
Maybe we will see an Ethereal NNUE engine in the near future.
-
- Posts: 1756
- Joined: Tue Apr 19, 2016 6:08 am
- Location: U.S.A
- Full name: Andrew Grant
Re: SPCC: Testrun of Ethereal 12.62 finished
Nope. NNUE is a one way ticket to kill the originality of the engine.
I'll stick to hand picked terms, slightly adjusted by a minimal NN.
Thanks for the tests. +34 in your testsing; +26 in our regression testing.
Good to see that the self-play gains are inline with the general gains.
#WeAreAllDraude #JusticeForDraude #RememberDraude #LeptirBigUltra
"Those who can't do, clone instead" - Eduard ( A real life friend, not this forum's Eduard )
"Those who can't do, clone instead" - Eduard ( A real life friend, not this forum's Eduard )
-
- Posts: 240
- Joined: Sat Mar 18, 2006 4:01 am
- Location: Cold
Re: SPCC: Testrun of Ethereal 12.62 finished
And what if NNUE trained on Ethereal games?AndrewGrant wrote: ↑Mon Sep 28, 2020 8:29 am
Nope. NNUE is a one way ticket to kill the originality of the engine.
-
- Posts: 6995
- Joined: Thu Aug 18, 2011 12:04 pm
Re: SPCC: Testrun of Ethereal 12.62 finished
I am not so sure that's true.AndrewGrant wrote: ↑Mon Sep 28, 2020 8:29 amNope. NNUE is a one way ticket to kill the originality of the engine.
Here is a sim-test at depth=1 with the current NNUE engines.
http://rebel13.nl/dump/mysim.html
I expected 70-80% but.....
90% of coding is debugging, the other 10% is writing bugs.
-
- Posts: 1756
- Joined: Tue Apr 19, 2016 6:08 am
- Location: U.S.A
- Full name: Andrew Grant
Re: SPCC: Testrun of Ethereal 12.62 finished
This might be a statement about the variety of qsearch() functions.Rebel wrote: ↑Mon Sep 28, 2020 4:53 pmI am not so sure that's true.AndrewGrant wrote: ↑Mon Sep 28, 2020 8:29 amNope. NNUE is a one way ticket to kill the originality of the engine.
Here is a sim-test at depth=1 with the current NNUE engines.
http://rebel13.nl/dump/mysim.html
I expected 70-80% but.....
But interesting data. I expected much higher.
#WeAreAllDraude #JusticeForDraude #RememberDraude #LeptirBigUltra
"Those who can't do, clone instead" - Eduard ( A real life friend, not this forum's Eduard )
"Those who can't do, clone instead" - Eduard ( A real life friend, not this forum's Eduard )
-
- Posts: 550
- Joined: Tue Nov 19, 2019 8:48 pm
- Full name: Alayan Feh
Re: SPCC: Testrun of Ethereal 12.62 finished
High similarity is evidence, but low similarity is absence of evidence rather than evidence of absence.
If the nets are different, the case might be made that the nets really are this different from each other, but if my understanding is correct, Rubi and Minic use explicitely the same NNUE net (so identical position ordering by evaluation) yet get a very low similarity in the test, which would point at qsearch for the difference.
Ideally, for eval similarity testing you'd want (1) a static command allowing to get an engine's static eval for a position (2) a set of positions to evaluate, trimmed to only have objectively drawn positions (3) measure the similarity in position ordering.
If the nets are different, the case might be made that the nets really are this different from each other, but if my understanding is correct, Rubi and Minic use explicitely the same NNUE net (so identical position ordering by evaluation) yet get a very low similarity in the test, which would point at qsearch for the difference.
Ideally, for eval similarity testing you'd want (1) a static command allowing to get an engine's static eval for a position (2) a set of positions to evaluate, trimmed to only have objectively drawn positions (3) measure the similarity in position ordering.
-
- Posts: 1871
- Joined: Sat Nov 25, 2017 2:28 pm
- Location: France
Re: SPCC: Testrun of Ethereal 12.62 finished
Didn't t you show some week ago that minicnnue is 86% like sf while minic is only 36%? It was just for opening position maybe?Rebel wrote: ↑Mon Sep 28, 2020 4:53 pmI am not so sure that's true.AndrewGrant wrote: ↑Mon Sep 28, 2020 8:29 amNope. NNUE is a one way ticket to kill the originality of the engine.
Here is a sim-test at depth=1 with the current NNUE engines.
http://rebel13.nl/dump/mysim.html
I expected 70-80% but.....
-
- Posts: 216
- Joined: Sun Jan 22, 2017 8:30 pm
- Location: Russia
Re: SPCC: Testrun of Ethereal 12.62 finished
When people talk about NNUE, I'm not always sure if they mean sparse nets in general or the exact architecture used in SF (it looks like you meant the latter, which is how I use the NNUE term too). The reuse of the exact network architecture is an originality killer indeed. The exploration of the general ideas of sparsity and dynamic inference resulting in original architecture(s) is, which is what you've been doing, is, on the contrary, a way of making an engine truly unique.AndrewGrant wrote: ↑Mon Sep 28, 2020 8:29 am Nope. NNUE is a one way ticket to kill the originality of the engine.
Don't give revenge ideas to Norman, please
-
- Posts: 6995
- Joined: Thu Aug 18, 2011 12:04 pm
Re: SPCC: Testrun of Ethereal 12.62 finished
Yes, I remember. I will look into it, hoping I still have the data.xr_a_y wrote: ↑Mon Sep 28, 2020 8:32 pmDidn't t you show some week ago that minicnnue is 86% like sf while minic is only 36%? It was just for opening position maybe?Rebel wrote: ↑Mon Sep 28, 2020 4:53 pmI am not so sure that's true.AndrewGrant wrote: ↑Mon Sep 28, 2020 8:29 amNope. NNUE is a one way ticket to kill the originality of the engine.
Here is a sim-test at depth=1 with the current NNUE engines.
http://rebel13.nl/dump/mysim.html
I expected 70-80% but.....
90% of coding is debugging, the other 10% is writing bugs.