REBEL NUUE BLOG

Subject: Re: REBEL NUUE BLOG Mon Dec 04, 2023 5:45 pm

Part-two - Rook-Rook endings

Code:: Extract Rook Endings

Reading epd\l2-1.epd : 1.305.000.000

Rooks : 136.238.409 [10.44%]
KRPKR : 8.781.599

So 10.4% of my data (2.7B) is about rook-rook-with-pawns endings and the draw rate is : 71.2%

And 175M is about the KRPKR ending with a draw rate of : 95.6%

Time to trim here also.

Subject: Re: REBEL NUUE BLOG Fri Dec 22, 2023 11:15 am

https://rebel7775.wixsite.com/rebel/kopie-van-nnue-blog

Well, well, well, it seems I have been playing with a bugged version the last ~3 months, I am getting old confused

OTOH, the sun is shining again and it explains why so little progress has been made.

Subject: Re: REBEL NUUE BLOG Sat Dec 23, 2023 3:50 pm

One of my kids posted this on Facebook, married for 49 years.

Posts : 112 Join date : 2022-07-19

any updates?

Subject: Re: REBEL NUUE BLOG Wed Feb 21, 2024 9:56 pm

texium wrote:: any updates?

Not much progress, the net is pretty much optimal, only ~30 elo to REBEL-EAS.

I am giving myself another month to test a few ideas and then I am done.

Subject: Re: REBEL NUUE BLOG Wed Feb 21, 2024 11:19 pm

I activated the NNUE training page again.

Test-9 : We are trying a new shuffle method. Shuffling data is important. We once tried to train a net without shuffling the data and lost ~25 elo points.

A second test on the blog page the new shuffle method is tested with a small net, see the end of the page.

Code:: MAIN | New Shuffle
EPOCH PERC GAMES | EPOCH PERC GAMES
40 46.0% 2100 | 40 46.6% 2100
60 48.6% 2100 | 60 47.6% 2100
80 48.2% 2100 | 80 49.0% 2100
100 48.5% 2100 | 100 % 2100
120 49.7% 2100 | 120 % 2100
MAIN 2000

So far (emphasis added) so good.

Subject: Re: REBEL NUUE BLOG Tue Feb 27, 2024 11:41 am

https://rebel7775.wixsite.com/rebel/kopie-van-nnue-blog

Doing some final tests.

Subject: Re: REBEL NUUE BLOG Fri Mar 01, 2024 8:28 am

Final results -

Code:: Final tests - Master version + a BETA version
. UHO openings
. versus 10 opponents
. CEGT elo pool : 3463
. CEGT elo rating REBEL EAS : 3465

Code:: AMD Ryzen 5 3600

Version Games TC Score Elo Gain
Master 2000 40/10 55.5% 3502 +37
Master 2000 40/20 54.6% 3495 +30

BETA 2000 40/10 55.7% 3503 +38
BETA 2000 40/20 55.8% 3504 +39

Code:: AMD Ryzen 9 5900X

Version Games TC Score Elo Gain
Master 2000 40/10 54.7% 3496 +31
Master 2000 40/20 55.1% 3499 +34
Master 2000 40/40 54.4% 3495 +30
Master 1000 40/80 56.5% 3508 +43

BETA 2000 40/10 54.0% 3491 +26
BETA 2000 40/20 54.5% 3495 +30
BETA 2000 40/40 55.9% 3505 +40
BETA 1000 40/80 56.6% 3509 +44

Not bad at all but not the desired +50 elo.

Posts : 214 Join date : 2021-08-28

+40 Elo is a lot at this level (especially since they are no self-play results that must be evaluated differently). I would be very happy about a new release from Rebel.

Subject: Re: REBEL NUUE BLOG Fri Mar 01, 2024 11:25 pm

Dio wrote:: +40 Elo is a lot at this level (especially since they are no self-play results that must be evaluated differently). I would be very happy about a new release from Rebel.

We are discussing this at the moment.

Posts : 41 Join date : 2022-06-08

You could also check rating lists.. like in my list with +40Elo it would gain 8 places !!
Is a other way to look at it in place to get each time +50Elo Wink

» Rebel in my FACE -Tournaments
» Rebel 14
» REBEL-EAS
» Rebel EAS and CS Tal Bug
» Rebel, CS Tal