Posts : 2608 Join date : 2020-11-17 Location : Netherlands
Subject: Rebel 15 development Mon Apr 18, 2022 12:20 am
Code:
No. Name Win Draw Loss Unf. Score Games % ------------------------------------------------------------ 1 Rebel-15-nn8 +428 =581 -152 *0 718.5 1161 61.9% 2 Rebel-14.2-avx2 +152 =581 -428 *0 442.5 1161 38.1%
This is the very first result of what I call an early alpha version of Rebel 15, currently +83 elo over 14.2
Next step is to tune the neural net finding the optimal value of the "NNUE Tuner" option.
To speed-up this process I can use help if you are willing to donate your computer time. Minimum requirements are 1) at least having 16 threads available else it makes little sense, 2) AVX2 compatibility, 3) Because of the GPL and the restrictions that come with it you automatically become a member of my test team because I don't want to compose the source code with every new test version I will be sending you, 4) in case you want your name will be mentioned as tester after release on the web page and last but not least, 5) you do exactly what I say
Toga 1.1 is about 50 Elos better than Rebel 14.2, per Frank Quisinsky's KI rating list, after 200 games at 40 mvs in 20 mn. It seems there should be more room for improvements.
Unfortunately, I don't fulfill any of your hardware conditions at the present.
Mclane
Posts : 3022 Join date : 2020-11-17 Age : 57 Location : United States of Europe, Germany, Ruhr area
No. Name Win Draw Loss Unf. Score Games % ------------------------------------------------------------ 1 Rebel-15-nn8 +428 =581 -152 *0 718.5 1161 61.9% 2 Rebel-14.2-avx2 +152 =581 -428 *0 442.5 1161 38.1%
This is the very first result of what I call an early alpha version of Rebel 15, currently +83 elo over 14.2
Next step is to tune the neural net finding the optimal value of the "NNUE Tuner" option.
To speed-up this process I can use help if you are willing to donate your computer time. Minimum requirements are 1) at least having 16 threads available else it makes little sense, 2) AVX2 compatibility, 3) Because of the GPL and the restrictions that come with it you automatically become a member of my test team because I don't want to compose the source code with every new test version I will be sending you, 4) in case you want your name will be mentioned as tester after release on the web page and last but not least, 5) you do exactly what I say
1.I do have a ryzen 16 core / 32 threads with nvidia 3060, 2.And another slower ryzen 6 core /12 threads with a radeon gpu.
They should be avx2.
3. Further i think i have a laptop i7 with 4 cores and nvidia 16 or 1050. I am not sure which chipset this is.
4. And also and old 6 core xeon with graphic card.
I guess i have some more pc‘s.
5. There is at least another ryzen 4 core with internal gpu the speed of the 6 core xeon.
They all do nothing in the moment.
Admin likes this post
Admin Admin
Posts : 2608 Join date : 2020-11-17 Location : Netherlands
Posts : 2608 Join date : 2020-11-17 Location : Netherlands
Subject: Re: Rebel 15 development Sun Apr 24, 2022 10:36 am
Thorsten has been so kind to test the tuner-value of the new net architecture for Rebel 15. It seems that every new and better [!] net architecture created by Chris needs its own tuning, at least that is true for Rebel. Rebel 14 and 14.1 used 100%, with 14.2 the best value was 40%. So here are the results, first engine-engine, the current Rebel-15-nn8 vs Toga 4.11
Code:
Thorsten testing epoch 202 vs Toga 4.11 tc=40/20 Error Bar 5000 games +7 | -7
TUNER GAMES SCORE ELO 90 5000 55.5% -2 80 5000 56.5% +4 70 5000 55.9% 0 BASE 60 5000 56.9% +7 50 5000 55.3% -4 40 5000 52.6% -22
As it seems for the moment 60 is best, Thorsten played 5000 games against a CCRL 3300+ elo pool, Toga 4.11 is rated 3307 on the 40/2 list.
This will be followed by 2 matches with the values 70 and 80 against the same opponents. The in total 9 matches will be a good indication what the final value of the tuner will be for Rebel 15.
A big thank you to Thorsten, in the meantime I can be working on an even better net.
BTW, there is an interesting observation about the results, do you see it?
adminx, matejst and Damir Desevac like this post
Admin Admin
Posts : 2608 Join date : 2020-11-17 Location : Netherlands
Subject: Re: Rebel 15 development Sun Apr 24, 2022 11:20 pm
Posts : 2608 Join date : 2020-11-17 Location : Netherlands
Subject: Re: Rebel 15 development Mon Apr 25, 2022 9:11 am
BTW, there is an interesting observation about the results, do you see it?
I will answermy own question
What you see normally when programmers give an estimated elo gain based on self-play (say 100 elo) it will end up in most cases (not always) lower on the rating lists. Programmers know it and most of them will be happy with a gain of 70-80 elo.
Example, the Rebel 14.2 website states : Self playing against 14.1 shows an elo increase of ~80-85. And on the rating lists it was not more than 40-45 elo (or so).
Now look at the current data, self-play indicates 55-56% which is an elo increase of (only) 35-40 elo, the match results against other engines give 63-64%, an elo increase of 90-100. It's the world upside down, never seen this before, and I am doing this kind of stuff for quite some time, computer chess never disappoints.
Mclane, TheSelfImprover and Eelco like this post
Admin Admin
Posts : 2608 Join date : 2020-11-17 Location : Netherlands
Subject: Re: Rebel 15 development Tue Apr 26, 2022 9:34 am
New results from Thorsten are in. Topic, testing the scaling. Moving from 40/20 to 40/120
The results (unexpectedly) are so good I decided to release this version as Rebel 15. Meanwhile I can proceed the version I have in mind in a more relaxed way. Thanks to Thorsten for his testing and Chris of course for the creation of a new better NNUE architecture.
Mclane, TheSelfImprover, adminx, Dio and Eelco like this post
Home 
