REBEL NUUE BLOG

As you possibly have already read the NNUE development of REBEL has come to a halt because my available hardware has become insufficient to make further progress in an acceptable time. My NNUE develop PC, has a Ryzen 12 cores, 24 threads plus a fast RTX 3080. That was good in the early days with Rebel 14 and 15 with small net sizes and few data but with REBEL-EAS using 28B data into a large net size the development time for a net increased from 2 days to 10-12 days and the fun slowly started to shrink.

After switching to Leela data I made quickly 20-25 elo progress but thereafter only 2 new ideas worked (see the good-data folder) and the fun reached a low point, spoiling 10-12 days too many times. That's no way of living. It's not something new, the stronger your engine becomes the harder it is to make further progress.

I still love NNUE programming, so much new to discover, so many new ideas to test, see the new-data folder. And so I setup a test environment of 4B data and small net size, test it with the new ideas. In this way testing a new idea usually takes only 2 days. And as my experience with NNUE programming has learned me, if new data works on 4B it will usually work adding the new data to the current best data of 26.5B as shown in the 2 improvements in the good-data folder.

So, there is this blog you can follow, usually updated once or twice a day with the latest results.

https://rebel7775.wixsite.com/rebel/kopie-van-adrl-blitz-2

Not much to see at the moment.

nice improvement wie epoch 80 (compared with epoch 40), about 26 Elo ?

I of course hope that Rebel EAS will not be the final version, the results of the previous DEV version looked good (+30). I also hope that there will be a change of mindset with Chris, at this strong level it is very difficult to make progress.

I also can't imagine that there aren't ways to improve the current search. There seem to be numerous opportunities for improvement in the new programs (of which I am only aware of a fraction), which may not individually bring about major improvements in rating, but should have an impact as a whole. Without these very small advances, there would be no measurable progress at Stockfish.

https://www-ikbenstil-nl.translate.goog/ikbenstil-workstation-hollandse-nieuwe-amd-threadripper-pro-5995wx-64-cores.html?_x_tr_sl=nl&_x_tr_tl=en&_x_tr_hl=nl&_x_tr_pto=wapp

This would be helpful

And with the latest RTX of course.

Estimated speed-up, factor 3.

Dio wrote:

nice improvement wie epoch 80 (compared with epoch 40), about 26 Elo ?

Early epochs learn fast.

Dio wrote:

I of course hope that Rebel EAS will not be the final version, the results of the previous DEV version looked good (+30). I also hope that there will be a change of mindset with Chris, at this strong level it is very difficult to make progress.

Temporarily moving to another game can bring new ideas.

Dio wrote:

I also can't imagine that there aren't ways to improve the current search. There seem to be numerous opportunities for improvement in the new programs (of which I am only aware of a fraction), which may not individually bring about major improvements in rating, but should have an impact as a whole. Without these very small advances, there would be no measurable progress at Stockfish.

True.

what's different in the cstal data from the rebel eas that makes them so much different

i also think your method while "strengthening" rebel/cstal, has also weakened its conversion skills. if you test rebel eas vs rebel 16 i'd be curious how it does in already winning position, as if you look at ccrl it has a very high drawrate like leela

texium wrote:

what's different in the cstal data from the rebel eas that makes them so much different

Chris and I talk a lot, exchange ideas and then we go different paths. Meaning, hard to tell the differences.

texium wrote:

i also think your method while "strengthening" rebel/cstal, has also weakened its conversion skills. if you test rebel eas vs rebel 16 i'd be curious how it does in already winning position, as if you look at ccrl it has a very high drawrate like leela

Draw rates are high in balanced positions, like in CCRL testing, go to Stefan's new list, or my anti-draw pilot for less draws.

Admin wrote:

texium wrote:

i also think your method while "strengthening" rebel/cstal, has also weakened its conversion skills. if you test rebel eas vs rebel 16 i'd be curious how it does in already winning position, as if you look at ccrl it has a very high drawrate like leela

Draw rates are high in balanced positions, like in CCRL testing, go to Stefan's new list, or my anti-draw pilot for less draws.

Perhaps I worded it wrong and misinterpreted data. I think cstal has an issue with generating advantages, be that it's aggressive style, or something else, but since ccrl tests balanced rather than winning openings, higher draw rate would be more indicative of not getting advantages for it to convert which is why it does better in adrl. I notic d when playing against weaker engines with normal openings it might struggle to get a win

No progress measured feeding the net with 10% high mobility situations, it only will make a net somewhat more aggressive but that I already have done with EAS. Now I want elo

texium wrote:

Admin wrote:

texium wrote:

i also think your method while "strengthening" rebel/cstal, has also weakened its conversion skills. if you test rebel eas vs rebel 16 i'd be curious how it does in already winning position, as if you look at ccrl it has a very high drawrate like leela

Draw rates are high in balanced positions, like in CCRL testing, go to Stefan's new list, or my anti-draw pilot for less draws.

Perhaps I worded it wrong and misinterpreted data. I think cstal has an issue with generating advantages, be that it's aggressive style, or something else, but since ccrl tests balanced rather than winning openings, higher draw rate would be more indicative of not getting advantages for it to convert which is why it does better in adrl. I notic d when playing against weaker engines  with normal openings it might struggle to get a win

It's hard to draw such a conclusion, current top engines hardly make mistakes, look at the number of SF losses, there are rare. Perhaps you could try the EAS version Chess System Tal.

First green (50.8%) in Rook-Imbalance.

This is great, very unique approach by improving strength and style through training data. Curious if it's possible for you to find positions that top engines struggle to evaluate due to complexity and train on that

texium wrote:

This is great, very unique approach by improving strength and style through training data. Curious if it's possible for you to find positions that top engines struggle to evaluate due to complexity and train on that

I don't even know the type of positions Rebel struggles with.

Finished Rook-Imbalance

Code:

   Rook Imbalnce
EPOCH  PERC  GAMES
 40   46.9%   2100
 60   47.2%   2100
 80   49.4%   2100
100   50.8%   2100
126   50.4%   2100
MAIN  50.7%   2000  

Testing Rook Imbalance, adding 400M positions  to the 4B MAIN data. Conclusion: nice but not convincing, but good enough to create 2B of those, merge them into the 26.5B best data, start the trainer and go fishing for 10-12 days.

The days of fast progress are over, gathering is what's left.

I got some more small improvements than expected and while I promised myself -- no more training sessions of 10-12 days -- it's going to happen again

Adding New Raw Data [2] in total 5.9B positions to the main base of 26.4B positions, making the data in total 32.3B positions and 1.25Tb of data. Then a full shuffle of the data which takes about one day.

Then start the trainer with some new setting, for insiders increasing the start LR from 0.000875 to 0.000975 and a gamma of 0.994.

So I noticed the rebel site is up, do you plan to update it

texium wrote:

So I noticed the rebel site is up, do you plan to update it

Active topics are -

https://rebel7775.wixsite.com/rebel/kopie-van-adrl-blitz-1
https://rebel7775.wixsite.com/rebel/kopie-van-adrl-blitz-2
https://rebel7775.wixsite.com/rebel/kopie-van-nnue-blog

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

First epoch (70) looks good, score is not only higher (49.0%) but it also needed less data 21% whereas Rebel-Dev used 27% of the data.

Nevertheless with only 2000 games the error bar (sometimes) is without mercy and scores have a tendency to swing. But because we play so many games the end result will be pretty reliable with and error margin of -5 | +5 elo maximum and likely less.

Admin wrote:

texium wrote:

So I noticed the rebel site is up, do you plan to update it

Active topics are -

https://rebel7775.wixsite.com/rebel/kopie-van-adrl-blitz-1
https://rebel7775.wixsite.com/rebel/kopie-van-adrl-blitz-2
https://rebel7775.wixsite.com/rebel/kopie-van-nnue-blog

i was referring to rebel13.nl still being up

rebel13.nl no longer exists. What you see is in the internet-archive, a search bot that scans the whole internet and copies every website.

Admin wrote:

rebel13.nl no longer exists. What you see is in the internet-archive, a search bot that scans the whole internet and copies every website.

im aware thats what was supposed to happen but the internet archive saves sites on the internet archive, not the official link of rebel



im not crazy right, it clearly appears to me like the site is running under rebel13.nl

You are not crazy it works with a so-called MTML redirect.

Code:

<meta http-equiv="Refresh" content="0; url='https://web.archive.org/web/20230622212023/http://www.rebel13.nl/'" />

Just try the link inside - https://web.archive.org/web/20230622212023/http://www.rebel13.nl

New NU version 1.4

[F12] - Tuning Piece Division,  this is how our 4B test data looks like when we make an overview from the 4B data with NU [F8]. As you see some are not well synchronized, specially the high odd piece counts. With NU [F12] we create extra data to fill the gaps in order that the data is better synchronized. Note the data before and after.



Operation
Edit the file balance-piece-count.txt and in millions enter the data that should be added for each of the low piece-counts. And good for ~10 elo, in this particular case.

Another use of NU [F12] is to set everything (for instance) on "1" in balance-piece-count.txt witch will create a new epd where every piece count has exactly 1.000.000 positions. In other words you can design any piece-count division data you wish, from midgame to endgame.

balance-piece-count.txt

Code:

32 : 1
31 : 1
30 : 1
29 : 1
28 : 1
27 : 1
26 : 1
25 : 1
24 : 1
23 : 1
22 : 1
21 : 1
20 : 1
19 : 1
18 : 1
17 : 1
16 : 1
15 : 1
14 : 1
13 : 1   
12 : 1
11 : 1
10 : 1
9  : 1
8  : 1
7  : 1
6  : 1
5  : 1
4  : 1
3  : 1

Some statistics of the data I am using.

Part-one - Pawnsless positions

Code:

Count Pawnless epd\l2-1.epd

Reading 1.305.000.000
 
Queen vs Queen     : 2.742.316
Queen vs two Rooks : 156.644
Queen vs Rook      : 329.140
Rook vs Rook       : 28.463.072
KBNK mates         : 401.139
Rook vs Minor      : 4.500.619
Pawnless Total     : 53.337.819  [4.09%]

Done...

This is 5% of my current (26.5B) best data and KR vs KR (28M) is laughable and in total 20x28=560M pointless positions that safely can (and will be deleted) and unnecessarily take space in the NN.

I wonder if databases of rating lists can beat my 4% of pawnless positions