Rebel 14.2

1. New and stronger neural network made from 1.6 billion position analyzed with Rebel 14.1, estimated elo gain 50-60.

2. Search changes, estimated elo gain 15-20.

3. NNUE tuner parameter in Engine Settings. Default setting now 40%.

Self playing against 14.1 shows an elo increase of ~80-85.

On the Gambit Rating List Rebel 14.2 is rated 3278 at place 15.

http://rebel13.nl/windows/rebel-14.html
http://rebel13.nl/grl-best-40-2.html

Just in case you downloaded Rebel 14.2 before 10:00 Amsterdam time on March 2, 2022, re-download for a faster compile.

Ed,

Will there be a Toga IV with the new net?

matejst wrote:

Ed,

Will there be a Toga IV with the new net?

Yes and with a good look at the Toga IV search.

Thanks, Ed.

Once again, if you invest a lot of work in the Toga IV search, developing two similar searches could be redundant.

I have started the first tests with Rebel 14.2 for CEGT's 40/4 list. The tests continue, of course. Please ignore unqualified comments from others regarding Toga IV 1.1 on talkchess.  https://cegt.forumieren.com/t1696-testing-rebel-14-2nn#3233

Nice, thank you for the service.

Dio wrote:

I have started the first tests with Rebel 14.2 for CEGT's 40/4 list. The tests continue, of course. Please ignore unqualified comments from others regarding Toga IV 1.1 on talkchess.  https://cegt.forumieren.com/t1696-testing-rebel-14-2nn#3233

I can imagine who you have in mind Jörg.

Some words of the inner works of the engine valid for Rebel and Toga.

1. The NNUE Tuner parameter. In previous versions its value was 100%, in the current versions the value is set to 40%. Why is that? Because the R14.2 net was no longer created from Benjamin but from R14.1 and those who had a good look at the games will have noticed the move scores it produced are much higher than Benjamin, up to a factor of 2 average, sometimes peeking with a factor of 3, or even 3.5. Meaning, the NNUE learner seems to exaggerate, likely because I did start from Benjamin and not from the more solid ProDeo evaluation. It's more an educated wild guess than a fact. Nobody exactly knows what's goes on in the mind of the learner, kind of quantum physics headache stuff.

Anyway, the default value 40 is tuned for elo. If you increase it to 45 you will lose ~5 elo but playing style will improve, similar story for 50, a loss of 10-15 elo but more active play. Set it to 100 and it will lose about every game.

2. I have the feeling some of the engine options are not fully understood. Because of NNUE -- which replaces the whole HCE evaluation -- the evaluation options are no longer active, you can change them as much as you want but nothing will happen. All of it from Material to Toga Rook Pawn Endgame Penalty is no longer active. However I like to keep them for the moment.

3. You can get a small improvement (5-10 elo) by setting Delta Margin from 50 to 100 and maybe 150 or 200 will be even better. Found that out yesterday.

-------------

Just read the Russians have been attacking a nuclear power, I don't mind any longer if NATO interferes.

The first result for Rebel 14.2 in the CEGT 40/4 list after 2500 games, where available, I have attached the results for Rebel 14.1 for comparison. https://cegt.forumieren.com/t1696-testing-rebel-14-2nn

Dio wrote:

The first result for Rebel 14.2 in the CEGT 40/4 list after 2500 games, where available, I have attached the results for Rebel 14.1 for comparison. https://cegt.forumieren.com/t1696-testing-rebel-14-2nn

Thanks for testing and your updates. I am satisfied with the result considering 14.1 scored a lot above my own expectation.

+46 ?!

Mclane wrote:

+46 ?!

I don't know exacly what you meant with the question Thorsten but if you combine it with the gain of Rebel 14.1 over Fruit, as Ed said, I think it is an enormous amount. Just for NNUE and Rebel's or Benjamin's eval. (I did not quite follow the NNUE thing for a while so I don't know exactly what the change was for Stockfish. But not this much I am sure).

I don't know if Pavel changed much in the evaluation of Fruit, but it wasn't bad before. But it is blown away by Rebel's NNUE evaluation. With classic eval you could not achieve it no matter how many improvements You have to factor in the material imbalances, the sudden jumps in eval if a piece goes missing, the gradual change you have to incorporate in classical eval if you go from midgame to endgame, all that the NNUE can take care of better, I think? Just a guess.

If the results against 14.1 alone were better, I think because you have many iterations of testing against the previous version alone. But any 'hole' remaining, both from Fruit's search and Rebel's original eval, will, apparent from the results against other engines, not go away with tuning alone. If the 'holes' are still in both versions playing against each other, they are invisible. They come back when you test against other engines. So that testing is also very valuable by all the testing groups. But the measured increase will always be less than the self testing.

(I can understand though the testers are a bit overwhelmed with all the new versions. And a bit confused. You already know the new version is better, just not by how much. Why not ask them to just test if there is a major new version, like a Rebel 15 or something?)

It's normal that there are many versions. Usually a programmer produces 1 version a day, sometimes more then one version a day. It depends how many of them you give to the testing team. Usually the inner circle of friends around an engine are used to these many versions...

Every programmer works different. I am used to work systematically. One project at the time. When it's finished and the elo is good enough release. Backup the version and start with ideas that come next. Other programmers (remember for instance Shredder 13 and recently Hiarcs 15) prefer to wait till they have 200-2050. It has of course its own charm.