TOGA IV with Rebel 14.2 NNUE evaluation

• TOGA IV now with Rebel 14.2 NNUE evaluation.
• Toga IV 1.1 is about 45 elo stronger than Rebel 14.2 and above that offers you:
• Multiple threading up to 64 threads;
• MultiPV up to 10 moves.

• TOGA IV 1.1 earned a rating of 3323 on the Gambit Rating List.

• Offered to you in cooperation with Chris Whittington.

http://rebel13.nl/windows/rebel-14.html#other

Admin wrote:

• TOGA IV now with Rebel 14.2 NNUE evaluation.
• Toga IV 1.1 is about 45 elo stronger than Rebel 14.2 and above that offers you:
• Multiple threading up to 64 threads;
• MultiPV up to 10 moves.

• TOGA IV 1.1 earned a rating of 3323 on the Gambit Rating List.

• Offered to you in cooperation with Chris Whittington.

http://rebel13.nl/windows/rebel-14.html#other

Well done!

Ed,

You should definitively switch to Toga as the engine for Rebel. It would be much simpler to test and improve just one search than two. Finally, there is no need to call it "Toga" if credits is given where it is due, but just continue with "Rebel 14.3, or 4" etc.

I understand that you want to make the origins transparent and try to be as honest as possible, but there is nothing wrong about using the Toga source and continuing its development under another name. Nobody complains about StockFish today and nobody calls it "Glaurung".

Go the mp way ed.

matejst wrote:

Ed,

You should definitively switch to Toga as the engine for Rebel. It would be much simpler to test and improve just one search than two. Finally, there is no need to call it "Toga" if credits is given where it is due, but just continue with "Rebel 14.3, or 4" etc.

I understand that you want to make the origins transparent and try to be as honest as possible, but there is nothing wrong about using the Toga source and continuing its development under another name. Nobody complains about StockFish today and nobody calls it "Glaurung".

Mclane wrote:

Go the mp way ed.

It's a good argument certainly now that I fixed the Toga search and added more search improvements. It also avoids name confusion as seen at Talkchess.

Toga's king safety is very bad. If this can be improved would be great.

Damir,

I think it has already improved compared to Toga 14.1. But, indeed, there seems to be lot of room to adapt the search to the evaluation. Then, I just discovered that multiPV has a big bug. In this known position of the Sicilian,

[pgn]

[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "?"]
[Black "?"]
[Result "*"]
[PlyCount "20"]

1. e4 c5 2. Nf3 Nc6 3. d4 cxd4 4. Nxd4 e6 5. Nc3 Nf6 6. Ndb5 Bc5 7. Bf4 O-O 8.
Bc7 Qe7 9. Bd6 Bxd6 10. Qxd6 Qxd6 *

[/pgn]

the second line it gives is a double blunder (11.Nd5, Ne4 etc.).

Anyway, it is a project that matures steadily, and with more insights in the way NNs behave I trust Ed to improve tremendously Rebel.

Damir Desevac wrote:

Toga's king safety is very bad. If this can be improved would be great.

Example?

For my part, after a few hours of analyzing and playing with Toga Rebel, I saw a better evaluation in every phase of the game. I compared it with some much higher ranked engines, and their advantage is, for me, obviously only the depth. The problems with blunders have been addressed, the level of play in the endgame is way better, and style wise, I feel it is more balanced. The style seems less risky but remained active.

I ran a short match against Zahak, and despite being outsearched sometimes even by four plies, Toga Rebel stood his ground easily. In simpler positions, where search was not crucial, TR improved steadily its positions and usually managed to win. In complicated positions it had difficulties, especially when the difference in search depths was big.

What I did not like is that it avoided draws in K+R vs K+R positions, but, hey!, nobody's perfect.


Edit: Definitively a huge improvement in simple positions and endings.

Thanks for your insight, much appreciated.

I am currently testing Toga IV 1.1 in comparison with Rebel 14.1 and 14.2 in an unofficial tournament with other AB engines.

Toga IV 1.1 is therefore running with 1 thread to make the results comparable. Toga IV 1.1 is currently slightly ahead of Rebel 14.2, which I expected.I would be happy if a new Rebel based on the Toga version appears to avoid future irritations.

My Rebel 14.2 test for CEGT 40/4 is still running, after 2500 games I calculate the preliminary first result for Rebel 14.2.

I see that Rebel 14.2 scales well. He is +60 on the CEGT 40/20 list.

Quote :

Toga IV 1.1 is currently slightly ahead of Rebel 14.2, which I expected.I would be happy if a new Rebel based on the Toga version appears to avoid future irritations.

I agree with you. Imho, Ed should switch to the Toga base for further development.

matejst wrote:

I see that Rebel 14.2 scales well. He is +60 on the CEGT 40/20 list.

Quote :

Toga IV 1.1 is currently slightly ahead of Rebel 14.2, which I expected.I would be happy if a new Rebel based on the Toga version appears to avoid future irritations.

I agree with you. Imho, Ed should switch to the Toga base for further development.

Don't worry, no more name confusion, next will be Rebel 14.3 based on Toga 4.11, in the between time I gained ~25 elo, shall I release?































Just kidding.

From my point of view, of course you should release.  

matejst wrote:

From my point of view, of course you should release.  

But there are sanctions against me

Hi Ed, after further unofficial tests against other engines on my old PC, Toga IV 1.1 (1 thread) is about 15 to 20 Elo ahead of Rebel 14.2 after 400 games so far. I would look forward to a new Rebel 14.3 (or better 15.0) based on Toga.

Dio wrote:

Hi Ed, after further unofficial tests against other engines on my old PC, Toga IV 1.1 (1 thread) is about 15 to 20 Elo ahead of Rebel 14.2 after 400 games so far. I would look forward to a new Rebel 14.3 (or better 15.0) based on Toga.

Thanks for testing, if you set "Delta Margin" to 200 (instead of 50) you get some 18 elo more based on 10,000 games.

From my notes:

Code:

******************************************************************
****                                                          ****
****          Rebel 14.3 testing using TOGA source code       ****
****             [FQ.PGN] [40/20] [5000 games]                ****
****                                                          ****
****                        AVX2     SSE                      ****
****        QS-100         51.8%   51.5%                      ****
****        QS-200         52.7%   52.4%                      ****
****        QS-300         52.3%   52.4%                      ****
****        QS-500         52.6%   51.9%                      ****
****                                                          ****
******************************************************************

Great news!

Then, my two cents. After two, three days of using Toga Rebel, here is where I see possible improvements:

1. Of course, the search. I think that the quality of the evaluation is, let's say, at the level of a 3450 (CCRL) engine, and the search is at the level of a 3100 engine. The NN used in Dark Toga was similar in strength to SF nets, but Dark Toga reached an Elo of only 3150 on the CCRL list. I feel there still remains a lot of room for easy improvements.

2. That's also what I saw when I tested Toga Rebel against Zahak and Wasp. All Toga's losses were in the middlegame, where Toga got 1-4 plies less than its opponents. Past that stage, Toga almost always achieved better positions in the ensuing late middlegames and endgames, or managed to improve worse positions. Toga treated very well positions with a Knight against a Bishop (usually with rooks) for both sides; it also made a good use of the bishop pair; Toga also understood positions with bishops of different colors.

3. The NN was especially improved for later phases of the game, but it lacks some elementary knowledge about drawn positions (e.g., in opposite colors bishops endings, or bishops + h pawn vs K in the wrong corner), and perhaps training a specialized endgame net could help.

4. Since I still do not care too much about speed, I am so far very happy how the development of Rebel goes. The net is very mature, the search is steadily improving, and with the Toga base the engine is comfortable. The style is less aggressive than it was, but it is more sound.

I would prefer Rebel 15.0 (based on Toga) instead of Rebel 14.3 as the name of the new release, to have a differentiation to Rebel 14.x (based on Fruit), just a suggestion, nothing more.

I think this is also important for future results in rating lists and for the normal user to clearly separate the two development directions (Fruit/Toga).

Of course this name giving is mainly relevant for elo lists.
And also the frequency of change is a pain in the ass for elo lists. We others… meanwhile..,.

I think the name is not only important for rating lists, it is meant to show a concrete difference between two development directions. Even experienced testers seem to have difficulty understanding that Toga IV 1.1 is nothing more than Rebel 14 with MP search. A Rebel 14.3 instead of 15.0 will cause further irritation, I am sure. Personally, I couldn't care less about the name...

Ok. I see what you mean Jörg. The name is Eds decision. Maybe you can convince him. From my experience with programmers they very often come with new versions. Not any version is a progress.
So the tester normally finds out before the rating guys know.
E.g. with don dailey or chris Whittington i had more then 1 version a day. Thats the normal process of trial and error. Only the best versions reach the public.

When i worked with chris on CSTAL it was the age of telephone, fax and floppy disc.
Nevertheless i had chris on telephone at least 2 times a day.
He was like a brother for me and i telephoned more with him then i did with my family.

see you later Thorsten..

Jörg i am very happy that we still do computerchess.

Admin wrote:

• TOGA IV now with Rebel 14.2 NNUE evaluation.
• Toga IV 1.1 is about 45 elo stronger than Rebel 14.2 and above that offers you:
• Multiple threading up to 64 threads;
• MultiPV up to 10 moves.

• TOGA IV 1.1 earned a rating of 3323 on the Gambit Rating List.

• Offered to you in cooperation with Chris Whittington.

http://rebel13.nl/windows/rebel-14.html#other

Ratinglist-testrun of Toga 4.11 finished: Sorry for the mismatch in naming: Toga 4.11... Toga 4 1.11 would be correct. But, I hope, a new version comes soon by Ed Schroeder and will be named "Rebel", then this Toga-version will be deleted out of the ratinglist.

https://www.sp-cc.de/