REBEL 15.1

REBEL 15.1
1. New network architecture, version 10.
2. Increased data set from 2.75 billion positions to 3.6 billion positions.
3. Search improvements.
4. Estimated elo gain based on GRL testing ~56 elo.
5. Rebel 15.1 development diary.

http://rebel13.nl/windows/rebel-15.html

Thanks ed.
Will never forget how i wrote the MM4 and MM5 article for Computer Chess and „Spiele“.

your fan since then.
And 1986…. In cologne. Saw you and Jan with the hair fan trying to cool the accelerator card.

Mclane wrote:

Thanks ed

Thank you for your contributions.

Today i was in Düsseldorf. Capitol of NRW. Got a new dedicated chess computer. In Düsseldorf is Arno Kreuzberg who repairs the old machines. This is my new toy:

Self moving Fidelity machine with Spracklen program and LCD Display.

I actually played a game against the Phantom at an exhibition around 1990. It was very impressive technology, but had a dark, negative side: it didn't show respect for its creator species, and cheekily won the game.

At that same exhibition, I also:

* met Stephen Wolfram who was touting his product, Mathematica (I shouted out from the crowd and challenged him with a problem I didn't think it would be able to solve: he accepted the impolite challenge and it took him something like 10 seconds, and solved it MUCH better than I was expecting)

* used a virtual reality headset: the computer was about the size of a toilet room, the headset was big and heavy, and the 3d images were very blocky. When I took the headset off, the real world looked completely weird!

Those were days

running...

Rebel 15.1NN x64 1CPU

LC0 0.28.2 DX12 Vega11 771721   3497  00,0  xxxx [W.B.]
SlowChess Blitz 2.9NN x64       3465  00,0  xxxx [W.B.]
Seer 2.5.0NN x64 1CPU           3424  00,0  xxxx [W.B.]
Igel 3.1.0NN x64 1CPU           3358  00,0  xxxx [W.B.]
Nemorino 6.05Net16 x64 1CPU     3323  00,0  xxxx [W.B.]
Clover 3.1NN x64 1CPU           3296  00,0  xxxx [W.B.]
Wasp 5.50NN x64 1CPU            3259  00,0  xxxx [W.B.]
Xiphos 0.6 x64 1CPU             3231  00,0  xxxx [W.B.]
Booot 6.5 x64 1CPU              3222  00,0  xxxx [W.B.]
Marvin 6.0.0NN x64 1CPU         3202  00,0  xxxx [W.B.]
Revenge 1.0NN x64 1CPU          3363  51,5  3373 [J.B.]
Uralochka 3.37cNN x64 1CPU      3357  49,0  3350 [J.B.] 
Komodo 13.02 x64 1CPU           3341  00,0  xxxx [J.B.]
Fritz 18NN x64 1CPU             3257  00,0  xxxx [J.B.]
Fire 8.2 x64 1CPU               3338  00,0  xxxx [J.B.]
Arasan 23.3NN x64 1CPU          3347  00,0  xxxx [J.B.]
RofChade 3.0NN x64 1CPU         3449  00,0  xxxx [J.B.]


more to follow

Performance = ca. ELO xxxx / xxxx games => +xy  to v15x2 (3291)
                                        => +xy  to v15.0 (3262)

https://cegt.forumieren.com/t1771-testing-rebel-15-1nn

65% against Fritz 18, wow...

updated after 1400 games.. nice progress.

Dio wrote:

updated after 1400 games.. nice progress.

Surprising good. I was worried about the scaling. Relieved...

I am very happy about the results so far, also against Rubichess it looks very good.

I have decided to test against Stoofvlees II a19 as well, but this test is much more elaborate as only 1 game can be played at a time. The same applies to the test against Lc0, here an iGPU from AMD (Vega 11) is used.

Quite a job rating Lc0.

CCRL -

https://tlcv.net/16094/result-table

just a test to show people that no expensive graphics card is needed for Lc0, but any (even very old) computer from AMD.

no question, due to the large amount of testing, Lc0 and Stoofvlees are only used against very strong programs and Rebel 15.1 is now one of them.

after 1500 games +80 Elo compared with Rebel 15x2.

Dio, can the CEGT software produce such a version history like CCRL?

Ed, the improvement of Rebel 15.1 is amazing. But still, the weakness in problems solving remains. I just had a position where Marvin/Benjamin NN finds mate in 5 instantly, while Rebel 15.1 simply cannot find the solution. Just try this position after move 17.Qe2.

[pgn]
[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "?"]
[Black "?"]
[Result "*"]
[GameNo "172"]

1. e4 c5 2. Nf3 Nc6 3. d4 cxd4 4. Nxd4 e5 5. Nb5 d6 6. c4 Be7 7. N1c3 a6 8.
Na3 f5 9. Bd3 f4 10. g3 Nf6 11. gxf4 exf4 12. Bxf4 Ng4 13. Qd2 O-O 14. Bg3
Nxf2 {!} (14... Nce5 15. Be2 {?}) 15. Bxf2 Rxf2 {!} 16. Kxf2 Bg5 17. Qe2 Bh4+
18. Kf1 Qf6+ *
[/pgn]

Hello Ed, I can't say.

The version list of CCRL is an advantage to our list.

The CEGT list is created by our webmaster Gerhard Sonnabend. According to my information, he first has to copy all new games and convert them for Chessbase. The recalculation of the respective lists then takes several hours according to my knowledge.

Dio wrote:

Hello Ed, I can't say.

The version list of CCRL is an advantage to our list.

The CEGT list is created by our webmaster Gerhard Sonnabend. According to my information, he first has to copy all new games and convert them for Chessbase. The recalculation of the respective lists then takes several hours according to my knowledge.

No need to convert games, all the info is in the html, I will see if I can create a util that does the job.

matejst wrote:

Ed, the improvement of Rebel 15.1 is amazing. But still, the weakness in problems solving remains. I just had a position where Marvin/Benjamin NN finds mate in 5 instantly, while Rebel 15.1 simply cannot find the solution. Just try this position after move 17.Qe2.

[pgn]
[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "?"]
[Black "?"]
[Result "*"]
[GameNo "172"]

1. e4 c5 2. Nf3 Nc6 3. d4 cxd4 4. Nxd4 e5 5. Nb5 d6 6. c4 Be7 7. N1c3 a6 8.
Na3 f5 9. Bd3 f4 10. g3 Nf6 11. gxf4 exf4 12. Bxf4 Ng4 13. Qd2 O-O 14. Bg3
Nxf2 {!} (14... Nce5 15. Be2 {?}) 15. Bxf2 Rxf2 {!} 16. Kxf2 Bg5 17. Qe2 Bh4+
18. Kf1 Qf6+ *
[/pgn]

Thanks for the position. That's a pretty weak point in the search. ProDeo (600 elo lower) spots it instantly. Doesn't feel good.

I just finished the test against Stoofvlees II a19, an extremely strong performance for Rebel 15.1 with only one thread.

I had a look at Rebel's results. A lots of draws for the quality of the evaluation, which seems to be almost at SF level or, at least, very close (I analyzed dozens of positions with several engines and compared). I am afraid that this search bug has an important impact on results, especially in complicated positions.

Ratinglist-testrun of Rebel 15.1 finished. Huge progress (in Elo and EAS) !


https://www.sp-cc.de

https://www.sp-cc.de/eas-ratinglist.htm

(Perhaps you have to clear your browsercache or reload the website)

pohl4711 wrote:

Ratinglist-testrun of Rebel 15.1 finished. Huge progress (in Elo and EAS) !


https://www.sp-cc.de

https://www.sp-cc.de/eas-ratinglist.htm

(Perhaps you have to clear your browsercache or reload the website)

Thanks for testing Stefan.

You never know for sure with rating lists, my testing only showed +59, now it's +103. Often it's the other way around.

Admin wrote:

pohl4711 wrote:

Ratinglist-testrun of Rebel 15.1 finished. Huge progress (in Elo and EAS) !


https://www.sp-cc.de

https://www.sp-cc.de/eas-ratinglist.htm

(Perhaps you have to clear your browsercache or reload the website)

Thanks for testing Stefan.

You never know for sure with rating lists, my testing only showed +59, now it's +103. Often it's the other way around.

Only explanation here is, that your gambit-openings are too special. I use 500 balanced HERT openings (HERT= Human and Engine relevant theory), means a bunch of "normal" chess openings.
Because Rebel (Toga) is outdated (compared to the other strong engines in the SPCC-list) a little in doing prunings etc, Rebel has perhaps weaker tactics. That could lead to a disadvantage, when using gambit-openings, which are more tactically.
My theory... (And in CEGT it looks like +100 Elo, too. And they use "normal" openings, too).