Release Rebel 15x2

I am quite ok with the speed, Ed, I don't look for anything better otherwise I would be using SF. I just wanted to help you with details I noticed. I am not a big fan of "fast and furious" engines. Rather slow with a good evaluation.

Nevertheless I still think you should add some check extensions to improve the tactical abilities of Rebel since it would fit well with its evaluation.

Now I use Orion (which is even slower) in the background to check tactics. It is a good combo, but I would prefer to use solely Rebel.

Ratinglist-testrun of Rebel 15x2 finished.

https://www.sp-cc.de

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

Thank you Stefan. Not bad at all on your time control.

pohl4711 wrote:

Ratinglist-testrun of Rebel 15x2 finished.

https://www.sp-cc.de

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

Apologies for an off-topic question, but, if there's a quick answer, how did Stockfish get so far clear at the top?

As top engines become more and more difficult to beat, that ought to be almost impossible.

Rebel 15x2 live (CCRL) div 1: https://tlcv.net/16091

first game, a draw against Koivisto 8.0

Whoops, division 1, Graham...!?

TheSelfImprover wrote:

pohl4711 wrote:

Ratinglist-testrun of Rebel 15x2 finished.

https://www.sp-cc.de

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

Apologies for an off-topic question, but, if there's a quick answer, how did Stockfish get so far clear at the top?

As top engines become more and more difficult to beat, that ought to be almost impossible.

Stockfish is a community project, lots of possible ideas as candidate improvements are contributed by people. Where 2 people cooperate 1+1 is already 3 and sometimes 4.

Dio wrote:

Rebel 15x2 live (CCRL) div 1: https://tlcv.net/16091

first game, a draw against Koivisto 8.0

Code:

RANK   ENGINE                                    GAMES  POINTS   1  2  3  4  5  6  7  8  9  10 11 12
1.   BERSERK 8.5.1 64-BIT                         4     3.0    *  =  .  .  .  .  .  .  .  =  1  1
2.   SLOWCHESS BLITZ 2.8 64-BIT                   4     3.0    =  *  .  .  .  .  .  =  .  .  1  1
3.   RUBICHESS 20220223 64-BIT                    5     3.0    .  .  *  =  =  .  =  .  =  .  .  1
4.   STOCKFISH 15 64-BIT                          4     2.5    .  .  =  *  =  =  .  .  .  .  .  1
5.   KOIVISTO 8.0 64-BIT                          4     2.5    .  .  =  =  *  .  =  .  .  .  .  1
6.   SEER 2.5.0 64-BIT                            4     2.5    .  .  .  =  .  *  =  .  1  =  .  .
7.   REBEL 15X2 64-BIT                            4     2.0    .  .  =  .  =  =  *  =  .  .  .  .
8.   IGEL 3.0.5 64-BIT                            4     2.0    .  =  .  .  .  .  =  *  =  =  .  .
9.   ARASAN 23.3 64-BIT                           4     2.0    .  .  =  .  .  0  .  =  *  .  1  .
10.  ROFCHADE 3.0 64-BIT                          4     2.0    =  .  .  .  .  =  .  =  .  *  =  .
11.  FIRE 8.2 64-BIT                              4     0.5    0  0  .  .  .  .  .  .  0  =  *  .
12.  BOOOT 7.0-BETA 64-BIT                        5     0.0    0  0  0  0  0  .  .  .  .  .  .  *
Total games = 25

4 draws so far

Dio wrote:

Rebel 15x2 live (CCRL) div 1: https://tlcv.net/16091

first game, a draw against Koivisto 8.0

Code:

RANK   ENGINE                                    GAMES  POINTS 1  2  3  4  5  6  7  8  9  10 11 12
1.   STOCKFISH 15 64-BIT                          7     5.0    *  =  .  1  =  .  =  .  .  1  =  1
2.   SEER 2.5.0 64-BIT                            8     5.0    =  *  .  .  =  =  =  1  =  =  .  1
3.   SLOWCHESS BLITZ 2.8 64-BIT                   7     4.5    .  .  *  =  .  =  .  =  =  =  1  1
4.   BERSERK 8.5.1 64-BIT                         7     4.0    0  .  =  *  .  =  .  =  =  .  1  1
5.   KOIVISTO 8.0 64-BIT                          7     4.0    =  =  .  .  *  =  =  .  .  =  =  1
6.   REBEL 15X2 64-BIT                            8     4.0    .  =  =  =  =  *  =  .  =  =  =  .
7.   RUBICHESS 20220223 64-BIT                    7     4.0    =  =  .  .  =  =  *  =  .  =  .  1
8.   ARASAN 23.3 64-BIT                           8     4.0    .  0  =  =  .  .  =  *  =  =  1  =
9.   ROFCHADE 3.0 64-BIT                          8     4.0    .  =  =  =  .  =  .  =  *  =  =  =
10.  IGEL 3.0.5 64-BIT                            8     3.5    0  =  =  .  =  =  =  =  =  *  .  .
11.  FIRE 8.2 64-BIT                              7     2.0    =  .  0  0  =  =  .  0  =  .  *  .
12.  BOOOT 7.0-BETA 64-BIT                        8     1.0    0  0  0  0  0  .  0  =  =  .  .  *
Total games = 45

8 draws now

Ed, please check the search. It could well be buggy, although I am not sure if is something to do with the GUI.

When I analyze, I sometimes get nonsensical lines, with systematic blunders. It doesn't happen immediately, when I start analyzing, but after some times. In such cases, I also noted that the search slows down.

I thought it was related with the complexity of the position, but I am not sure, I still cannot find which kind of positions triggers this behaviour. Another hypothesis is that has something to do with engine switching -- e.g., today I compare Rebel's lines with SF, and the problem started after that.

Anyway, I think it could have an impact on Rebel's results. In general, its evaluation is outstanding -- in the opening and middlegame, it is at SF's level or very close.

Edit: the endgame play is still much worse than the middlegame.

Can you please post an example, EPD or PGN so I can check?

Admin wrote:

Can you please post an example, EPD or PGN so I can check?

I will try, give me just a few minutes.

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

1. e4 Nf6 2. e5 Nd5 3. d4 d6 4. c4 Nb6 5. exd6 cxd6 6. Nc3 g6 7. Be3 Bg7 8.
Rc1 O-O 9. b3 Nc6 10. d5 Ne5 11. Be2 Bf5 12. f4 Ned7 13. Nf3 Nf6 14. Nd4 Bd7
15. O-O Qc8 16. h3 Re8 17. Qd2 e5 18. dxe6 Bxe6 19. f5 Bd7 20. Kh1 Bc6 21.
Nxc6 bxc6 22. Bg5 Qd8 23. fxg6 hxg6 24. Bg4 Re5 25. Bxf6 Bxf6 26. Qf4 Bg7 27.
Qxf7+ Kh7 28. Ne2 Qg5 29. Qc7 Rae8 30. Ng1 R8e7 31. Qxd6 Re3 32. Qxc6 Rg3 33.
Nf3 Qf4 34. Rce1 Rxe1 35. Rxe1 Qh6 36. Re4 Qc1+ 37. Kh2 Rxf3 38. Bxf3 Be5+ 39.
Rxe5 Qf4+ 40. g3 Qxe5 41. c5 Qb2+ 42. Bg2 Qxa2 43. cxb6 axb6 44. Qb7+ Kh6 45.
Qxb6 Qb2 46. b4 *

[/pgn]

The lines after 36.Re4.

[pgn]
[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "?"]
[Black "?"]
[Result "*"]
[FEN "8/p5bk/1nQ3pq/8/2P1R1B1/1P3NrP/P5P1/7K b - - 0 36"]
[SetUp "1"]

36... Na4 (36... Qc1+ 37. Kh2 Nd5 (37... Rxg2+ 38. Kxg2 Nd5 39. cxd5 Qb2+ 40.
Nd2 Qxd2+ 41. Be2 (41. Re2 Qf4 42. Qa4 Qh2+ 43. Kxh2 Be5+ 44. Rxe5 Kg8 45. Qa3
{(0:01:08) 8936kN}{[%eval 622,13,Rebel 15x2]})41... Qg5+ 42. Rg4 Kh6 (42...
Qxd5+ 43. Qxd5 Bd4 {(0:01:15) 9783kN}{[%eval 638,14,Rebel 15x2]})43. Rxg5 Bf6
44. Qe8 Bg7 45. Rxg6+ Kh7 46. Qh8+ Kxg6 {(0:01:24) 10965kN}{[%eval
654,14,Rebel 15x2]}) (37... Na4 38. Kxg3 Nc3 39. Re7 Kh6 40. Qb7 Qf4+ 41. Kf2
{(0:00:50) 6531kN}{[%eval 636,12,Rebel 15x2]}) (37... Nxc4 38. Kxg3 Qb2 39.
Re2 (39. Ng5+ Kh6 40. Qa8 {(0:00:23) 2991kN}{[%eval 620,11,Rebel 15x2]})(39.
bxc4 Qb8+ 40. Kf2 Qb2+ 41. Kg1 (41. Ke3 Qb6+ 42. Qxb6 {(0:00:17) 2218kN}
{[%eval 639,10,Rebel 15x2]})41... Qb1+ 42. Kh2 Qf1 43. Kg3 {(0:00:18) 2454kN}
{[%eval 656,10,Rebel 15x2]})39... Be5+ 40. Nxe5 Nd2 (40... Qc1 41. bxc4 Qc3+
42. Kf2 Qd4+ 43. Re3 Kg8 44. Nxg6 Qb2+ 45. Re2 {(0:00:29) 3779kN}{[%eval
636,12,Rebel 15x2]})41. Qxg6+ Kh8 42. Kf2 Qd4+ 43. Ke1 Qc5 44. Kxd2 {(0:00:37)
4878kN}{[%eval 652,12,Rebel 15x2]})(37... Be5 38. Rxe5 Qh1+ 39. Kxg3 Qxg2+ 40.
Kf4 Qxg4+ 41. Ke3 Qg1+ 42. Ke4 Qe1+ {(0:00:03) 508kN}{[%eval 666,8,Rebel 15x2]}
)38. Kxg3 Nc3 39. Re7 Kh6 40. Qc7 Qe1+ 41. Kh2 Qh1+ 42. Kxh1 Nb1 43. Qf4+ Kh7
{(0:01:50) 14129kN}{[%eval 672,14,Rebel 15x2]})(36... Rxf3 37. Bxf3 Na4 (37...
Qc1+ 38. Kh2 Be5+ 39. g3 Bxg3+ 40. Kxg3 Qg5+ 41. Rg4 Qe5+ 42. Rf4 (42. Kf2
Qg3+ 43. Ke2 Qf2+ 44. Kxf2 Kg8 45. Qa4 {(0:00:04) 551kN}{[%eval 623,8,Rebel
15x2]})42... Qe1+ 43. Kh2 Qd2+ 44. Kg1 Qxf4 45. Qb7+ Kh6 {(0:00:04) 567kN}
{[%eval 587,8,Rebel 15x2]})38. Kh2 Nc3 39. Re7 Qf4+ 40. g3 {(0:00:04) 588kN}
{[%eval 651,8,Rebel 15x2]}) 37. Qxa4 a6 (37... Qc1+ 38. Re1 Qc2 39. Re2 Qc1+
40. Ng1 Rxg2 41. Qxa7 Qxg1+ 42. Qxg1 Rh2+ 43. Rxh2 Bf8 44. Kg2 Kh6 45. Bd1
{(0:01:54) 14546kN}{[%eval 664,14,Rebel 15x2]}) (37... Rxg2 38. Re7 Rg3 39.
Qxa7 Rxh3+ 40. Kg2 Rh2+ 41. Nxh2 Qxh2+ 42. Kxh2 Kh6 43. Re2 {(0:00:15) 1950kN}
{[%eval 657,10,Rebel 15x2]})(37... Rxf3 38. Bxf3 Qc1+ (38... a6 39. Kh2 Qf4+
40. Rxf4 Kh6 41. Rh4+ Kg5 42. Qa5+ Kf6 {(0:00:04) 605kN}{[%eval 619,8,Rebel
15x2]})39. Kh2 Qg1+ 40. Kg3 Qf2+ (40... g5 41. Qa6 Qh2+ 42. Kxh2 Bf8 43. Kh1
{(0:00:05) 704kN}{[%eval 639,9,Rebel 15x2]})41. Kxf2 Kh6 42. Qxa7 Bf8 43. Rh4+
Kg5 44. Qe7+ Kf5 45. Qxf8+ {(0:00:05) 783kN}{[%eval 655,10,Rebel 15x2]}) 38.
Qxa6 Qc1+ 39. Kh2 Rxf3 40. Bxf3 g5 41. Qa7 Qh1+ 42. Kg3 Qxh3+ 43. Kf2 Qxg2+
44. Bxg2 Kh6 {(0:02:29) 19097kN}{[%eval 685,15,Rebel 15x2]} *

[/pgn]

Suddenly, Rebel 15x2 needs 3mn to achieve depth 14, while it is usually almost instantaneous. It could be the GUI, I will check.

Same behaviour under Aquarium and Shredder GUI.

Edit: Rebel 15 behaves in a similar way.

matejst wrote:

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

1. e4 Nf6 2. e5 Nd5 3. d4 d6 4. c4 Nb6 5. exd6 cxd6 6. Nc3 g6 7. Be3 Bg7 8.
Rc1 O-O 9. b3 Nc6 10. d5 Ne5 11. Be2 Bf5 12. f4 Ned7 13. Nf3 Nf6 14. Nd4 Bd7
15. O-O Qc8 16. h3 Re8 17. Qd2 e5 18. dxe6 Bxe6 19. f5 Bd7 20. Kh1 Bc6 21.
Nxc6 bxc6 22. Bg5 Qd8 23. fxg6 hxg6 24. Bg4 Re5 25. Bxf6 Bxf6 26. Qf4 Bg7 27.
Qxf7+ Kh7 28. Ne2 Qg5 29. Qc7 Rae8 30. Ng1 R8e7 31. Qxd6 Re3 32. Qxc6 Rg3 33.
Nf3 Qf4 34. Rce1 Rxe1 35. Rxe1 Qh6 36. Re4 Qc1+ 37. Kh2 Rxf3 38. Bxf3 Be5+ 39.
Rxe5 Qf4+ 40. g3 Qxe5 41. c5 Qb2+ 42. Bg2 Qxa2 43. cxb6 axb6 44. Qb7+ Kh6 45.
Qxb6 Qb2 46. b4 *

[/pgn]

The lines after 36.Re4.

Finally have the time to look into this after 2 days of debugging.

What's the problem after 36.Re4, I don't see it and what color is Rebel?

[pgn]
[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "?"]
[Black "?"]
[Result "*"]
[FEN "8/p5bk/1nQ3pq/8/2P1R1B1/1P3NrP/P5P1/7K b - - 0 36"]
[SetUp "1"]

36... Na4 (36... Qc1+ 37. Kh2 Nd5 (37... Rxg2+ 38. Kxg2 Nd5 39. cxd5 Qb2+ 40.
Nd2 Qxd2+ 41. Be2 (41. Re2 Qf4 42. Qa4 Qh2+ 43. Kxh2 Be5+ 44. Rxe5 Kg8 45. Qa3
{(0:01:08) 8936kN}{[%eval 622,13,Rebel 15x2]})41... Qg5+ 42. Rg4 Kh6 (42...
Qxd5+ 43. Qxd5 Bd4 {(0:01:15) 9783kN}{[%eval 638,14,Rebel 15x2]})43. Rxg5 Bf6
44. Qe8 Bg7 45. Rxg6+ Kh7 46. Qh8+ Kxg6 {(0:01:24) 10965kN}{[%eval
654,14,Rebel 15x2]}) (37... Na4 38. Kxg3 Nc3 39. Re7 Kh6 40. Qb7 Qf4+ 41. Kf2
{(0:00:50) 6531kN}{[%eval 636,12,Rebel 15x2]}) (37... Nxc4 38. Kxg3 Qb2 39.
Re2 (39. Ng5+ Kh6 40. Qa8 {(0:00:23) 2991kN}{[%eval 620,11,Rebel 15x2]})(39.
bxc4 Qb8+ 40. Kf2 Qb2+ 41. Kg1 (41. Ke3 Qb6+ 42. Qxb6 {(0:00:17) 2218kN}
{[%eval 639,10,Rebel 15x2]})41... Qb1+ 42. Kh2 Qf1 43. Kg3 {(0:00:18) 2454kN}
{[%eval 656,10,Rebel 15x2]})39... Be5+ 40. Nxe5 Nd2 (40... Qc1 41. bxc4 Qc3+
42. Kf2 Qd4+ 43. Re3 Kg8 44. Nxg6 Qb2+ 45. Re2 {(0:00:29) 3779kN}{[%eval
636,12,Rebel 15x2]})41. Qxg6+ Kh8 42. Kf2 Qd4+ 43. Ke1 Qc5 44. Kxd2 {(0:00:37)
4878kN}{[%eval 652,12,Rebel 15x2]})(37... Be5 38. Rxe5 Qh1+ 39. Kxg3 Qxg2+ 40.
Kf4 Qxg4+ 41. Ke3 Qg1+ 42. Ke4 Qe1+ {(0:00:03) 508kN}{[%eval 666,8,Rebel 15x2]}
)38. Kxg3 Nc3 39. Re7 Kh6 40. Qc7 Qe1+ 41. Kh2 Qh1+ 42. Kxh1 Nb1 43. Qf4+ Kh7
{(0:01:50) 14129kN}{[%eval 672,14,Rebel 15x2]})(36... Rxf3 37. Bxf3 Na4 (37...
Qc1+ 38. Kh2 Be5+ 39. g3 Bxg3+ 40. Kxg3 Qg5+ 41. Rg4 Qe5+ 42. Rf4 (42. Kf2
Qg3+ 43. Ke2 Qf2+ 44. Kxf2 Kg8 45. Qa4 {(0:00:04) 551kN}{[%eval 623,8,Rebel
15x2]})42... Qe1+ 43. Kh2 Qd2+ 44. Kg1 Qxf4 45. Qb7+ Kh6 {(0:00:04) 567kN}
{[%eval 587,8,Rebel 15x2]})38. Kh2 Nc3 39. Re7 Qf4+ 40. g3 {(0:00:04) 588kN}
{[%eval 651,8,Rebel 15x2]}) 37. Qxa4 a6 (37... Qc1+ 38. Re1 Qc2 39. Re2 Qc1+
40. Ng1 Rxg2 41. Qxa7 Qxg1+ 42. Qxg1 Rh2+ 43. Rxh2 Bf8 44. Kg2 Kh6 45. Bd1
{(0:01:54) 14546kN}{[%eval 664,14,Rebel 15x2]}) (37... Rxg2 38. Re7 Rg3 39.
Qxa7 Rxh3+ 40. Kg2 Rh2+ 41. Nxh2 Qxh2+ 42. Kxh2 Kh6 43. Re2 {(0:00:15) 1950kN}
{[%eval 657,10,Rebel 15x2]})(37... Rxf3 38. Bxf3 Qc1+ (38... a6 39. Kh2 Qf4+
40. Rxf4 Kh6 41. Rh4+ Kg5 42. Qa5+ Kf6 {(0:00:04) 605kN}{[%eval 619,8,Rebel
15x2]})39. Kh2 Qg1+ 40. Kg3 Qf2+ (40... g5 41. Qa6 Qh2+ 42. Kxh2 Bf8 43. Kh1
{(0:00:05) 704kN}{[%eval 639,9,Rebel 15x2]})41. Kxf2 Kh6 42. Qxa7 Bf8 43. Rh4+
Kg5 44. Qe7+ Kf5 45. Qxf8+ {(0:00:05) 783kN}{[%eval 655,10,Rebel 15x2]}) 38.
Qxa6 Qc1+ 39. Kh2 Rxf3 40. Bxf3 g5 41. Qa7 Qh1+ 42. Kg3 Qxh3+ 43. Kf2 Qxg2+
44. Bxg2 Kh6 {(0:02:29) 19097kN}{[%eval 685,15,Rebel 15x2]} *

[/pgn]

Quote :

Suddenly, Rebel 15x2 needs 3mn to achieve depth 14, while it is usually almost instantaneous. It could be the GUI, I will check.

It's true, the Toga search is not good with checks and pawn promotions that end up with checks causing search explosions. I will try my own check system, fixed that problem in the 80's.

http://talkchess.com/forum3/viewtopic.php?p=926848#p926848

Managed to remain in division one

Meanwhile... Did Rebel lost the "Benjamin style"? I still use Marvin with the Benjamin78 NN, and while it seems weaker in the endings, and sometimes too optimistic in the early middlegame (the openings are a topic that should be treated apart), when I compare it with Rebel15x2, its style of play seems much more active, more aggressive than the style of Rebel15x2 (and Rebel 15), which became somehow a bit drawish.

I think that it is due in part because of the inadequacies of the Toga search -- it simply misses tactical shots. I wrote several time here that Rebel 15 misses mates seen not only by other engines, but "felt" by me, if not outright calculated (although only once, when it was a direct, forced mate in six, with checks along the way). Marvin/Benjamin found this particular mate right away.

Then, there is another thing. Short story long, to try to explain myself. I also tried two engines whose nets were based mainly on ending positions: Seer and SCB. Their analysis in the opening are... let's say correct, but when the position is not straightforward, Rebel is way better (in all versions). Despite this disadvantage in openings, and while not being ultra-fast, both engines fare well on different rating lists and in the TCEC competition. I attribute their good results -- following John Stanback remark at CCC -- to their outstanding game in simple positions and endings. Having tested SCB a bit more (Seer has an unpleasant bug), I can attest that it is really the case.

So, I was thinking. Was it not possible to safeguard the Benjamin aggressive style in openings, and add lots of ending knowledge to improve his results (and fix the Toga search). Could a selective approach to data/training retain the Rebel idiosyncrasies, which made it not only so valuable for human analysis, but also so original? Can Rebel remain more than "just another NN engine", like it was at his beginnings?

Finally, a disclaimer. All my remarks are made without knowledge of the way Rebel was done. It is just an impression, and I could be utterly wrong.

Ed, just to confirm. I think now that the reason for Rebel's lack of aggressiveness is because it simply does not find tactical shots. Like I wrote before, I noticed it when analyzing, but then I decided to test both Marvin/Rebel and Rebel 15x2 with the "Tactics for intermediate players" base (exercises I do myself) and Rebel 15x2 had difficulties finding solutions longer than mate in 4, while Marvin/Rebel found them immediately.

You can try for yourself with problems suited for chess engines, I believe the problem would become clear. I believe it has an important impact on the results of Rebel.

Endgames -- while Rebel 15x2 is better in simple positions, it still lacks behind the first versions of CSTal NN in endgame analysis. I worked on rook endgames and transition to pawn endgames yesterday, and analyzed the positions with both engines. Despite the handicap of not having TBs support implemented, CSTal still plays endings clearly better than Rebel 15x2. Rebel somehow still lacks specialized endgame knowledge. Let's have a look at this position:

[puzzle]
[SetUp "1"]
[FEN "8/3b2k1/8/3B4/8/r7/5RPP/6K1 b - -"]

1... Ra1+ {2:36m} 2. Rf1 {+2.08/4 0s} Rxf1+ {2s} 3. Kxf1
{+2.08/4 0s} Bh3 {1s} 4. gxh3 {+2.08/27 2s} *
[/puzzle]

SlowChess, even without use of TBs solves it easily. So, this knowledge can be trained, and specialized ending knowledge would probably improve drastically the results of the engine. Why not add TB positions to the training data, like Connor Macmonigle did?

Recapitulation.

Rebel's search misses tactical shots and not only probably alters drastically the style of the engine, but influences its results vastly. With a similar net, Rebel 14 was 100 Elos weaker than Marvin/Rebel. Solving the problems with the search will probably not yield so much Elos, but could well hugely improve the results.

Rebel still lacks technical, robust endgame knowledge. It is not only about bishops of different colors -- there, the situation is much better -- but it seems it is simply not trained on enough ending positions.

it took a while until the games of Rebel 15x2 were evaluated, because our webmaster was very busy, the official first entry of Rebel 15x2 in the 40/4 CEGT list. More results will follow...

http://www.cegt.net/40_4_Ratinglist/40_4_single/262.html

http://www.cegt.net/40_4_Ratinglist/40_4_single/rangliste.html

matejst wrote:

Endgames -- while Rebel 15x2 is better in simple positions, it still lacks behind the first versions of CSTal NN in endgame analysis. I worked on rook endgames and transition to pawn endgames yesterday, and analyzed the positions with both engines. Despite the handicap of not having TBs support implemented, CSTal still plays endings clearly better than Rebel 15x2. Rebel somehow still lacks specialized endgame knowledge. Let's have a look at this position:

[puzzle]
[SetUp "1"]
[FEN "8/3b2k1/8/3B4/8/r7/5RPP/6K1 b - -"]

1... Ra1+ {2:36m} 2. Rf1 {+2.08/4 0s} Rxf1+ {2s} 3. Kxf1
{+2.08/4 0s} Bh3 {1s} 4. gxh3 {+2.08/27 2s} *
[/puzzle]

SlowChess, even without use of TBs solves it easily. So, this knowledge can be trained, and specialized ending knowledge would probably improve drastically the results of the engine. Why not add TB positions to the training data, like Connor Macmonigle did?

Recapitulation.

Rebel's search misses tactical shots and not only probably alters drastically the style of the engine, but influences its results vastly. With a similar net, Rebel 14 was 100 Elos weaker than Marvin/Rebel. Solving the problems with the search will probably not yield so much Elos, but could well hugely improve the results.

Rebel still lacks technical, robust endgame knowledge. It is not only about bishops of different colors -- there, the situation is much better -- but it seems it is simply not trained on enough ending positions.

Berserk dev. version has recently added some simple endgame knowledge, by-passing the NNUE eval for these cases.

https://github.com/jhonnold/berserk/blob/main/src/endgame.c

It's all "known" ideas, usually found in HCE engines in one form or another, typically Stockfish and Ethereal, probably many others. Possibly overkill to use it with a NNUE engine, because (theoretically) the NNUE ought to know these things anyway.
I guess if one's engine screwed up an ending in a tournament, one might be tempted to program it in, or it might allow for slightly earlier termination when generating training games. Dunno. Anyway, I have no plans. I suspect Ed might have though.

Dio wrote:

it took a while until the games of Rebel 15x2 were evaluated, because our webmaster was very busy, the official first entry of Rebel 15x2 in the 40/4 CEGT list. More results will follow...

http://www.cegt.net/40_4_Ratinglist/40_4_single/262.html

http://www.cegt.net/40_4_Ratinglist/40_4_single/rangliste.html

Thanks for testing.

Much better than my own testing.

Rebel 14 - 3017
Rebel 15x2 - 3297

About 300 elo in 6 months, hard to believe, needed 10 years for that in the 80's and 90's.

in the CEGT 40/20 list Rebel 15x2 (3282 Elo, rank 165) is also very clearly ahead of Rebel 14 (3038 Elo, rank 412) = 244 Elo. http://www.cegt.net/40_40%20Rating%20List/40_40%20All%20Versions/rangliste.html