BoCC -- Beauty of Computer Chess

[Rebel]

New tool, extract from PGN (with or without comments) interesting king-attacks.

In the above example I used Million Base 3.45M, human games.

The output is controlled by the parameter King Attack Margin, a lower value than 100 will produce more games, a higher value less games. It's not exact science, just a matter of taste. An extreme low value of 10 might extract not so convincing games. The parameter is found in the king-attack.ini file.

Download :

[pohl4711]

Buggy: I filtered my Super 3 tournament games, got 170 games (said the tool). But it were 171 (1 win with black). Said ORDO. I trust Ordo more. FritzGUI also says 171 games.

Sadly, I have to say, I am still not convinced by this idea/algorithm. I looked into these found 171 games and already game 6 is not a king attack. Mostly the opposite is true here (nearly everything in this game happens on queenside of the board). Just having a simple pattern like queen and bishop on the a1/h8 diagonal (happens in this game) is not a king-attack. I can be. But here it is definitly not the case. And, if a tool only delivers 171 games out of a 19350 games gamebase, and games mostly played by aggressive playing engines (Patrica, Rebel Extreme, CS Tal 2.1, Revenge 1, Cerberus), these 171 games should really nail it... And that does not happen.

[pgn]
[Event "Super 3 tournament"]
[Site "?"]
[Date "2024.06.01"]
[Round "138"]
[White "Komodo 14.1 HCE"]
[Black "Lc0 791921 CPU"]
[Result "1-0"]
[ECO "A57"]
[WhiteFideId "-1"]
[WhiteFideId "-1"]
[PlyCount "101"]
[GameId "2204558239056107"]
[EventDate "2024.??.??"]

1. d4 {book} Nf6 {book} 2. c4 {book} c5 {book} 3. d5 {book} b5 {book} 4. cxb5 {book} g6 {book} 5. Nc3 {book} Bg7 {book} 6. e4 {book} d6 {book} 7. f4 {book} Nbd7 {book} 8. Nf3 {book} O-O {book} 9. a4 {1.07/28 69s} a6 {-0.48/11 34} 10. Be2 {0.87/29 66s} axb5 {-0.42/13 27} 11. Bxb5 {0.81/29 20} Ba6 {-0.41/12 10} 12. Bxa6 {1.11/28 11} Rxa6 {-0.44/12 9.0s} 13. O-O {0.90/28 63s} Qa8 {-0.43/11 16} 14. Ra3 {0.96/26 15} Nb6 {-0.42/12 25} 15. Qc2 {0.92/29 45} Nfd7 {-0.40/16 24} 16. b3 {0.92/30 14} c4 {-0.40/17 9.1s} 17. b4 {0.76/31 14} Bxc3 {-0.41/18 8.9s} 18. Qxc3 {0.99/31 9.8s} Rxa4 {-0.43/18 12} 19. Bb2 {1.05/30 11} Nf6 {-0.50/18 35} 20. Rfa1 {0.99/30 15} Rc8 {-0.56/18 58} 21. g4 {1.87/33 9.9s} Qa7 {-0.39/18 9.7s} 22. Kg2 {2.23/32 10} Rxa3 {-0.74/16 7.9s} 23. Rxa3 {2.65/32 13} Na4 {-1.07/15 9.4s} 24. Rxa4 {2.69/34 19} Qxa4 {-2.03/13 13} 25. g5 {3.09/32 14} h5 {-1.97/13 35} 26. gxf6 {6.58/27 7.7s} exf6 {-2.41/14 5.8s} 27. Qxf6 {7.43/30 11} c3 {-5.82/14 19} 28. Ng5 {7.65/29 10} Qd7 {-8.33/14 18} 29. Bxc3 {8.66/30 15} Rxc3 {-9.71/13 7.7s} 30. Qxc3 {9.18/29 21} Qg4+ {-10.18/12 7.9s} 31. Qg3 {9.84/30 18} Qe2+ {-10.33/10 17} 32. Kh3 {9.71/29 13} f6 {-8.53/9 37} 33. Ne6 {10.63/29 14} Kf7 {-7.50/12 23} 34. f5 {10.67/27 12} g5 {-8.19/13 17} 35. Nxg5+ {10.72/25 8.6s} fxg5 {-10.24/12 9.4s} 36. Qxg5 {10.57/24 8.9s} Qf3+ {-11.13/11 31} 37. Qg3 {13.88/28 17} Qf1+ {-11.25/11 24} 38. Kh4 {15.01/25 9.7s} Qe2 {-9.86/9 6.6s} 39. Qg6+ {15.86/24 12} Kf8 {-11.25/9 20} 40. Qxd6+ {19.67/22 6.9s} Kf7 {-14.44/8 16} 41. Qd7+ {25.88/25 8.1s} Kf8 {-8.99/9 17} 42. Kg5 {32.93/23 8.5s} Qxh2 {-7.18/9 14} 43. Qc8+ {#25/25 1.6s} Kf7 {-7.23/10 12} 44. Qe6+ {#23/26 1.6s} Kf8 {-8.59/1 0.001s} 45. f6 {#21/29 2.1s} Qc7 {-19.48/13 20} 46. d6 {#11/49 1.2s} Qf7 {-#12/6 1.2s} 47. Qxf7+ {#9/99 0.31s} Kxf7 {-117.70/1 0.001s} 48. d7 {#7/99 0.010s} Ke6 {-#6/4 0.13s} 49. d8=Q {#5/99 0.016s} Kf7 {-#4/1 0.001s} 50. Qe7+ {#3/99 0.007s} Kg8 {-#2/1 0.001s} 51. Qg7# {#1/99 0.006s, White mates} 1-0
[/pgn]

[Rebel]

Buggy: I filtered my Super 3 tournament games, got 170 games (said the tool). But it were 171 (1 win with black). Said ORDO. I trust Ordo more. FritzGUI also says 171 games.

Sadly, I have to say, I am still not convinced by this idea/algorithm. I looked into these found 171 games and already game 6 is not a king attack. Mostly the opposite is true here (nearly everything in this game happens on queenside of the board). Just having a simple pattern like queen and bishop on the a1/h8 diagonal (happens in this game) is not a king-attack. I can be. But here it is definitly not the case. And, if a tool only delivers 171 games out of a 19350 games gamebase, and games mostly played by aggressive playing engines (Patrica, Rebel Extreme, CS Tal 2.1, Revenge 1, Cerberus), these 171 games should really nail it... And that does not happen.

[pgn]
[Event "Super 3 tournament"]
[Site "?"]
[Date "2024.06.01"]
[Round "138"]
[White "Komodo 14.1 HCE"]
[Black "Lc0 791921 CPU"]
[Result "1-0"]
[ECO "A57"]
[WhiteFideId "-1"]
[WhiteFideId "-1"]
[PlyCount "101"]
[GameId "2204558239056107"]
[EventDate "2024.??.??"]

1. d4 {book} Nf6 {book} 2. c4 {book} c5 {book} 3. d5 {book} b5 {book} 4. cxb5 {book} g6 {book} 5. Nc3 {book} Bg7 {book} 6. e4 {book} d6 {book} 7. f4 {book} Nbd7 {book} 8. Nf3 {book} O-O {book} 9. a4 {1.07/28 69s} a6 {-0.48/11 34} 10. Be2 {0.87/29 66s} axb5 {-0.42/13 27} 11. Bxb5 {0.81/29 20} Ba6 {-0.41/12 10} 12. Bxa6 {1.11/28 11} Rxa6 {-0.44/12 9.0s} 13. O-O {0.90/28 63s} Qa8 {-0.43/11 16} 14. Ra3 {0.96/26 15} Nb6 {-0.42/12 25} 15. Qc2 {0.92/29 45} Nfd7 {-0.40/16 24} 16. b3 {0.92/30 14} c4 {-0.40/17 9.1s} 17. b4 {0.76/31 14} Bxc3 {-0.41/18 8.9s} 18. Qxc3 {0.99/31 9.8s} Rxa4 {-0.43/18 12} 19. Bb2 {1.05/30 11} Nf6 {-0.50/18 35} 20. Rfa1 {0.99/30 15} Rc8 {-0.56/18 58} 21. g4 {1.87/33 9.9s} Qa7 {-0.39/18 9.7s} 22. Kg2 {2.23/32 10} Rxa3 {-0.74/16 7.9s} 23. Rxa3 {2.65/32 13} Na4 {-1.07/15 9.4s} 24. Rxa4 {2.69/34 19} Qxa4 {-2.03/13 13} 25. g5 {3.09/32 14} h5 {-1.97/13 35} 26. gxf6 {6.58/27 7.7s} exf6 {-2.41/14 5.8s} 27. Qxf6 {7.43/30 11} c3 {-5.82/14 19} 28. Ng5 {7.65/29 10} Qd7 {-8.33/14 18} 29. Bxc3 {8.66/30 15} Rxc3 {-9.71/13 7.7s} 30. Qxc3 {9.18/29 21} Qg4+ {-10.18/12 7.9s} 31. Qg3 {9.84/30 18} Qe2+ {-10.33/10 17} 32. Kh3 {9.71/29 13} f6 {-8.53/9 37} 33. Ne6 {10.63/29 14} Kf7 {-7.50/12 23} 34. f5 {10.67/27 12} g5 {-8.19/13 17} 35. Nxg5+ {10.72/25 8.6s} fxg5 {-10.24/12 9.4s} 36. Qxg5 {10.57/24 8.9s} Qf3+ {-11.13/11 31} 37. Qg3 {13.88/28 17} Qf1+ {-11.25/11 24} 38. Kh4 {15.01/25 9.7s} Qe2 {-9.86/9 6.6s} 39. Qg6+ {15.86/24 12} Kf8 {-11.25/9 20} 40. Qxd6+ {19.67/22 6.9s} Kf7 {-14.44/8 16} 41. Qd7+ {25.88/25 8.1s} Kf8 {-8.99/9 17} 42. Kg5 {32.93/23 8.5s} Qxh2 {-7.18/9 14} 43. Qc8+ {#25/25 1.6s} Kf7 {-7.23/10 12} 44. Qe6+ {#23/26 1.6s} Kf8 {-8.59/1 0.001s} 45. f6 {#21/29 2.1s} Qc7 {-19.48/13 20} 46. d6 {#11/49 1.2s} Qf7 {-#12/6 1.2s} 47. Qxf7+ {#9/99 0.31s} Kxf7 {-117.70/1 0.001s} 48. d7 {#7/99 0.010s} Ke6 {-#6/4 0.13s} 49. d8=Q {#5/99 0.016s} Kf7 {-#4/1 0.001s} 50. Qe7+ {#3/99 0.007s} Kg8 {-#2/1 0.001s} 51. Qg7# {#1/99 0.006s, White mates} 1-0[/pgn]

We must have a different understanding about the definition of a king attack.

After 24.Rxa4 score 2.69
25.g5 score 3.09
26.gxf6 score 6.58
27.Qxf6 score 7.43

The game is over.

A classic king attack.

I am more concerned about the games the tool did not found, but should have.

There is work to do.

[pohl4711]

I am more concerned about the games the tool did not found, but should have.

There is work to do.

Indeed, 171 out of 19350 games is a very small margin. Only 0.88%. That is way too less to build a working scoring-system for a EAS-like-tool. Because if the number of input-games is "only" 1000 or so, only some few games (10 or less) will be found. Not all gamebases are that big.

[pohl4711]

We must have a different understanding about the definition of a king attack.

Indeed, for me a king-attack does not happen in or close before the endgame like in this game here. This king-attack was just a possible threat in the whole middlegame. But it did not happen. When it finally became a real threat, the game was already in the endgame.
Thats exactly what the core of the problem is, IMO: The simple HCE-eval sees a possible king-attack, but that does not mean, it really happens, because the eval of the playing engine can be totally different - perhaps it wants just to win the endgame and uses the threat of a king-attack just for some positional progress. Who knows? Your tool does not, my tool does not. So, IMHO it is way better not to use any evals. Especially not using a simple HCE-eval when deciding, if a superstrong neuralnet engine launches a king-attack or not. That will never work properly, I presume. A neuralnet evaluation is on another planet, compared to a simple HCE-eval. Thats why HCE evals are no longer used in engines.

But, for me, it seems legit, that an engine, which is a great king-attacking-player, should score very good in my sub-EAS-stat "Very short wins", because real king-attacks (launched right after the opening!) will lead to more wins in this category (mating the king with an attack launched right after the opening is a very short win...)

Here my Super 3 Tournament

Code: Select all

C: Very short wins (40 moves or less)                : [1]:04.00% CSTal 2.1 EAS  
                                                       [2]:02.43% Patricia 250510 a512  
                                                       [3]:01.92% Rebel Extreme 1.0  
                                                       [4]:01.68% Revenge 1.0 avx2  
                                                       [5]:01.66% Cerberus 21124081r81 

Clear lead for CSTal 2.1 EAS here.

Or take a look of the C-category of my full UHO-Ratinglist (1.8 million games). Same here:

Code: Select all

C: Very short wins (45 moves or less)                : [1]:11.05% CSTal 2.1 EAS  
                                                       [2]:09.00% Patricia 3.1 avx2  
                                                       [3]:07.97% Patricia 250510 a512  
                                                       [4]:07.32% Stockfish final HCE  
                                                       [5]:06.16% Slow Chess 2.9 avx2  

So, it is possible to find the best king-attacker without using any evals? Not sure about that, but it looks promising. This category alone seems to find the good king-attacking engines. To find the king-attack games, you should look only into these supershort wins (using a way smaller king-attack margin than 100) and search for king-attack patterns only in these games. This could work. Perhaps.

So, I would try this: Calculate the average length of all won games and sub -15 or -10 or -5 (experimental...). Then look only in games for a king-attack, which are that short, but do this with a smaller king-attack-margin to find more games.

[Rebel]

@pohl4711

Your judgement is way out of control, you haven't understood the difference between between the rating part (BoCC) and the pgn-king-attack-extraction tool.

The output is controlled by the parameter King Attack Margin, a lower value than 100 will produce more games, a higher value less games. It's not exact science, just a matter of taste. An extreme low value of 10 might extract not so convincing games. The parameter is found in the king-attack.ini file.

Just set the King Attack Margin = 10 and run again.

On BoCC however all games from 1 to infinity are calculated, the vast majority is below 100 and the value 100 is just for PGN output only in BoCC.

[Rebel]

UPDATE

New version, copy BoCC-1.1.exe in the right folders.

Changes:

1. Bug-fix in the king-attack evaluation when both players castled long, hardly has impact on ratings and pgn output.

2. Considerable change in the shorties evaluation, not counting book moves. Meaning the output is equal if games are played from EPD or from PGN. The result is a major shift in higher shorties evaluation scores and increased PGN output.

3. Because of the massive increased shorties evaluation scores the king-attack scores needed a simple factor update to bring the 2 evaluations back into balance.

4. Addition in the PGN output.
. [King "14736"]
. [Short "3200"]
Evaluation score of the game.
Meant for future PGN sorting and debugging.

All in all the numbers look quite different but it hardly has an effect on the order of the rating output.

Example of the 3578 elo pool -

Code: Select all

BoCC Rating-list  Sun Aug  3 11:01:43 2025

PGN   : pgn\bocc-3578.pgn
Games : 135000
Won   : 77849
Time  : 120 seconds

  BoCC   King          Short     
 Total  Attack   %     Games   %    Engine
364913  190656 25.6%  174257 61.5%  CSTal 2.1 EAS
265994  152287 23.1%  113707 54.4%  Rebel Extreme Dev
245999  131053 25.2%  114946 62.1%  Patricia 3.01
244358  114639 22.6%  129719 55.2%  Rebel Extreme
195111  142081 28.3%   53030 37.9%  Patricia 4 dev
190880  137846 27.3%   53034 36.0%  Patricia 4
181919   73977 20.3%  107942 53.9%  Rebel EAS 2.0
178248   91199 20.2%   87049 54.1%  Velvet 8.1 risky
156166   78829 18.4%   77337 54.5%  SF 17.1
101218   35700 08.2%   65518 48.6%  Titan
 90863   34304 07.6%   56559 46.8%  Obsidian130
 90775   32810 07.2%   57965 48.2%  Clover.8.0.2
 88934   30213 06.3%   58721 50.2%  viridithas 14.0.1
 87066   29060 05.9%   58006 47.5%  seer_v2.8
 86484   31565 07.3%   54919 47.5%  berserk 13
 85460   31599 06.8%   53861 45.5%  PlentyChess 2.1.0
 84809   30303 06.2%   54506 46.0%  caissa 1.20
 82938   29149 06.6%   53789 45.7%  Alexandria 7.0
 79982   27925 05.7%   52057 40.5%  Titan 1.1
 79669   32318 07.4%   47351 43.3%  Lizard 11_0

Forgot to add, ignore the nowadays silly virus warnings, nothing wrong with compiler output.

[Rebel]

Did a similarity test for the style engines, I was particulary interested in the similarity between the Rebel's and the Cstal one because they originate from the same source code of 2023 and only the neural nets are considerable different.

I do find the similarity of the beta version of Rebel-Extreme with Cstal-2.1-EAS a bit too high to my taste and so it won't see the daylight. However it is useful for my testing.