Revisiting Testposition

Discussion of anything and everything relating to chess playing software and machines.

Moderator: Ras

rj329
Posts: 18
Joined: Sat May 28, 2022 6:30 am
Full name: Randall Jackson

Revisiting Testposition

Post by rj329 »

[d]r2n1r1k/1pp5/3p1n2/P2pp2p/2P1P1p1/1Q2R1Pq/R2N1PN1/6K1 w
The winning continuation is 1. exd5 Nf7 2. c5 Ng5 3. Nh4 Rf7 4. c6!

This position is 2 moves into the previously known position, [d]r2n1r1k/1pp5/3ppn2/P2Pp2p/2P1P1pq/1Q2R2P/R2N1PP1/4N1K1 w
Even more impressive to get a winning evaluation from the start.
Top
peter
Posts: 3361
Joined: Sat Feb 16, 2008 7:38 am
Full name: Peter Martan

Re: Revisiting Testposition

Post by peter »

rj329 wrote: ↑Thu Mar 20, 2025 9:35 pm This position is 2 moves into the previously known position, [d]r2n1r1k/1pp5/3ppn2/P2Pp2p/2P1P1pq/1Q2R2P/R2N1PP1/4N1K1 w
Even more impressive to get a winning evaluation from the start:
Yet no problem anymore for SF dev. nor for Lc0, as for finding best continuations, "winning eval" is a matter of definition, hardware- time and depth to go into the lines, which aren't forced single ones only. The 2 moves you go into the original test position starting with 29th...Qxh3 e.g. aren't the ones, that were played in high level corr. chess way back then.
The game the positions originate from:
[pgn][Event "corr WC30/final"]
[Site "ICCF"]
[Date "2017.06.20"]
[Round "?"]
[White "Bauer, Robert"]
[Black "Zhak, Boris Mikhailovich"]
[Result "1-0"]
[ECO "C54"]
[WhiteElo "2558"]
[BlackElo "2528"]
[PlyCount "131"]
[GameId "320796609318"]
[EventDate "2017.??.??"]
[EventType "tourn (corr)"]
[SourceTitle "Fernschach2025"]
[SourceDate "2024.10.15"]
[SourceVersion "1"]
[SourceVersionDate "2024.10.15"]
[SourceQuality "1"]

1. e4 e5 2. Nf3 Nc6 3. Bc4 Nf6 4. d3 Bc5 5. O-O d6 6. c3 a6 7. a4 h6 8. Re1 Be6 9. Bxe6 fxe6 10. b4 Ba7 11. Be3 Bxe3 12. Rxe3 g5 13. Qb3 Qe7 14. Nbd2 Nd7 15. d4 Qf6 16. h3 h5 17. b5 axb5 18. Qxb5 Ra7 19. d5 Nd8 20. a5 g4 21. Ne1 Qg5 22. Nb3 Qg7 23. c4 O-O 24. Nd2 Nf6 25. Raa3 Ra8 26. Qb4 Qg5 27. Ra2 Kh8 28. Qb3 Qh4 29. g3 Qg5 30. h4 Qg6 31. Nd3 Qf7 32. Re1 Kg7 33. Rc1 Nd7 34. Rf1 Nf6 35. Nb4 Qg6 36. Qe3 Qh6 37. Qe2 Nd7 38. Rfa1 Qg6 39. Nd3 Ra6 40. Ra3 Qf6 41. Nb1 c6 42. Nc3 exd5 43. exd5 Qf3 44. Qd2 Qf5 45. Nb4 Ra7 46. Qe2 Nc5 47. Rd1 Qf6 48. Ra2 Qg6 49. a6 Kh7 50. Qe1 Rf7 51. Ra5 Rf8 52. Rxc5 dxc5 53. Nd3 Rxa6 54. Nxc5 Ra8 55. N3e4 cxd5 56. Rxd5 Nf7 57. Nd7 Rfe8 58. Ndf6+ Kg7 59. Nxe8+ Rxe8 60. Rb5 b6 61. Qe3 Re6 62. Ng5 Nxg5 63. hxg5 Kf7 64. c5 bxc5 65. Qxc5 Qxg5 66. Rb8 1-0
[/pgn]
White's 29th move g3! SF nowadays finds single threaded (3.5GHz) almost at once:

Code: Select all

r2n1r1k/1pp5/3ppn2/P2Pp2p/2P1P1pq/1Q2R2P/R2N1PP1/4N1K1 w - - 0 1

Analysis by Stockfish dev-20250202-d46c0b6f:

29.Sf1 exd5 30.cxd5 gxh3 31.Txh3 Dxe4 32.Se3 Tg8 33.Tc2 Sf7 34.Tc4 Dh7 35.Txc7 
  Weiß steht etwas besser.  
  +/= (0.33)  Tiefe: 10/18   00:00:00
29.Sf1 exd5 30.exd5 Sf7 31.hxg4 Dxg4 32.Tg3 
  Weiß steht etwas besser.  
  +/= (0.36)  Tiefe: 11/18   00:00:00
29.g3 Dxh3 30.Sg2 Sd7 31.Sh4 Sc5 32.Dd1 Kg8 33.Sb3 
  Weiß steht etwas besser.  
  +/= (0.70)  Tiefe: 12/19   00:00:00
29.g3 Dxh3 30.Sg2 Sh7 31.Sh4 Tf6 32.c5 Sg5 33.Db5 Sdf7 34.Dxb7 Tf8 
  Weiß steht etwas besser.  
  +/= (0.67)  Tiefe: 13/20   00:00:00  109kN
...
29.g3 Dg5 30.h4 Dg7 31.Sd3 Df7 32.Te1 Kg8 33.Tb1 Sd7 34.Tf1 Sf6 35.Sb4 Sd7 36.Taa1 c5 37.Sd3 Sf6 38.Kg2 Dh7 
  Weiß steht besser.  
  +/- (1.02)  Tiefe: 21/42   00:00:06  1990kN
29.g3 Dxh3 30.Sg2 exd5 31.exd5 Sf7 32.Sh4 Sg5 33.c5 Tg8 34.Ta4 Tab8 35.Dd3 Tg7 36.Sg6+ Kg8 37.Df5 Sfh7 38.cxd6 cxd6 39.Se4 Sxe4 40.Taxe4 Tf7 
  Weiß steht besser.  
  +/- (1.06)  Tiefe: 22/43   00:00:09  3236kN
...
29.g3 Dxh3 30.Sg2 exd5 31.cxd5 Sf7 32.Sh4 Sg5 33.Dxb7 Tg8 34.Db1 Sf3+ 35.Shxf3 gxf3 36.Txf3 Tg6 37.a6 h4 38.Df1 Dg4 39.Dg2 Sh7 40.Tfa3 Sg5 41.Sf3 hxg3 42.Sxg5 Dd1+ 43.Df1 gxf2+ 44.Txf2 Txg5+ 45.Kh2 Dxf1 46.Txf1 Kg7 47.Tc1 Ta7 48.Tb1 Th5+ 49.Kg1 Th8 50.Tb7 Txb7 51.axb7 Tb8 52.Ta7 Kf6 
  Weiß steht besser.  
  +/- (1.02 ++)  Tiefe: 29/61   00:01:02  22579kN
29.g3 Dg5 30.h4 Dg7 31.Sd3 Df7 32.Te1 Sd7 33.Tf1 Kg8 34.Sb4 Dg6 35.De3 Kg7 36.Sd3 Sf6 37.De2 Kg8 38.Tb1 Dh6 39.Tba1 Sd7 40.c5 Sxc5 41.Sxc5 dxc5 42.Tb1 
  Weiß steht besser.  
  +/- (1.14)  Tiefe: 29/61   00:01:07  24437kN
29.g3 Dxh3 
  Weiß steht besser.  
  +/- (1.01 --)  Tiefe: 30/48   00:01:11  26007kN
...
29.g3 Dxh3 
  Weiß steht besser.  
  +/- (0.96 ++)  Tiefe: 30/59   00:01:20  29172kN
And Lc0 with 2 threads, a 3070ti GPU and the 6147500PT- net:

Code: Select all

r2n1r1k/1pp5/3ppn2/P2Pp2p/2P1P1pq/1Q2R2P/R2N1PP1/4N1K1 w - - 0 1

Analysis by Lc0 v0.31.2:

29.Sf1 gxh3 
  Weiß steht etwas besser.  
  +/= (0.33)  Tiefe: 1/2   00:00:05  2Knoten
29.Sf1 gxh3 30.Txh3 
  Weiß steht etwas besser.  
  +/= (0.34)  Tiefe: 2/3   00:00:05  4Knoten
29.g3 Dxh3 30.Sg2 
  Weiß steht besser.  
  +/- (0.90)  Tiefe: 2/3   00:00:05  8Knoten
29.g3 Dg5 30.h4 Dg8 
  Weiß steht besser.  
  +/- (0.84)  Tiefe: 2/4   00:00:05  9Knoten
29.g3 Dxh3 30.Sg2 Sh7 31.Sh4 
  Weiß steht besser.  
  +/- (0.94)  Tiefe: 3/5   00:00:05  23Knoten
...
29.g3 Dxh3 30.Sg2 Sh7 31.Sh4 Tf6 32.c5 
  Weiß steht besser.  
  +/- (0.93)  Tiefe: 4/7   00:00:05  44Knoten
29.g3 Dg5 30.h4 Dg8 31.Sd3 exd5 32.cxd5 c6 
  Weiß steht besser.  
  +/- (1.01)  Tiefe: 4/8   00:00:05  64Knoten
...
29.g3 Dg5 30.h4 Dg8 31.Sd3 Df7 32.Te1 Kg8 33.Ta3 exd5 
  Weiß steht besser.  
  +/- (1.01)  Tiefe: 7/11   00:00:05  350Knoten
29.g3 Dxh3 30.Sg2 Sd7 31.Db5 Sc5 32.Sb3 Sxb3 33.Txb3 Sf7 34.dxe6 Sg5 
  Weiß steht besser.  
  +/- (0.99)  Tiefe: 7/12   00:00:05  418Knoten
...
29.g3 Dxh3 30.Sg2 exd5 31.cxd5 Sf7 32.Sh4 Sg5 33.Dxb7 Tg8 34.Db1 Sf3+ 35.Shxf3 gxf3 36.Txf3 Tg6 37.Df1 Dg4 38.De2 h4 39.a6 Sh5 40.Te3 Dxe2 41.Txe2 Ta7 42.Te3 Tg8 43.Tb3 Tga8 44.gxh4 Txa6 45.Txa6 Txa6 46.Tc3 Ta2 47.Sf1 Ta4 48.Sg3 Sxg3 49.fxg3 Txe4 50.Txc7 Td4 51.Kg2 Txd5 52.g4 Td4 53.Kf3 d5 54.h5 
  Weiß steht besser.  
  +/- (1.01)  Tiefe: 18/52   00:00:38  136kN
Peter.
Top
rj329
Posts: 18
Joined: Sat May 28, 2022 6:30 am
Full name: Randall Jackson

Re: Revisiting Testposition

Post by rj329 »

Yes, Stockfish and Leela find the initial move g3 almost instantly, but the evaluation stagnates and enters more drawish territory since they don't find the best continuation there. The line by Leela has 3. cxd5 which is drawish, and Stockfish doesn't find the correct line quickly either.

The winning continuation with ~1.5 evaluation (almost twice that of the next best line) is difficult to find with a good amount of only moves on both sides.
Top
peter
Posts: 3361
Joined: Sat Feb 16, 2008 7:38 am
Full name: Peter Martan

Re: Revisiting Testposition

Post by peter »

rj329 wrote: ↑Fri Mar 21, 2025 2:18 am Yes, Stockfish and Leela find the initial move g3 almost instantly, but the evaluation stagnates and enters more drawish territory since they don't find the best continuation there. The line by Leela has 3. cxd5 which is drawish, and Stockfish doesn't find the correct line quickly either.
You saw the output- lines were from STC- pondering, started with empty hash (NN-cache) didn't you? The one from SF is single threaded, as I wrote, because I wanted to make it reproducible, Lc0's is very short as for the finally pasted one (3 plies only, the very last one of SF's has only 2).
SF's latest a little longer output- line at depth 29 is suboptimal at 31.(your first) cxd5?!, that's true.
After your suggested
1. exd5 Nf7 2. c5 Ng5 3. Nh4 Rf7 4. c6!
SF after some Forward- Backward with 30 threads of 16x3.5GHz CPU at the end of that starting with 31 instead of 1:

[d]r6k/1pp2r2/2Pp1n2/P2Pp1np/6pN/1Q2R1Pq/R2N1P2/6K1 b - - 0 1

Analysis by Stockfish dev-20250202-d46c0b6f:

34...bxc6 35.dxc6 Taf8 36.Tb2 Kg7 37.Dd3 e4 38.Dd4 Sf3+ 39.Sdxf3 gxf3 40.Txe4 Kg8 41.Tb5 Sxe4 42.Dxe4 Dg4 43.Db1 Tg7 44.Kh2 De6 45.Txh5 d5 46.Dd1 Dxc6 47.a6 De6 48.a7 c6 49.Da1 Ta8 50.Sf5 Taxa7 51.Db2 Tab7 52.Sxg7 Txg7 53.Db8+ Kf7 54.Th4 Dg6 55.Db7+ Kg8 56.Dc8+ Kf7 57.Tf4+ Ke7 58.Txf3 Th7+ 59.Kg2 Kd6 60.Tf4 Te7 61.Db8+ Ke6 62.Df8 Tb7
Tiefe: 40/54 00:05:31 4430MN
Weiß steht deutlich besser.
+- (2.02 --)

So, yes, a somwhat winning eval, if you take e.g. cb- definitions for +-, but at the end of the output- line it's even a "bingo- eval" of >199, so what?

Anyhow, the latest pasted SF- line of the single thread STC- pondering didn't include the blunder 31.cxd anymore neither, it was a fail high then before that, so let's see, how SF does SMP (on same hardware as above) started with empty hash (restart of GUI) again at the position after 30....exd5:

[d]r2n1r1k/1pp5/3p1n2/P2pp2p/2P1P1p1/1Q2R1Pq/R2N1PN1/6K1 w - - 0 1

Code: Select all

Analysis by Stockfish dev-20250202-d46c0b6f:

31.exd5 Sf7 32.Db1 Sg5 33.Sh4 Tg8 34.Df5 Taf8 35.Sg6+ Txg6 36.Dxg6 Sfh7 37.Taa3 Tg8 
  Weiß steht besser.  
  +/- (0.90)  Tiefe: 10/31   00:00:00
...
31.exd5 Sf7 32.Sh4 Sg5 33.Db1 Sf3+ 34.Sdxf3 gxf3 35.Txf3 Sg4 36.Dg6 Txf3 37.Dxh5+ Kg8 38.Dg6+ Kh8 39.Sxf3 Tf8 40.De4 Sf6 41.De2 Sg4 42.Tb2 Tf6 43.De4 Kg8 44.c5 Th6 45.Sh4 Dh2+ 46.Kf1 
  Weiß steht besser.  
  +/- (1.03)  Tiefe: 17/47   00:00:01  16596kN
31.cxd5 Sf7 32.Sh4 Sg5 33.Dxb7 Tab8 34.a6 Kg8 35.Tb3 Txb7 36.axb7 Sd7 37.Ta8 Sb8 38.Tc3 Sf3+ 39.Sdxf3 gxf3 40.Txf3 Td8 41.Sf5 Kh8 42.Tc3 Dg4 43.f3 Dg5 44.Txc7 Dd2 45.Tc8 Kh7 46.Txd8 De1+ 47.Kg2 
  Weiß steht besser.  
  +/- (1.08)  Tiefe: 18/58   00:00:01  21412kN
31.exd5 Sf7 32.Db1 Sg5 33.Sh4 Sf3+ 34.Sdxf3 gxf3 35.Txf3 Sg4 36.Dg6 Txf3 37.Dxh5+ Kg8 38.Dg6+ Kh8 39.Sxf3 Tf8 40.Ta3 Tg8 41.Df5 e4 42.Dxe4 Dh1+ 43.Kxh1 Sxf2+ 44.Kg2 Sxe4 45.Sd4 Kg7 46.Se6+ Kf7 
  Weiß steht besser.  
  +/- (0.99)  Tiefe: 19/58   00:00:02  33467kN
...
31.exd5 Sf7 32.Sh4 Sg5 33.c5 Tf7 34.Dxb7 Tg8 35.Ta4 Sf3+ 36.Sdxf3 gxf3 37.Txf3 Se4 38.Txe4 Txf3 39.Sxf3 Txg3+ 40.fxg3 Dxg3+ 41.Kh1 Dxf3+ 42.Kh2 Df2+ 43.Kh3 Df3+ 44.Kh4 Dxe4+ 45.Kg5 De1 
  Weiß steht deutlich besser.  
  +/- (1.59 ++)  Tiefe: 32/85   00:02:33  2017MN
31.exd5 Sf7 32.Sh4 Sg5 33.c5 Tf7 34.c6 bxc6 35.dxc6 Taf8 36.Tb2 Kg7 37.Dd3 e4 38.Dd4 Kg8 39.a6 Sf3+ 40.Sdxf3 gxf3 41.Txe4 Sg4 42.Txg4+ Dxg4 43.Dd5 Da4 44.Ta2 Db4 45.De6 Ta8 46.Ta1 Dc5 47.Te1 Dd4 48.Te3 Df6 49.Da2 Kf8 50.Da5 Kg8 51.Te4 
  Weiß steht deutlich besser.  
  +/- (1.55)  Tiefe: 32/85   00:02:34  2037MN
...
31.exd5 Sf7 32.c5 Sg5 33.Sh4 Tf7 34.c6 bxc6 35.dxc6 Taf8 36.Tb2 Kg7 37.Dd3 e4 38.Dd4 Sf3+ 39.Sdxf3 gxf3 40.Txe4 Kg8 41.a6 Sg4 42.Txg4+ Dxg4 43.Dd5 Da4 44.Ta2 Db4 45.Kh2 Dc5 46.De6 Ta8 47.Td2 Kf8 48.Dh6+ Kg8 49.Dg6+ Kf8 50.De6 Da5 51.Td3 Da1 52.Td5 Tf6 53.Tf5 Txf5 54.Dxf5+ Kg7 55.Dg6+ Kf8 56.Dxh5 De1 
  Weiß steht deutlich besser.  
  +- (1.70 ++)  Tiefe: 36/19   00:04:30  3590MN
The weaker move 31.cxd5? is pondered about only from depth 18 to depth 19, you see?
From depth 32 the crucial 34.c6! is in output, at the end of the latest pasted line, SF gives #20, if you let it jump there, of course not believing all these many output- plies were best moves, just to show the difference between an eval only and a longer one output- line together with it one more time.
Robert Bauer won Olympiad with his team following the ICCF WC30 final round, within which the game was played then, so not winning with this really beautiful move 29.g3! in less then 65 moves in such a tournament is at least a hint, the game wasn't as easily won, as you seem to suggest with your short line. And that B.M.Zhak prefered 29...Qg5 before your 29...Qxh3?! might have had it's causes too, mightn't it?
:)
Peter.
Top

Return to “Computer Chess Club: General Topics”