Piket vs Smirin 1993 : CB claims 31.Qxd7 is better than Qa4

[Eelco de Groot]

H6.03 Pro needs d=38 to prefer 31.Qxd7.

Code: Select all

Piket--Smirin, Biel 1993
r2r2q1/2Rn2bk/b2N2pp/pQ2p3/4Pp2/B4N1P/5PP1/2R3K1 w - - 2 31

Analysis by Houdini 6.03 Pro x64-popcnt (15 cores, Tactical=5, 8GB hash)

1. +- (3.50): 31.Qxd7 Rxd7 32.Rxd7 Be2 33.Rcc7 Bxf3 34.Bb2 Rd8 35.Re7 Rxd6 36.Bxe5 Rd1+ 37.Kh2 Be2 38.Rxg7+ Qxg7 39.Rxg7+ Kh8 40.Rxg6+ Kh7 41.Rg7+ Kh8 42.f3 a4 43.Ra7+ Kg8 44.Ra8+ Kf7 45.Rxa4 Rb1 46.Ra7+ Ke8 47.Bxf4 Rb6 48.h4 Ra6 49.Rh7 Kf8 50.Kg3 Kg8 51.Rd7 Bb5 52.Rc7 Be8 53.Kg4 Ra2 54.g3 Ra6 55.Be5 Bb5 56.Kf5 Rc6 57.Rb7 Ba6 58.Ra7 Kf8 59.g4 Ke8 60.g5 
2. +- (3.31): 31.Qa4 Qe6 32.Kh2 Nb6 33.Qxa5 Rxd6 34.Bxd6 Qxd6 35.R1c6 Nc4 36.Rxd6 Nxa5 37.Nxe5 Rc8 38.Rxg7+ Kxg7 39.Rxg6+ Kh7 40.Rxa6 Nb3 41.Ng4 Nd2 42.Nf6+ Kg6 43.h4 f3 44.Nd5+ Kg7 45.Ra7+ Kf8 46.gxf3 Nxf3+ 47.Kg3 Nd4 48.f4 Re8 49.e5 Re6 50.Kg4 Nc6 51.Rh7 Kg8 52.Rc7 Kf8 53.Nf6 Ne7 54.Ra7 Rc6 55.Ne4 Rc4 56.Nd6 Rc2 57.f5 
Depth: 38/89   02:47:46  293164 MN, tb=25245253

The bottom line is:

The move played was certainly brilliant but it was not the only winning move, and not necessarily the most accurate, though that hardly matters here. It certainly wasn't risk-free.

Is that after 2 hours, 47 minutes and 46 seconds Wow...

[MikeGL]

I do not understand why do you trust the position that you see in shredder GUI in multi-pv
It is known that programs do not find the best move in long pv.

I don't quite understand Uri, if Bf6 is better, White is more winning. But it is a waiting move, and after h5, planning Kh6 I think it still has to be Rxg7? I agree it is a nice move from a human point of view and that Stockfish thinks h5 weakens the position probably at least in KS calculations, and that Stockfish probably would not find the move it unless you first show it because it is Zugzwang for Black.

Sorry for an offtopic post. Maybe my question would merit a different thread on its own or
maybe even already discussed on old threads previously and heavily debated. If I understood
engines correctly, the eval score was taken from the last leaf of any PV. If the whole PV
cannot be trusted up to the end, then logically the eval score of that PV cannot be trusted
too, or is this understanding of mine incorrect?

Serious and honest question since I am only a newbie and not a serious engine programmer
and only understood some programming fundamentals with few actual programming
experience.

[Dirt]

Sorry for an offtopic post. Maybe my question would merit a different thread on its own or
maybe even already discussed on old threads previously and heavily debated. If I understood
engines correctly, the eval score was taken from the last leaf of any PV. If the whole PV
cannot be trusted up to the end, then logically the eval score of that PV cannot be trusted
too, or is this understanding of mine incorrect?

Nah, engines will happily tack on bad moves at the end of an analysis line. They can't tell it's bad, but their value for a few plies before is almost certainly nearly correct. Consider that an engine may allow a fork from a knight on the last move, but if it could see one ply deeper it would have many ways to avoid it.

[tpoppins]

H6.03 Pro needs d=38 to prefer 31.Qxd7.

Code: Select all

Piket--Smirin, Biel 1993
r2r2q1/2Rn2bk/b2N2pp/pQ2p3/4Pp2/B4N1P/5PP1/2R3K1 w - - 2 31

Analysis by Houdini 6.03 Pro x64-popcnt (15 cores, Tactical=5, 8GB hash)

1. +- (3.50): 31.Qxd7 Rxd7 32.Rxd7 Be2 33.Rcc7 Bxf3 34.Bb2 Rd8 35.Re7 Rxd6 36.Bxe5 Rd1+ 37.Kh2 Be2 38.Rxg7+ Qxg7 39.Rxg7+ Kh8 40.Rxg6+ Kh7 41.Rg7+ Kh8 42.f3 a4 43.Ra7+ Kg8 44.Ra8+ Kf7 45.Rxa4 Rb1 46.Ra7+ Ke8 47.Bxf4 Rb6 48.h4 Ra6 49.Rh7 Kf8 50.Kg3 Kg8 51.Rd7 Bb5 52.Rc7 Be8 53.Kg4 Ra2 54.g3 Ra6 55.Be5 Bb5 56.Kf5 Rc6 57.Rb7 Ba6 58.Ra7 Kf8 59.g4 Ke8 60.g5 
2. +- (3.31): 31.Qa4 Qe6 32.Kh2 Nb6 33.Qxa5 Rxd6 34.Bxd6 Qxd6 35.R1c6 Nc4 36.Rxd6 Nxa5 37.Nxe5 Rc8 38.Rxg7+ Kxg7 39.Rxg6+ Kh7 40.Rxa6 Nb3 41.Ng4 Nd2 42.Nf6+ Kg6 43.h4 f3 44.Nd5+ Kg7 45.Ra7+ Kf8 46.gxf3 Nxf3+ 47.Kg3 Nd4 48.f4 Re8 49.e5 Re6 50.Kg4 Nc6 51.Rh7 Kg8 52.Rc7 Kf8 53.Nf6 Ne7 54.Ra7 Rc6 55.Ne4 Rc4 56.Nd6 Rc2 57.f5 
Depth: 38/89   02:47:46  293164 MN, tb=25245253

[...]

Is that after 2 hours, 47 minutes and 46 seconds :) Wow...

Yes, Eeelco, it took that long to establish a certain level of certainty. Note that I didn't say "needs d=38 to find" -- 31.Qxd7 is at #1 already at d=22 after only 26 seconds. Also at d=24-28 and 33. The gap between it and 31.Qa4 never exceeded 40 cp, though.

[Eelco de Groot]

I don't quite understand Uri, if Bf6 is better, White is more winning. But it is a waiting move, and after h5, planning Kh6 I think it still has to be Rxg7? I agree it is a nice move from a human point of view and that Stockfish thinks h5 weakens the position probably at least in KS calculations, and that Stockfish probably would not find the move it unless you first show it because it is Zugzwang for Black.

After showing Kaissa Bf6:

[D]6q1/2R1R1bk/6pp/p3B3/4Pp2/7P/4bPPK/3r4 w - -

Engine: Kaissa HT (512 MB)
by T. Romstad, M. Costalba, J. Kiiski, G. Linscott

33 9:20 +5.97 38.Bf6 h5 39.Rxg7+ Qxg7 40.Rxg7+ Kh6
41.h4 Rd6 42.Rf7 Rxf6 43.Rxf6 a4
44.Rf7 a3 45.Ra7 a2 46.Rxa2 Bb5
47.e5 Kg7 48.Ra5 Be8 49.Ra6 Kf7
50.Rf6+ Ke7 51.f3 (3.404.519.023) 6077

33 9:20 +4.98 38.Rxg7+ Qxg7 39.Rxg7+ Kh8 40.Rxg6+ Kh7
41.Rg7+ Kh8 42.Ra7+ Kg8 43.Bxf4 Rd4
44.f3 Ra4 45.Bxh6 Ra2 46.Kg1 Bb5
47.Rg7+ Kh8 48.Rb7 Bc6 49.Ra7 a4
50.h4 Ra1+ 51.Kf2 (3.404.519.023) 6077

33 9:20 +4.98 38.f3 a4 39.Rxg7+ Qxg7 40.Rxg7+ Kh8
41.Rxg6+ Kh7 42.Rg7+ Kh8 43.Ra7+ Kg8
44.Rxa4 Rb1 45.Bxf4 Rb2 46.Kg1 Rb6
47.Be3 Rb1+ 48.Kf2 Bb5 49.Ra8+ Kg7
50.Rb8 h5 51.Kg3 (3.404.519.023) 6077

32 9:20 +3.99 38.Rb7 h5 39.Rxg7+ Qxg7 40.Rxg7+ Kh8
41.f3 a4 42.h4 Re1 43.Rxg6+ Kh7
44.Rg5 Bf1 45.Bxf4 a3 46.Rxh5+ Kg8
47.Ra5 Re2 48.Kg1 Bxg2 49.Rxa3 Bh3
50.Ra8+ Kh7 51.Ra7+ (3.404.519.023) 6077

I repeated it, stopped and started the search again without clearing the hash, Uri's Bf6 is really better even after deep search.


6q1/2R1R1bk/6pp/p3B3/4Pp2/7P/4bPPK/3r4 w - -

Engine: Kaissa HT (512 MB)
by T. Romstad, M. Costalba, J. Kiiski, G. Linscott

43 243:39 +7.12 38.Bf6 g5 39.Rxg7+ Qxg7 40.Rxg7+ Kh8
41.Rxg5+ Kh7 42.Rg7+ Kh8 43.e5 Rf1
44.e6 Bd1 45.Be5 Re1 46.Bc3 Re2
47.h4 h5 48.Rd7+ Kg8 49.Rxd1 Rxe6
50.Bxa5 Kf7 51.Rd5 (92.099.025.380) 6299

43 243:39 +5.99 38.Rxg7+ Qxg7 39.Rxg7+ Kh8 40.Rxg6+ Kh7
41.Rg7+ Kh8 42.Ra7+ Kg8 43.Bxf4 h5
44.Rxa5 Rb1 45.f3 Rb2 46.Kg1 Bb5
47.Ra8+ Kh7 48.Rb8 Rb1+ 49.Kh2 Rb3
50.Be5 Bc4 51.Rxb3 (92.099.025.380) 6299

43 243:39 +5.99 38.f3 h5 39.Rxg7+ Qxg7 40.Rxg7+ Kh8
41.Rxg6+ Kh7 42.Rg7+ Kh6 43.Ra7 a4
44.Rxa4 Kg6 45.Ra5 Rd8 46.Bxf4 Rh8
47.h4 Kf7 48.Kg3 Bd3 49.Rf5+ Ke7
50.Bg5+ Ke6 51.Rf6+ (92.099.025.380) 6299

43 243:39 +5.99 38.h4 a4 39.Rxg7+ Qxg7 40.Rxg7+ Kh8
41.Rxg6+ Kh7 42.Rg7+ Kh8 43.f3 Rd3
44.Ra7+ Kg8 45.Rxa4 Rb3 46.Ra8+ Kf7
47.Ra7+ Kg6 48.Bxf4 h5 49.Be5 Bd3
50.Re7 Rb5 51.Rg7+ (92.099.025.380) 6299

[pgn]

[Event "Biel Interzonal"]
[Site "Biel SUI"]
[Date "1993.07.22"]
[Round "6"]
[White "Jeroen Piket"]
[Black "Ilya Smirin"]
[WhiteElo "2590"]
[BlackElo "2640"]
[ECO "E94"]
[Result "1-0"]

1. d4 Nf6 2. c4 g6 3. Nc3 Bg7 4. e4 d6 5. Nf3 O-O 6. Be2 e5
7. O-O Na6 8. Be3 Ng4 9. Bg5 Qe8 10. dxe5 dxe5 11. h3 h6
12. Bd2 Nf6 13. Be3 Nd7 14. a3 f5 15. b4 c6 16. c5 f4
17. Bc1 Kh7 18. Bb2 Nc7 19. Nb1 Ne6 20. Nbd2 Nd4 21. Nc4
Nxe2+ 22. Qxe2 b6 23. Nd6 Qe6 24. Rfd1 bxc5 25. Rac1 cxb4
26. Rxc6 bxa3 27. Bxa3 Rd8 28. Rdc1 Qg8 29. Rc7 a5 30. Qb5
Ba6 31. Qxd7 Rxd7 32. Rxd7 Rd8 (32... Be2 33. Rcc7 Rd8
34. Re7 Bxf3 35. Bb2 Rxd6 36. Bxe5 Rd1+ 37. Kh2 Be2 {diagrammed position} 38. Bf6
h5) 33. Re7 Be2 34. Nxe5 Qa2 35. Rcc7 Rg8 36. Ne8 1-0[/pgn]

[Eelco de Groot]

H6.03 Pro needs d=38 to prefer 31.Qxd7.

Code: Select all

Piket--Smirin, Biel 1993
r2r2q1/2Rn2bk/b2N2pp/pQ2p3/4Pp2/B4N1P/5PP1/2R3K1 w - - 2 31

Analysis by Houdini 6.03 Pro x64-popcnt (15 cores, Tactical=5, 8GB hash)

1. +- (3.50): 31.Qxd7 Rxd7 32.Rxd7 Be2 33.Rcc7 Bxf3 34.Bb2 Rd8 35.Re7 Rxd6 36.Bxe5 Rd1+ 37.Kh2 Be2 38.Rxg7+ Qxg7 39.Rxg7+ Kh8 40.Rxg6+ Kh7 41.Rg7+ Kh8 42.f3 a4 43.Ra7+ Kg8 44.Ra8+ Kf7 45.Rxa4 Rb1 46.Ra7+ Ke8 47.Bxf4 Rb6 48.h4 Ra6 49.Rh7 Kf8 50.Kg3 Kg8 51.Rd7 Bb5 52.Rc7 Be8 53.Kg4 Ra2 54.g3 Ra6 55.Be5 Bb5 56.Kf5 Rc6 57.Rb7 Ba6 58.Ra7 Kf8 59.g4 Ke8 60.g5 
2. +- (3.31): 31.Qa4 Qe6 32.Kh2 Nb6 33.Qxa5 Rxd6 34.Bxd6 Qxd6 35.R1c6 Nc4 36.Rxd6 Nxa5 37.Nxe5 Rc8 38.Rxg7+ Kxg7 39.Rxg6+ Kh7 40.Rxa6 Nb3 41.Ng4 Nd2 42.Nf6+ Kg6 43.h4 f3 44.Nd5+ Kg7 45.Ra7+ Kf8 46.gxf3 Nxf3+ 47.Kg3 Nd4 48.f4 Re8 49.e5 Re6 50.Kg4 Nc6 51.Rh7 Kg8 52.Rc7 Kf8 53.Nf6 Ne7 54.Ra7 Rc6 55.Ne4 Rc4 56.Nd6 Rc2 57.f5 
Depth: 38/89   02:47:46  293164 MN, tb=25245253

[...]

Is that after 2 hours, 47 minutes and 46 seconds Wow...

Yes, Eeelco, it took that long to establish a certain level of certainty. Note that I didn't say "needs d=38 to find" -- 31.Qxd7 is at #1 already at d=22 after only 26 seconds. Also at d=24-28 and 33. The gap between it and 31.Qa4 never exceeded 40 cp, though.

I see, thanks for running it! So it seems from this Houdini does slightly less downscaling for opp. color bishops but not enough to clearly favor Qxd7. It can be noted that evaluation of 3.50 for Qd7 is very much what Vincent found for Qa4 with Stockfish. In this case Qd7 is clearly a win though, just look at the scores if you play 38. Bf6 But no engine would find Uri's move with nullmove on...

Crafty's choice is still the best. Never underestimate Crafty in an endgame, that was true already in Rebel times during the "Chess in 2000" test from Ed if my memory serves me right.

[Eelco de Groot]

34/25 3:53 +3.22-- 31.Qa4 Qe6 (2.363.730.117) 10125

34/53 5:18 +3.30++ 31.Qxd7 (3.264.332.799) 10239

Please, use the 2 best moves mode to compare the 2 eval.

Here are the results from Kaissa at depth 42 (best five moves).


r2r2q1/2Rn2bk/b2N2pp/pQ2p3/4Pp2/B4N1P/5PP1/2R3K1 w - -

Engine: Kaissa HT (512 MB)
by T. Romstad, M. Costalba, J. Kiiski, G. Linscott

42 519:38 +4.98 31.Qxd7 Rxd7 32.Rxd7 Be2 33.Rcc7 Rd8
34.Re7 Bxf3 35.Bb2 Rxd6 36.Bxe5 Rd1+
37.Kh2 Be2 38.Rxg7+ Qxg7 39.Rxg7+ Kh8
40.Rxg6+ Kh7 41.Rg7+ Kh8 42.Ra7+ Kg8
43.Bxf4 a4 44.Rxa4 (224.448.783.721) 7198

42 519:38 +4.03 31.Qa4 Qe6 32.Kh2 Nb6 33.Qxa5 Rxd6
34.Bxd6 Qxd6 35.R1c6 Nc4 36.Rxd6 Nxa5
37.Nxe5 Rc8 38.Rxg7+ Kxg7 39.Rxg6+ Kh7
40.Rxa6 Nb3 41.f3 Rc2 42.Rf6 Nd2
43.Rxf4 Nf1+ 44.Kg1 (224.448.783.721) 7198

42 519:38 +3.76 31.Qb1 Rab8 32.Qa1 Qe6 33.R1c6 Be2
34.Nf5 Qxc6 35.Rxc6 gxf5 36.Rc7 fxe4
37.Nh4 e3 38.fxe3 Bd3 39.Be7 Rb1+
40.Qxb1 Bxb1 41.Bxd8 Nf8 42.Bf6 Ne6
43.Rb7 fxe3 44.Bxe5 (224.448.783.721) 7198

42 519:38 +3.67 31.Qb2 Rab8 32.Qa1 Qe6 33.R1c6 Bd3
34.Kh2 Rb1 35.Qc3 Qb3 36.Nd2 Qxc3
37.Rxc3 Rd1 38.Rxd3 Nf6 39.Bb2 Kg8
40.Rxg7+ Kxg7 41.Bxe5 Rd7 42.Bc3 Re7
43.e5 Ne8 44.Bxa5 (224.448.783.721) 7198

42 519:38 +2.32 31.Qc6 Qe6 32.Qa4 Rab8 33.Qxa5 Be2
34.Nb7 Rg8 35.Qb4 Bxf3 36.R1c6 Qxc6
37.Rxc6 Bf8 38.Qa4 Rxb7 39.gxf3 Bxa3
40.Qxa3 Rg7 41.Qd6 Rb8 42.Rc7 Rd8
43.Qd5 Re7 44.h4 (224.448.783.721) 7198

[Vinvin]

Thanks to all for your insight !
Both moves seems close in strength.
And the next sentence on the Chessbase site in certainly wrong because it's not "clear" :

...1.Qxd7! ...still shows the clearly weaker move 1.Qa4.

[MikeGL]

Sorry for an offtopic post. Maybe my question would merit a different thread on its own or
maybe even already discussed on old threads previously and heavily debated. If I understood
engines correctly, the eval score was taken from the last leaf of any PV. If the whole PV
cannot be trusted up to the end, then logically the eval score of that PV cannot be trusted
too, or is this understanding of mine incorrect?

Nah, engines will happily tack on bad moves at the end of an analysis line. They can't tell it's bad, but their value for a few plies before is almost certainly nearly correct. Consider that an engine may allow a fork from a knight on the last move, but if it could see one ply deeper it would have many ways to avoid it.

Ah yes, thanks for your reply. I get it now.

regards,

[Paloma]

Here are my examinations
The best move switched around

Analysis by Houdini 6.03 x64-pext Std x64-pext:

1. +- (3.84): 31.Qxd7 Rxd7 32.Rxd7 Be2 33.Rcc7 Bxf3 34.Bb2 Rd8 35.Re7 Rxd6 36.Bxe5 Rd1+ 37.Kh2 Be2 38.Rxg7+ ... snip
2. +- (3.77): 31.Qa4 Qe6 32.Kh2 Nb6 33.Qxa5 Rxd6 34.Bxd6 Qxd6 35.R1c6 Nc4 36.Rxd6 Nxa5 37.Nxe5 Rc8 38.Rxg7+ ... snip

3 Core @3.8GHz, 8192MB Hash, Depth: 44 (2/60)

One Iteration further:

1. +- (3.92): 31.Qa4 Qe6 32.Kh2 Nb6 33.Qxa5 Rxd6 34.Bxd6 Qxd6 35.R1c6 Nc4 36.Rxd6 Nxa5 37.Nxe5 Rc8 38.Rxg7+ ... snip
2. +- (3.84): 31.Qxd7 Rxd7 32.Rxd7 Be2 33.Rcc7 Bxf3 34.Bb2 Rd8 35.Re7 Rxd6 36.Bxe5 Rd1+ 37.Kh2 Be2 38.Rxg7+ ... snip

3 Core @3.8GHz, 8192MB Hash, Depth: 45 (1/60)