Winning after Qf6+!

[zullil]

A much discussed position:

[d]1r3r2/4bpkp/1qb1p1p1/3pP1P1/p1pP1Q2/PpP2N1R/1Pn1B2P/3RB2K w - - bm Qf6
Here we see Lc0 finding Qf6+ instantly. And the claim is made that its line is correct. But is it? The line given starts with Qf6+ Bxf6 gxf6+ Kg8 Ng5.

Here Stockfish seems to show that Bd2 (rather than Ng5) is indeed winning. But how does White win after Ng5 Nxa3?

[jp]

And could someone with the necessary GPU power please post what Lc0 says after it calculates for as long as possible? Does it stick to that line and evaluation, or does one or both of them change?

[Zenmastur]

I actually solved this once. The I accidently destroyed the data base the analysis was in along with the solutions to about 100 other such positions. I still have the DB but can't read it and haven't had time to try to repair it so it can be read

In the mean time I do remember that the line Lc0 gave has an error in it. Anyway I will solve it again when I have had a little more time.

Regards,

Zenmastur

[peter]

But how does White win after Ng5 Nxa3?

24.Ng5 only draws.

After that and Forward Backward of some according lines I had in my database already:

[d]1r3rk1/5p1p/1qb1pPp1/3pP1N1/p1pP4/PpP4R/1Pn1B2P/3RB2K b - - 0 1

Analysis by asmFishWCP_2019-10-11_popcnt:

24...Nxa3 25.Bf2 Nb1 26.Rxh7 Nxc3 27.Rg7+ Kh8 28.bxc3 b2 29.Kg2 a3 30.Rg1 a2 31.Kh3 b1B 32.Rh7+ Kg8 33.Be3 Bf5+ 34.Bg4 Rb7 35.Bxf5 exf5 36.Rg7+ Kh8 37.Rf1 a1Q 38.Rf4 Qf1+ 39.Rxf1 Ba4 40.Rf4 Bd1 41.Kg3 Qb1 42.Rh7+ Kg8 43.Rfh4 Bh5 44.Rg7+ Kh8 45.Rxh5+ gxh5 46.Rh7+ Kg8 47.Rg7+
= (0.00) Depth: 53/47 00:20:16 31415MN

Yet after 21...Kg7(?) with full hash from a slow Forward Backward:

[d]1r3r2/4bpkp/1qb1p1p1/3pP1P1/p1pP1Q2/PpP2N1R/1Pn1B2P/3RB2K w - - 0 1

Analysis by asmFishWCP_2019-10-11_popcnt:

22.Qf6+ Bxf6 23.gxf6+ Kg8 24.Bd2 Qa5 25.Bh6 Qxc3 26.bxc3 b2 27.Bxf8 Rxf8 28.Rb1 Rb8 29.Ng5 h6 30.Rxh6 Nxa3 31.Rh7 Be8 32.Rxb2 Rxb2 33.Bg4 Nc2 34.Rg7+ Kh8 35.Bxe6 fxe6 36.Ra7 Rb8 37.Nxe6 Ne1 38.h4 Nf3 39.Kg2 Nxh4+ 40.Kh3 Nf3 41.Kg3 Nd2 42.Ng5 Ne4+ 43.Nxe4 dxe4 44.d5 g5 45.e6 Bg6 46.d6 Bf5 47.e7 Kg8 48.e8Q+ Rxe8 49.f7+ Kf8 50.fxe8Q+ Kxe8 51.Rxa4 Kd7 52.Ra5 Be6 53.Rxg5 Kxd6 54.Kf4 Bh3 55.Kxe4 Bf1 56.Kd4 Kc6 57.Ra5 Kb6 58.Rc5 Be2 59.Re5 Bf1 60.Re6+ Kb5 61.Rf6 Bd3 62.Kd5 Ka4 63.Rb6
+- (2.70) Depth: 47/86 00:14:37 23997MN

[jp]

But how does White win after Ng5 Nxa3?

24.Ng5 only draws.

And what happens if Black replies to the apparently bad Ng5 with the apparently also bad h5?
Is White then winning again with Bxh5, and what is the main line?
Old SF can only find a draw at depth 46.

1r3rk1/5p2/1qb1pPp1/3pP1NB/p1pP4/PpP4R/1Pn4P/3RB2K b - -

[peter]

And what happens if Black replies to the apparently bad Ng5 with the apparently also bad h5?
Is White then winning again with Bxh5, and what is the main line?

22. Qf6+ Bxf6 23.gxf6+ Kg8

24. Ng5 (?) h5 (?) 25. Bxh5 Rb7 (25...Nxa3 (?!) 26. Bd2 Nb5 27. Rg1 +-) 26. Bf2 Nb4 27. Rf1 Nd3 28. Bg3 Bb5 29. Be2 Qxd4 30. cxd4 c3 31. bxc3 Nf4 +-

A line like this should be long enough and not too long to keep the +- eval in hash at backwarding, if hash is big enough and Backward has enough hardware- time.
Still you'll have to give the Fish some pondering- time after coming back to the dia- position to try alternative moves after 25.Bxh5 and maybe it will spare hardware- time there again to go into and out of the one or the other one side lide instead of stand alone computing, but the output should become a winning one sooner or later.

[d]1r3rk1/5p2/1qb1pPp1/3pP1NB/p1pP4/PpP4R/1Pn4P/3RB2K b - - 0 1

Analysis by asmFishWCP_2019-10-11_popcnt:

25...Tb7 26.Lf2 Sb4 27.Tf1 Sd3 28.Lg3 Dxd4 29.cxd4 c3 30.Le2 Lb5 31.bxc3 b2 32.Tb1 Sf2+ 33.Lxf2 Lxe2 34.Kg1 Lg4 35.Th7 Lf5 36.Tg7+ Kh8 37.Sxf7+ Tbxf7 38.Txf7 Txf7 39.Txb2 g5 40.Le3 Kh7 41.Lxg5 Kg6 42.Ld2 Kh5 43.Kf2 Kg4 44.Tb8 Lg6 45.Tg8 Kf5 46.Th8 Ke4 47.h4 Kd3 48.Le1 Kc2 49.h5 Lf5 50.Tg8 Kb3 51.h6 Kxa3 52.Tg7 Tf8 53.Tb7 Ka2 54.h7 a3 55.Ke3
+- (4.37) Tiefe: 40/72 00:14:15 22768MN

Which one's the main line with both sides blundering is difficult to say, but after 24.Ng5(?) h5(?) White's winning still, yet only because of Black's blunder 24...h5(?).
24.Ng5(?) Nxa3 is drawing.

[peter]

And could someone with the necessary GPU power please post what Lc0 says after it calculates for as long as possible? Does it stick to that line and evaluation, or does one or both of them change?

I borrowed a 2xRTX2080ti- installation using net 49921 and 8 x i9 CPU with 64 G hash here.

After 2 minutes output changes from 24.Ng5 (?) to 24.Bd2 (!), that's impressive:

[d]1r3r2/4bpkp/1qb1p1p1/3pP1P1/p1pP1Q2/PpP2N1R/1Pn1B2P/3RB2K w - - 0 1
Analysis by Lc0 v0.22.0 2x RTX2080 Ti syzygy:

Code: Select all

22.Df6+ Lxf6 23.gxf6+ Kg8 24.Sg5 Se3 25.Txh7 Sxd1 26.Tg7+ Kh8 27.Th7+ Kg8 28.Lxd1 Tb7 29.Tg7+ Kh8 30.Th7+ Kg8 31.Kg2 Tc8 32.Ld2 Dd8 
  +/-  (1.29)   Tiefe: 9/22   00:00:00
22.Df6+ Lxf6 23.gxf6+ Kg8 24.Sg5 Se3 25.Txh7 Sxd1 26.Tg7+ Kh8 27.Lxd1 Tb7 28.Ld2 Da7 29.Kg2 Le8 
  +/-  (1.34)   Tiefe: 10/22   00:00:01  3kN
22.Df6+ Lxf6 23.gxf6+ Kg8 24.Sg5 h6 25.Txh6 Se3 26.Td2 Tfc8 27.Th3 Sf5 28.Lg4 Le8 29.Th7 Tb7 30.Lxf5 
  +/-  (1.36)   Tiefe: 11/22   00:00:01  7kN
...
22.Df6+ Lxf6 23.gxf6+ Kg8 24.Sg5 Sxa3 25.Txh7 Sb5 26.Tg7+ Kh8 27.Th7+ Kg8 28.Tg7+ Kh8 29.Kg2 a3 30.Th7+ Kg8 31.h4 axb2 32.Tg7+ Kh8 33.Th7+ Kg8 34.h5 Sxd4 35.Tg7+ Kh8 36.Th7+ Kg8 37.Tg7+ Kh8 38.Lf2 Sxe2 39.hxg6 Sf4+ 40.Kf3 Dxf2+ 41.Kxf2 
  +/-  (0.79)   Tiefe: 21/55   00:02:09  4302kN
22.Df6+ Lxf6 23.gxf6+ Kg8 24.Ld2 Dc7 25.Sg5 Da7 26.Lc1 Sxd4 27.cxd4 h5 28.Lxh5 Lb5 29.Lg4 c3 30.Th7 Dc7 31.Tg7+ Kh8 32.Th7+ Kg8 33.Tg1 cxb2 34.Lxb2 Dc2 35.Tg3 Db1+ 36.Tg1 Dc2 37.Tg7+ Kh8 38.Tg3 Lf1 39.Th7+ Kg8 40.Tg7+ Kh8 
  +/-  (0.81)   Tiefe: 21/55   00:02:14  4571kN
...
22.Df6+ Lxf6 23.gxf6+ Kg8 24.Ld2 Dc7 25.Lh6 Sxd4 26.Txd4 Tfd8 27.Lg7 h5 28.Txh5 gxh5 29.Th4 Dxe5 30.Sxe5 d4+ 31.Sxc6 dxc3 32.Txh5 cxb2 33.Th8# 
  +/-  (1.53)   Tiefe: 19/57   00:05:41  18283kN
22.Df6+ Lxf6 23.gxf6+ Kg8 24.Ld2 Tfc8 25.Lh6 Sxd4 26.Txd4 Dc5 27.Kg1 Da7 28.Kf2 Le8 29.Sg5 Dxd4+ 30.cxd4 Tb4 31.Tc3 Tbb8 32.Ke1 Tc7 33.Kd2 Tcc8 34.h4 Tc7 
  +/-  (1.53)   Tiefe: 19/58   00:05:41  18291kN
...
22.Df6+ Lxf6 23.gxf6+ Kg8 24.Ld2 Tfc8 25.Lh6 Sxd4 26.Txd4 Dc5 27.Kg1 Da7 28.Kf2 Le8 29.Sg5 Dxd4+ 30.cxd4 Tb4 31.Tc3 Tbb8 32.Ke1 Tc7 33.Kd2 Tcc8 34.h4 Tc7 
  +/-  (1.52)   Tiefe: 20/69   00:05:58  19039kN
22.Df6+ Lxf6 23.gxf6+ Kg8 24.Ld2 Dd8 25.Tg1 Dc7 26.Sg5 Sxd4 27.cxd4 c3 28.Lxc3 Lb5 29.Lg4 Dc4 30.Sxh7 Tfe8 31.Sg5 Tec8 32.Sxf7 Kxf7 33.Th7+ Kg8 34.Th6 Kf7 35.Lh5 Ke8 36.Th7 Tc7 37.Lxg6+ Kd8 38.f7 Txf7 39.Txf7 Le8 40.La5+ Kc8 41.Ld3 Lxf7 42.Lxc4 dxc4 43.Tg7 Tb7 44.Kg2 Td7 45.Kf2 Lh5 46.Txd7 Kxd7 47.Ke3 Lg4 48.h4 Lh5 49.Le1 Le8 50.Kf4 Ke7 51.Kg5 Kf7 52.h5 Lc6 53.h6 
  +/-  (1.52)   Tiefe: 21/69   00:06:03  19307kN
...
22.Df6+ Lxf6 23.gxf6+ Kg8 24.Ld2 Dd8 25.Tg1 Dc7 26.Sg5 Sxd4 27.cxd4 c3 28.Lxc3 Lb5 29.Lg4 Dc4 30.Sxh7 Tfe8 31.Sg5 Tec8 32.Sxf7 Kxf7 33.Th7+ Kg8 34.Th6 Kf7 35.Lh5 Ke8 36.Th7 Tc7 37.Lxg6+ Kd8 38.f7 Txf7 39.Txf7 Le8 40.La5+ Kc8 41.Ld3 Lxf7 42.Lxc4 dxc4 43.Tg7 Tb7 44.Kg2 Td7 45.Kf2 Lh5 46.Txd7 Kxd7 47.Ke3 Lg4 48.h4 Lh5 49.Le1 Le8 50.Kf4 Ke7 51.Kg5 Kf7 52.h5 Lc6 53.h6 
  +/-  (1.48)   Tiefe: 21/69   00:06:48  23245kN
...
22.Df6+ Lxf6 23.gxf6+ Kg8 24.Ld2 Da5 25.Lh6 Dxc3 26.bxc3 b2 27.Lxf8 Txf8 28.Tb1 Tb8 29.Sg5 h5 30.Lxh5 Le8 31.Txb2 Txb2 32.Lxg6 Tb1+ 33.Kg2 fxg6 34.Th7 Se3+ 35.Kf3 Sf5 36.Tc7 Tg1 37.Sxe6 Tc1 38.Kg4 Tg1+ 39.Kf4 Tf1+ 40.Kg5 Tg1+ 41.Kf4 Tf1+ 42.Kg4 Tg1+ 43.Kf3 Tc1 44.Tc8 Txc3+ 45.Kf2 Tc2+ 46.Kg1 Tc1+ 47.Kg2 Tc2+ 48.Kg1 Tc1+ 49.Kf2 Tc2+ 50.Ke1 Tg2 51.Sc7 Kf7 52.Sxe8 c3 53.Sd6+ Sxd6 54.Tc7+ Ke6 55.Tc6 c2 56.Txd6+ 
+/-  (1.43)    Tiefe: 25/77   00:10:31  41046kN, tb=145

And after giving in 22.Qf6+ Bxf6 23.gxf6+ Kg8 24.Ng5 (?):

1r3rk1/5p1p/1qb1pPp1/3pP1N1/p1pP4/PpP4R/1Pn1B2P/3RB2K b - - 0 1

Analysis by Lc0 v0.22.0 2x RTX2080 Ti syzygy:

24...Sxa3 25.Txh7 Sb5 26.h4 Sd6 27.Tg7+ Kh8 28.Kg2 Se8 29.Th7+ Kg8 30.h5 Dd8 31.hxg6 fxg6 32.Lg4 Txf6 33.Ld2 Dd7 34.exf6 Sxf6 35.Txd7 Lxd7 36.Lxe6+ Lxe6 37.Sxe6 Ta8 38.Sc7 Ta7 39.Sb5 Ta5 40.Sa3 Se4 41.Lc1 Ta7 42.Tf1 Th7 43.Tf4 Te7 44.Tf3 Kg7 45.Tf1 Te8 46.Te1 Kf6 47.Sb5 Te7 48.Tf1+ Ke6 49.Tf8 g5 50.Ta8 Tf7 51.Ta6+
+/= (0.47) Tiefe: 31/68 00:02:24 11469kN, tb=1679

BTW Fat Fritz at 2xRTX2080 4 threads and 2G hash sticked to 24.Ng5(?) still after 10 minutes, stopped it then.

1r3r2/4bpkp/1qb1p1p1/3pP1P1/p1pP1Q2/PpP2N1R/1Pn1B2P/3RB2K w - - 0 1

Analysis by Fat Fritz v1.22.2:

22.Df6+ Lxf6 23.gxf6+ Kg8 24.Sg5 h5 25.Lxh5 Sxa3 26.Lf2 Tb7 27.Lf3 Sb5 28.Th7 Sd6 29.Tg7+ Kh8 30.Th7+ Kg8 31.Tg1 Se8 32.Tg4 Dd8 33.Tf4 a3 34.Le3 axb2 35.Tfh4 Dxf6 36.exf6 b1D+
+- (2.04) Tiefe: 17/40 00:00:07 204kN
...
22.Df6+ Lxf6 23.gxf6+ Kg8 24.Sg5 Sxa3 25.Txh7 Sb5 26.Tg7+ Kh8 27.h4 Sd6 28.Kg2 Dd8 29.Lg3 a3 30.h5 Se4 31.Tf1 axb2 32.Th7+ Kg8 33.hxg6 Sxf6 34.exf6 b1D 35.Tg7+ Kh8 36.Tf4 Dxf6 37.Th7+ Kg8 38.Txf6 Dc2 39.gxf7+ Txf7 40.Tfxf7 Dxe2+ 41.Kh3 Tb7 42.Thg7+ Kh8 43.Le5 De3+ 44.Kh2 De2+ 45.Kg3 De3+ 46.Kg4 De2+ 47.Kg3 De1+ 48.Kg4 De2+ 49.Kh4 De1+ 50.Kh5 Dd1+ 51.Kg6 Db1+ 52.Kh5 Dh1+ 53.Kg4 Dg1+ 54.Kh5 Dh1+ 55.Kg4 Dg1+ 56.Kh4 De1+ 57.Kh3 Dh1+ 58.Kg3 Dg1+ 59.Kh4
+/- (1.21) Tiefe: 41/86 00:10:29 21130kN, tb=166

27.h4 is still drawn, but 27...Nd6 (?) (...Nxd4 instead holds) loses, yet not as soon as after 28..Qd8( ?) (better 28...Ne8).

[jp]

Thank you, Peter! I'll put some of those lines in pgn so we can view them easily. Here is your asmFishWCP_2019-10-11_popcnt PV:

[pgn][FEN "1r3r2/4bpkp/1qb1p1p1/3pP1P1/p1pP1Q2/PpP2N1R/1Pn1B2P/3RB2K w - - 0 22"] 22.Qf6+ Bxf6 23.gxf6+ Kg8 24.Bd2 Qa5 25.Bh6 Qxc3 26.bxc3 b2 27.Bxf8 Rxf8 28.Rb1 Rb8 29.Ng5 h6 30.Rxh6 Nxa3 31.Rh7 Be8 32.Rxb2 Rxb2 33.Bg4 Nc2 34.Rg7+ Kh8 35.Bxe6 fxe6 36.Ra7 Rb8 37.Nxe6 Ne1 38.h4 Nf3 39.Kg2 Nxh4+ 40.Kh3 Nf3 41.Kg3 Nd2 42.Ng5 Ne4+ 43.Nxe4 dxe4 44.d5 g5 45.e6 Bg6 46.d6 Bf5 47.e7 Kg8 48.e8Q+ Rxe8 49.f7+ Kf8 50.fxe8Q+ Kxe8 51.Rxa4 Kd7 52.Ra5 Be6 53.Rxg5 Kxd6 54.Kf4 Bh3 55.Kxe4 Bf1 56.Kd4 Kc6 57.Ra5 Kb6 58.Rc5 Be2 59.Re5 Bf1 60.Re6+ Kb5 61.Rf6 Bd3 62.Kd5 Ka4 63.Rb6[/pgn]

[peter]

Thank you, Peter! I'll put some of those lines in pgn so we can view them easily.

You're welcome.
I have a rather big .pgn of my own stored since 2014 when this game was played between Lyudmil and Louis' SF of that time. Over the years lots of lines came in addition, now there's one more output- line from LC0 added. Even if no new insight within the first decisive moves are included, it's just for historical interest.

Obviously it's a typical position where NN- engines with big enough net and hardware succeed faster and better than SF, even if you have to consider, that Lyudmil did choose this special one rather selectively to show a certain weakness of SF's that way.

[zullil]

27.h4 is still drawn, but 27...Nd6 (?) (...Nxd4 instead holds) loses, yet not as soon as after 28..Qd8( ?) (better 28...Ne8).

Hi Peter,

Are you saying that White wins here?

[d]1r3rk1/5p1R/1qbnpPp1/3pP1N1/p1pP3P/1pP5/1P2B3/3RB2K w - - 1 6

Stockfish-dev isn't seeing it.