Any engine out there can see gxf6 is a draw in this position?

[pocpit]

From a game David Antón vs Dimitrij Kollars

[fen]8/p5p1/Pp1k1n2/1Pp2BP1/2Pp1K2/3P4/8/8 w - - 0 1[/fen]

I've just tried SF at ply 33 (my hardware is really old) and the score goes way above +3.

[pocpit]

Ok, I answer myself; a Crystal version I have finds the draw in 18(!) plys.

Hmmm, it seems I can't copy/paste any analysis line from ScidVsPC :/

https://imgur.com/a/G1Nrtk8

[pocpit]

Ok, this is funny; I updated my version of Crystal and now it can't see the draw any more (?) (at least up to ply 22)

This is the update info, if it could be of use to anyone:

$ git pull
remote: Enumerating objects: 5139, done.
remote: Counting objects: 100% (1531/1531), done.
remote: Total 5139 (delta 1531), reused 1531 (delta 1531), pack-reused 3608
Receiving objects: 100% (5139/5139), 3.19 MiB | 1.51 MiB/s, done.
Resolving deltas: 100% (3968/3968), completed with 136 local objects.
From https://github.com/jhellis3/Stockfish
3a748915..054c2e7c crystal -> origin/crystal
* [new tag] v4.1 -> v4.1
* [new tag] v2.0 -> v2.0
* [new tag] v3.0 -> v3.0
* [new tag] v3.1 -> v3.1
* [new tag] v3.2 -> v3.2
* [new tag] v4.0 -> v4.0
Updating 3a748915..054c2e7c

[Ovyron]

Ok, this is funny; I updated my version of Crystal and now it can't see the draw any more (?) (at least up to ply 22)

It's not funny, many engines that were updated lost what made them special and now they're mostly equivalent to stockfish depth with a different name (at least you could see more of a difference by switching to a net like nn-0fd8a64be200.nnue).

[fishpov]

You are perfectly right and it's a pity. After the 3.2 version of Crystal as: 4.0, 4.1.
These last two versions have lost what made Crystal so special.

[jhellis3]

Thanks guys, glad to be of service.

[peter]

From a game David Antón vs Dimitrij Kollars

[fen]8/p5p1/Pp1k1n2/1Pp2BP1/2Pp1K2/3P4/8/8 w - - 0 1[/fen]

I've just tried SF at ply 33 (my hardware is really old) and the score goes way above +3.

Hi!
Modification of Crystal's called Crystallite from OpenChess- user dorsz

https://www.open-chess.org/viewtopic.php?p=25642#p25642

does much better here then the latest version's download of mine of the orginal:

8/p5p1/Pp1k1n2/1Pp2BP1/2Pp1K2/3P4/8/8 w - - 0 1

Analysis by Crystallite 1.1a-bmi2:

Code: Select all

1.gxf6 gxf6 2.Ke4 Kc7 3.Kd5 Kd8 4.Ke6 Kc7 5.Kxf6 
  +-  (4.32)   Depth: 7/11   00:00:00  41kN
1.gxf6 gxf6 2.Ke4 Ke7 3.Kd5 Kf7 4.Kd6 Ke8 5.Kc7 Kf8 
  +-  (4.46)   Depth: 8/12   00:00:00  63kN
1.gxf6 gxf6 2.Ke4 Ke7 3.Kd5 Kf8 4.Kd6 Kf7 5.Bd7 Kg8 6.Ke7 Kg7 
  +-  (4.63)   Depth: 9/15   00:00:00  126kN
1.gxf6 gxf6 2.Ke4 Kc7 3.Kd5 Kd8 4.Kc6 Ke7 5.Bc8 Kd8 6.Kb7 f5 7.Bxf5 Ke8 8.Kc7 Ke7 
  +-  (5.94)   Depth: 10/15   00:00:00  143kN
1.gxf6 gxf6 2.Ke4 Kc7 3.Kd5 Kb8 4.Kc6 Ka8 5.Kd7 Kb8 6.Kd8 Ka8 7.Bd7 Kb8 8.Bg4 f5 9.Bf3 f4 10.Ke8 Kc8 11.Be2 
  +-  (5.86)   Depth: 11/25   00:00:00  277kN
1.gxf6 gxf6 2.Ke4 Kc7 3.Kd5 Kb8 4.Kc6 Ka8 5.Kd6 Kb8 6.Bd7 f5 7.Be6 f4 8.Bg4 f3 9.Ke5 Kc7 10.Bh5 Kb8 
  +/-  (1.58)   Depth: 12/52   00:00:00  729kN
1.gxf6 gxf6 2.Ke4 Kc7 3.Kd5 Kb8 4.Kc6 Ka8 5.Kd6 Kb8 
  =  (0.00)   Depth: 13/52   00:00:00  12009kN
...
1.gxf6 gxf6 2.Ke4 Kc7 3.Kd5 Kb8 4.Kc6 Ka8 5.Kd6 Kb8 
  =  (0.00)   Depth: 58/30   00:00:56  4411MN, tb=24864139

Latest Crystal I have (Ipman's compile):

8/p5p1/Pp1k1n2/1Pp2BP1/2Pp1K2/3P4/8/8 w - - 0 1

Analysis by Crystal 040722:

Code: Select all

1.gxf6 gxf6 2.Ke4 Kc7 3.Kd5 Kb8 4.Ke6 Kc8 5.Kf7+ Kc7 
  +-  (3.43)   Depth: 7/13   00:00:00  94kN
...
1.gxf6 gxf6 2.Ke4 Kc7 3.Kd5 Kb8 4.Kd6 Ka8 5.Kd7 Kb8 6.Be6 f5 7.Bxf5 Ka8 8.Kd8 Kb8 9.Be6 Ka8 10.Bh3 Kb8 11.Kd7 Ka8 12.Kd6 Kb8 13.Bf5 Ka8 
  +-  (3.70)   Depth: 20/42   00:00:01  29789kN
1.gxf6 gxf6 2.Ke4 Kc7 3.Kd5 Kb8 4.Kc6 Ka8 5.Kd7 Kb8 6.Kd8 Ka8 7.Ke8 Kb8 8.Kd7 Ka8 9.Kd8 Kb8 10.Bd7 f5 11.Bc8 f4 12.Kd7 f3 13.Bb7 f2 14.Bg2 f1B 15.Bxf1 Ka8 16.Bh3 Kb8 17.Kd6 Ka8 18.Ke5 Kb8 19.Kf4 Ka8 
  +-  (3.47)   Depth: 21/42   00:00:01  56128kN
1.gxf6 gxf6 2.Ke4 Kc7 3.Kd5 Kb8 4.Kd6 Ka8 5.Kd7 Kb8 6.Ke7 Ka8 7.Kd8 Kb8 8.Kd7 Ka8 9.Bh3 Kb8 10.Bf1 f5 11.Kd8 f4 12.Bh3 f3 13.Kd7 f2 14.Bf1 Ka8 15.Bg2+ Kb8 16.Kd8 f1Q 
  +-  (3.86)   Depth: 22/42   00:00:02  85858kN
1.gxf6 gxf6 
  +-  (3.99 --)   Depth: 23/44   00:00:03  142MN
1.gxf6 gxf6 2.Ke4 Kc7 3.Kd5 Kb8 4.Kd6 Ka8 5.Ke7 Kb8 6.Kd7 Ka8 7.Kd8 Kb8 8.Bc8 f5 9.Bd7 f4 10.Be6 f3 11.Bh3 f2 12.Bg2 f1R 13.Bxf1 Ka8 14.Kd7 Kb8 15.Bh3 Ka8 16.Kd6 Kb8 17.Bg2 Kc8 18.Bf1 Kb8 19.Kd7 
  +-  (4.33)   Depth: 23/44   00:00:04  156MN
...
1.gxf6 gxf6 2.Ke4 Kc7 3.Kd5 Kb8 4.Kc6 Ka8 5.Kd6 Kb8 6.Kd7 Ka8 7.Bh3 Kb8 8.Bf1 f5 9.Kd8 f4 10.Kd7 f3 11.Bh3 f2 12.Bg2 f1Q 13.Bxf1 Ka8 14.Kd8 Kb8 15.Bh3 Ka8 16.Ke8 Kb8 17.Kd7 Ka8 18.Bf5 Kb8 19.Kc6 Ka8 20.Kc7 
  +-  (4.76)   Depth: 27/53   00:01:20  3610MN
...
1.gxf6 gxf6 2.Ke4 Kc7 3.Kd5 Kb8 4.Ke6 Kc8 5.Kf7+ Kb8 6.Ke7 Ka8 7.Ke6 Kb8 8.Kd7 Ka8 9.Ke7 Kb8 10.Kd8 Ka8 11.Bd7 Kb8 12.Be6 f5 13.Kd7 f4 14.Bg4 f3 15.Bh3 f2 16.Bf1 Ka8 17.Kd8 Kb8 18.Bg2 f1R 19.Bxf1 Ka8 20.Bh3 Kb8 21.Bc8 Ka8 22.Bd7 Kb8 
  +-  (4.67)   Depth: 32/56   00:02:18  6323MN
1.gxf6 
  +-  (4.65 ++)   Depth: 33/52   00:02:18  6326MN

Took about 80 moves forward to make the engine get 0.00- eval.

Crystal using this one net:
nn-3c0aa92af1da.nnue
Crystallite this one:
nn-793ebdeaf2e5.nnue
, being not able to use later ones.

Both engines with 30 threads of 16x3.5GHz- CPU and 32G hash, 6men Syzygys for both of them didn't make any difference (of course) as for ProbeLimit 6 or 0.

LC0 doesn't evaluate 0.00 but about that and also almost at once at 3070ti GPU:

8/p5p1/Pp1k1n2/1Pp2BP1/2Pp1K2/3P4/8/8 w - - 0 1

Analysis by Lc0 v0.30.0-dev+git.75c6ada:

Code: Select all

1.gxf6 gxf6 2.Ke4 Kc7 3.Kd5 Kb8 4.Ke6 Kc7 5.Ke7 Kb8 6.Kd6 Ka8 
  +/=  (0.67)   Depth: 7/15   00:00:02  0kN
...
1.gxf6 gxf6 2.Ke4 Kc7 3.Kd5 Kb8 4.Kc6 Ka8 5.Bh3 Kb8 6.Bd7 f5 7.Kd6 f4 8.Bg4 f3 9.Bxf3 Kc8 10.Bg4+ Kb8 11.Be6 Ka8 12.Bd5+ Kb8 13.Bb7 
  +/=  (0.50)   Depth: 10/26   00:00:03  51kN
...
1.gxf6 gxf6 2.Ke4 Kc7 3.Kd5 Kb8 4.Ke6 Ka8 5.Kd6 Kb8 6.Be6 f5 7.Ke7 f4 8.Kf6 f3 9.Kg5 f2 10.Bh3 f1Q 11.Bxf1 Kc7 12.Kf6 Kb8 13.Bg2 Kc7 14.Ke7 Kc8 15.Kd6 Kb8 16.Kc6 Kc8 17.Kd6 Kb8 18.Bh3 Ka8 19.Bd7 Kb8 20.Bc6 Kc8 21.Bb7+ 
  =  (0.23)   Depth: 15/51   00:00:15  1820kN
1.gxf6 gxf6 2.Bh7 Kd7 3.Kf5 Ke7 4.Kg6 Kd7 5.Kf7 Kd8 6.Be4 Kc7 7.Kg8 Kb8 8.Kf8 Kc7 9.Bh1 Kb8 10.Ke7 Kc7 11.Bg2 f5 12.Ke6 Kb8 13.Kf6 f4 14.Ke5 f3 15.Bxf3 Kc7 16.Kd5 Kb8 17.Kc6 Kc8 18.Kd6 Kb8 19.Be4 Kc8 20.Bf5+ Kb8 21.Be6 Ka8 22.Bd5+ Kb8 23.Kc6 Kc8 24.Be6+ Kb8 25.Kd5 
  =  (0.20)   Depth: 15/51   00:00:20  2581kN
...
1.gxf6 gxf6 2.Be4 Kc7 3.Kf5 Kb8 4.Kg6 Kc8 5.Kf7 Kb8 6.Ke8 Kc8 7.Ke7 Kb8 8.Kxf6 Kc8 9.Kg7 Kb8 10.Kf7 Kc7 11.Ke7 Kc8 12.Kd6 Kb8 13.Kc6 Kc8 14.Kd6 Kb8 15.Bd5 Kc8 16.Bb7+ Kb8 17.Bd5 Kc8 18.Kc6 Kb8 19.Be6 Ka8 20.Kc7 
  =  (0.16)   Depth: 16/60   00:00:41  5213kN, tb=8
1.gxf6 gxf6 2.Be4 Kc7 3.Kf5 Kb8 4.Kg6 Kc8 5.Kf7 Kb8 6.Ke8 Kc8 7.Ke7 Kb8 8.Kxf6 Kc8 9.Kg7 Kc7 10.Kf8 Kd8 11.Bf5 Kc7 12.Ke8 Kb8 13.Kd8 Ka8 14.Be4+ Kb8 15.Ke7 Kc7 16.Bd5 Kc8 17.Kd6 Kb8 18.Be6 Ka8 19.Bc8 Kb8 20.Bb7 
  =  (0.15)   Depth: 16/60   00:00:46  5773kN, tb=13
...
1.gxf6 gxf6 2.Be4 Ke6 3.Bd5+ Kd7 4.Kf5 Kd8 5.Ke6 Kc7 6.Ke7 Kc8 7.Kd6 Kb8 8.Be6 f5 9.Ke5 f4 10.Bh3 Ka8 11.Kf5 Kb8 12.Kg4 Kc7 13.Kh4 Kb8 14.Bf5 Kc7 15.Kg4 Kb8 
  =  (0.11)   Depth: 18/70   00:01:29  10303kN, tb=41

regards

[hgm]

This position is easy to recognize as a fortress: the Pawn chain is only vulnerable at a7 for attack by a King, and by keeping his King in the area a8-b8-c8-c7 black can cut off the approach of the white King to it. I suppose it is even more worthwile to recognize this pattern in evaluation than it is to recognize KBPK with the wrong Bishop.

Of course on-the-fly EGT generation would also solve this in a fraction of a second, as after the trade there are only 3 mobile non-pawns, and no way to force conversion other than capturing the g-Pawn. Which then itself leads to a 3-men P-slice.


[fen]2k5/p7/P7/4K3/4B3/8/8/8 w[/fen]
The basic pattern is like above; any black Pawn in a chain starting from a7 is invulnerable, and if any such Pawns are present white can have an extra Pawn blocking each as well. Extra passers for Black would also not invalidate the pattern. The information whether the Pawn structure satisfies this criterion can be stored as a flag in the Pawn hash. This flag would be tested in a special evaluation for the KBK + Pawns material combination (where KBPK is another special case).

[hgm]

Interesting... I put this pattern in KingSlayer, and indeed it avoids gxf6 even at d=1.

Code: Select all

dep	score	nodes	time	(not shown:  tbhits	knps	seldep)
 27	+1.57 	4.93M  	0:02.66	g5f6 g7f6 f5g4 d6c7 f4f5 c7d7 f5f6 d7d6 g4h3 d6c7 f6e7 c7b8 e7d8 b8a8 h3g4 a8b8 g4e6 b8a8 d8d7 a8b8 e6g4 b8a8 g4f3 a8b8 f3a8
 26	+1.57 	4.46M  	0:02.41	g5f6 g7f6 f5g4 d6c7 f4f5 c7d7 f5f6 d7d6 g4h3 d6c7 f6e7 c7b8 e7d8 b8a8 h3g4 a8b8 g4f5
 25	+1.59 	4.16M  	0:02.26	g5f6 g7f6 f5g4 d6c7 f4f5 c7d7 f5f6 d7d6 g4h3 d6c7 f6e7 c7b8 e7d8 b8a8 h3g2 a8b8 g2a8 b8a8 d8e8 a8b8 e8d7 b8a8
 24	+1.59 	3.92M  	0:02.13	g5f6 g7f6 f5g4 d6c7 f4f5 c7d7 f5f6 d7d6 g4h3 d6c7 f6e7 c7b8 e7d8 b8a8 h3g2 a8b8 g2a8 b8a8 d8e8 a8b8 e8d7 b8a8
 23	+0.04 	3.11M  	0:01.68	g5f6 g7f6 f4e4 d6c7 e4d5 c7b8 d5d6 b8a8 f5e6 a8b8 d6d7 f6f5 e6f5 b8a8 f5e6 a8b8 d7d6 b8a8 e6d7
 22	+0.04 	2.93M  	0:01.59	g5f6 g7f6 f4e4 d6c7 e4d5 c7b8 d5d6 b8a8 f5e6 a8b8 d6d7 f6f5 e6f5 b8a8 f5e6 a8b8 d7d6 b8a8 e6d7
 21	+0.04 	2.76M  	0:01.49	g5f6 g7f6 f4e4 d6c7 e4d5 c7b8 d5d6 b8a8 f5e6 a8b8 d6d7 f6f5 e6f5 b8a8 f5e6 a8b8 e6f7 b8a8 d7e6 a8b8
 20	+0.04 	1.86M  	0:01.02	g5f6 g7f6 f5g6 d6e6 g6h5 e6d6 f4f5 d6e7 h5g6 e7d7
 19	+0.05 	1.49M  	0:00.84	g5f6 g7f6 f5g6 d6e6 g6h5 e6d6 f4f5 d6e7 h5g6 e7d7 f5f6
 18	+0.05 	1.28M  	0:00.71	g5f6 g7f6 f5g6 d6e6 g6h5 e6d6 f4f5 d6e7 h5g6 e7d7
 17	+0.05 	1.10M  	0:00.63	g5f6 g7f6 f5g6 d6e6 g6h5 e6d6 f4f5 d6e7 h5g6 e7d7
 16	+0.05 	974300	0:00.56	g5f6 g7f6 f5g6 d6e6 g6h5 e6d6 f4f5 d6e7 h5g6 e7d7
 15	+0.06 	933913	0:00.54	g5f6 g7f6
 15	+0.05 	876882	0:00.51	f4f3 d6e5 f5g6 e5d6 g5f6 g7f6 f3f4 d6e6 g6e8 e6d6 f4f5 d6e7 e8g6 e7d6
 14	+0.08 	618450	0:00.39	f4f3 d6e5 f5g6 f6d7 f3g4 e5d6 g4h5 d7e5 g6h7 e5f3 h5g6 d6e5 g6g7 f3g5
 13	+0.41 	482431	0:00.31	f4f3 d6e5 f5g6 f6d7 f3g4 e5d6 g4h5 d7f6 h5h4 f6d7 g6h7 d7e5 h4h5
 12	+0.54 	413232	0:00.28	f4f3 d6e5 f5g6 e5e6 f3g3 f6d7 g6e4
 11	+0.54 	394379	0:00.28	f4f3 d6e5 f5g6 e5e6 f3g3 f6d7 g6e4
 11	+0.05 	173819	0:00.14	f5g6 f6d7 g6e4 d7e5 f4f5 e5f7 f5g6 f7g5 g6g5 d6e7
 10	+0.75 	86513  	0:00.07	f5g6 d6e6 g6e4 f6h5 f4g4 h5f6 g4f3 f6d7 f3f4 d7e5 e4f5 e6d6
  9	+0.75 	64716  	0:00.07	f5g6 d6e6 g6e4 f6h5 f4g4 h5f6 g4f3 f6d7 f3f4 d7e5 e4f5 e6d6
  8	+0.77 	43096  	0:00.06	f5g6 f6g8 g6h7
  8	+0.75 	33017  	0:00.04	f5e4 f6h5 f4g4 h5f6 g4f3
  7	+0.81 	18412  	0:00.03	f5e4 f6g8 e4h7 g8e7 h7f5 g7g6 f5e4
  6	+0.88 	11077  	0:00.03	f5e4 f6d7 f4g4 d6e7
  5	+0.79 	7866    0:00.01	f5e4 f6d7 e4d5 d7e5 f4e4
  5	+0.76 	4590    0:00.01	f5g6 d6e6 g6e4 f6h5 f4f3 h5f6
  5	+0.69 	2550    0:00.00	f4f3 d6e5 f5e4 f6e4 d3e4
  5	+0.08 	1833    0:00.00	g5f6 g7f6
  5	+0.07 	1631    0:00.00	g5g6 f6g8 f4f3 g8e7 f3f4 e7g6 f5g6
  4	+1.18 	513     0:00.00	g5g6 f6h5 f4f3 h5f6
  3	+1.18 	177     0:00.00	g5g6 f6h5 f4e4 h5f6 e4f3
  2	+1.28 	45      0:00.00	g5g6 d6e7
  1	+1.25 	23      0:00.00	g5g6

This is of course expected, as at d=1 QS will see the recapture that will create the position that matches the pattern. It takes 14 ply (0.39 sec) for it to see that it cannot protect the g-pawn from a Knight sacrifice that would create the pattern anyway.

Surprising thing is that the score bounces up again at 24 ply. Apparently white found a way there to force black taking the Bishop, to escape from the pattern. But even then, with each side 4 Pawns in a Pawn ending, 1.5 seems a gross misevaluation. This is probably coming from a bonus for having a Pawn on 6th rank, which is way too large, and also causes the inflated evaluation at low depth, when both minors and both g-Pawns remain present.

But of course it really should have recognize the pattern without the Bishop as a draw too.

[flok]

POS 1.20: g5g6

Feeks sticks at g5f6 (stopped after 148 seconds)

Micah as well (depth 25)

Embla also g5f6 (depth 28)