Stockfish get too deep and see nothing

[peter]

I wonder if there are chess programs who can solve the starting position let say at 3 minutes per move not in the meaning of finding one best move but in the meaning of beating every opponent with white.

I also wonder if you can define some simple algorithm for humans how to win with white so human who follow it can win with white without computer help or at least get a position that is a simple win based on engine-engine game for every engine with rating of at least 3000.

I'd bet, LC0 wins that with White against any opponent of your choice, if length of game doesn't matter and you give it enough hardware- time.
That's started with empty hash (which doesn't matter as for LC0 anyhow, NN- cache was empty too, which doesn't really matter neither):

2k5/2p5/1q1p4/pPpPp1pp/N1P1Pp2/P4PbP/KQ4P1/8 w - - 0 1

Analysis by Lc0 v0.28.2: (net 780486, NVIDIA 3070ti, 8G NN-chache)

46.Kb3 Kb7 47.Kc2 h4 48.Kd3 Lh2 49.Ke2 Lg3 50.Kf1 Kc8 51.Db3 Kb7 52.Kg1 Da7 53.b6 Da6
+/- (1.13) Tiefe: 11/29 00:00:19 280kN
...
46.Kb3 Kb7 47.Kc2 h4 48.Kd3 Bh2 49.Ke2 Bg3 50.Kf1 Kc8 51.Qc2 Kb7 52.Qd1 Qa7 53.Qb3 Qb6 54.Kg1 Qa7 55.b6 Qa6 56.bxc7+ Kxc7 57.Kf1
+/- (1.07) Depth: 15/46 00:03:02 9371kN

Of course you must not care about eval.

And the "algorithms" for humans (which you see in the line above too, it's just Black's moves, that aren't the best ones) I described often enough already in this thread

forum3/viewtopic.php?p=918315#p918315

forum3/viewtopic.php?p=919226#p919226

forum3/viewtopic.php?p=919326#p919326

About Crystal (with or without NNUE) I'd be rather sure it wouldn't go over 50 moves- boundary nor into 3fold- traps neither in LTC, you'd just have to give it enough hardware- time to overcome the first and only possible crucial blunder (if one really wants to call it like that) 46.Nxb6+ (?) regards...

Edit: Game LC0- SF is already running, SF 220122 at 30 threads with 45'+5" TC has eval of about +6 just after LC0's first move 46.Kb3, so what?

[Milos]

Yes Qb6 is always there, problem is with extra tempo prevented by cutting out Bh2 option a4 eventually gets pushed since just moving king (and queen) can not prevent it. Once a4 is pushed it's just a matter of time.

59...Qb6 is the only move to stop 60.b6 now.
After 59...Qb6 I can already give 60.g4 as premove, so let me know the other day, which way you go regards

As I already mentioned couple of times 59...Qb6 is the only option coz I can't allow 60.b6
However, I definitively cannot prevent a4 coming, and this is also certain loss for black with perfect play by white. Still, I am gonna play couple more moves just to see which path do you select towards the win.
So my moves: 59... Qb6 60.g4 Bg3

[fen]8/1kp5/1q1p4/pPpPp1p1/N1P1PpPp/PQ3PbP/8/5K2 w - - 1 61 [/fen]

[Milos]

I wonder if there are chess programs who can solve the starting position let say at 3 minutes per move not in the meaning of finding one best move but in the meaning of beating every opponent with white.

I also wonder if you can define some simple algorithm for humans how to win with white so human who follow it can win with white without computer help or at least get a position that is a simple win based on engine-engine game for every engine with rating of at least 3000.

For A/B engines I'm certain there is none that can win even with 1h per move. Lc0 maybe, but again I'm really skeptical.
For humans if they can do a quick check with a helper engine on the side it's rather easy. Without engine, probably not so easy.

[Milos]

I'd bet, LC0 wins that with White against any opponent of your choice, if length of game doesn't matter and you give it enough hardware- time.
That's started with empty hash (which doesn't matter as for LC0 anyhow, NN- cache was empty too, which doesn't really matter neither):

The question is whether Lc0 can win against perfect black defense or against engine assisted human that understands the position and all defense strategies. Not against another engine. And this I strongly doubt.

[peter]

I wonder if there are chess programs who can solve the starting position let say at 3 minutes per move not in the meaning of finding one best move but in the meaning of beating every opponent with white.

I also wonder if you can define some simple algorithm for humans how to win with white so human who follow it can win with white without computer help or at least get a position that is a simple win based on engine-engine game for every engine with rating of at least 3000.

For A/B engines I'm certain there is none that can win even with 1h per move. Lc0 maybe, but again I'm really skeptical.
For humans if they can do a quick check with a helper engine on the side it's rather easy. Without engine, probably not so easy.

As for the game LC0- SF, pity LC0 did let Black's Bishop come to first rank yet and draw happened, so software-hardware- TC- combi was too short or weak for this one experiment, if I'll find the time, I'll give Crystal a second one try, don't know when regards

[peter]

So my moves: 59... Qb6 60.g4 Bg3

[fen]8/1kp5/1q1p4/pPpPp1p1/N1P1PpPp/PQ3PbP/8/5K2 w - - 1 61 [/fen]

8/1kp5/1q1p4/pPpPp1p1/N1P1PpPp/PQ3PbP/8/5K2 w - - 0 1

61.Dc2 Kb8 62.Sb2 Kc8 63.a4


[fen]2k5/2p5/1q1p4/pPpPp1p1/P1P1PpPp/5PbP/1NQ5/5K2 b - - 0 63[/fen]

[peter]

As for the game LC0- SF, pity LC0 did let Black's Bishop come to first rank yet and draw happened, so software-hardware- TC- combi was too short or weak for this one experiment, if I'll find the time, I'll give Crystal a second one try, don't know when regards

This one went much better:

[pgn][Event "AMD, Schnellschach 45m+5s"]
[Site "DESKTOP"]
[Date "2022.01.25"]
[Round "1"]
[White "Crystal 251221"]
[Black "Stockfish 220122"]
[Result "1-0"]
[ECO "E87"]
[Annotator "1.40;4.40"]
[PlyCount "101"]
[TimeControl "2700+5"]

1. c4 g6 2. Nc3 Bg7 3. d4 Nf6 4. e4 d6 5. f3 e5 6. d5 Na6 7. Be3 Nh5 8. Qd2
Qh4+ 9. Bf2 Qe7 10. O-O-O O-O 11. Be3 f5 12. Bd3 f4 13. Bf2 Bf6 14. Nge2 Bh4
15. Bg1 Bd7 16. Kb1 b6 17. Nc1 Nc5 18. Bc2 a5 19. Bxc5 bxc5 20. Ba4 Bxa4 21.
Nxa4 Rfb8 22. Qc2 Kh8 23. Rd3 Rb4 24. Ra3 Qe8 25. Nd3 g5 26. h3 Nf6 27. Rc1 Kg7
28. b3 Nd7 29. Qd2 Rb7 30. Ndb2 Nb6 31. Ka1 h5 32. Nd1 Nxa4 33. Rxa4 Rb4 34.
Nc3 Qc8 35. Nb5 Bg3 36. Qe2 Kf6 37. Rb1 Ke7 38. a3 Rxa4 39. bxa4 Kd8 40. Nc3
Rb8 41. Rb5 Rxb5 42. axb5 Qa8 43. Na4 Qa7 44. Ka2 Kc8 45. Qb2 Qb6 46. Kb3 {1.
40/35 191} h4 {4.40/57 341 (Da7)} 47. Qd2 {4.18/40 95 (Kc2)} Qa7 {4.40/63 69
(Kb7)} 48. Nb2 {5.40/39 62 (Ka2)} Bh2 {8.79/37 327} 49. Ka4 {5.96/35 67} Bg1 {
10.44/50 483} 50. Qxa5 {6.41/39 84} Qb7 {10.56/44 70} 51. Kb3 {7.07/38 95} 1-0
[/pgn]
Of course I should have set GUI to not adjudicate with high eval from both sides, yet at this end White had won anyhow, hadn't it?

And the game doesn't tell so much, because Black's defence wasn't what humans would like to see, anyhow, as presumed, Crystal (without NNUE and with Syzygy Probe Limit 0) avoiding 46.Nxb6+ (?) didn't have any difficulties against SF, which (as one can see from the evals) "saw" the breakthrough coming all the time, and didn't find a way to deny it at this not so small hardware- TC, so what?

Reason on the other hand will have been, SF expected Nxb6 here and there once and again and therefore didn't have ponder- hits as for the alternatives coming, 46...h4 (?1, not even game changing already, but symptomatic) e.g. shows this misevaluation of the Queen- sac, as for the eval and as for the progress. It's an example for underperformance because of overestimation of an expected move, not even underestimation of the alternatives but giving in therefore to otherwise longer avoidable ones

45'+5", 30 threads of a 16x3.5GHz Ryzen, ponder off, 8G hash each, SF Syzygy Probe Limit 6, full 6men loaded, default params and net.

[Milos]

8/1kp5/1q1p4/pPpPp1p1/N1P1PpPp/PQ3PbP/8/5K2 w - - 0 1

61.Dc2 Kb8 62.Sb2 Kc8 63.a4


[fen]2k5/2p5/1q1p4/pPpPp1p1/P1P1PpPp/5PbP/1NQ5/5K2 b - - 0 63[/fen]

It's very straight forward from here, I'm just interested in what kind of path Crystal chooses, so the moves (I'll write premoves too):
61.Qc2 Kc8 62. Nb2 Qa7 63. a4 Qa8 64. Nd3 Kb7
[fen]q7/1kp5/3p4/pPpPp1p1/P1P1PpPp/3N1PbP/2Q5/5K2 w - - 3 65 [/fen]

[peter]

8/1kp5/1q1p4/pPpPp1p1/N1P1PpPp/PQ3PbP/8/5K2 w - - 0 1

61.Dc2 Kb8 62.Sb2 Kc8 63.a4


[fen]2k5/2p5/1q1p4/pPpPp1p1/P1P1PpPp/5PbP/1NQ5/5K2 b - - 0 63[/fen]

It's very straight forward from here, I'm just interested in what kind of path Crystal chooses, so the moves (I'll write premoves too):
61.Qc2 Kc8 62. Nb2 Qa7 63. a4 Qa8 64. Nd3 Kb7
[fen]q7/1kp5/3p4/pPpPp1p1/P1P1PpPp/3N1PbP/2Q5/5K2 w - - 3 65 [/fen]

Wouldn't it be somewhat easier to install a Crystal 4.0 of your own? I can really recommend it

https://github.com/jhellis3/Stockfish/releases/tag/v4.0


[fen]q7/1kp5/3p4/pPpPp1p1/P1P1PpPp/3N1PbP/2Q5/5K2 w - - 0 1[/fen]


Analysis by Crystal 251221:

65.Qc3 Bh2 66.Nxc5+ dxc5 67.Qxe5 Qg8 68.Qf6 Kb8 69.Qh6 Qd8 70.e5 Qe7 71.e6 Kc8 72.Kg2 Bg3 73.Qh8+ Kb7 74.Kf1 Bh2 75.Kf2 Ka7 76.Kg2 Bg3 77.Qe5 Kb7 78.d6 cxd6 79.Qf5 Kb6 80.Qe4 d5 81.cxd5 c4 82.Qd4+ Kc7 83.Qe5+ Kb7 84.d6 Qh7 85.Qd5+ Kb8 86.Qd2 c3 87.Qe2 c2 88.d7 Kb7 89.d8Q
+- (151.24) Depth: 39/134 00:03:04 14697MN

Yet after the position of your premoves, SF with classical eval (faster mate search) would serve as well, here I took SF11 for a quick Backward of the line above:

q7/1kp5/3p4/pPpPp1p1/P1P1PpPp/3N1PbP/2Q5/5K2 w - - 0 1

Analysis by Stockfish 11 64 POPCNT:

65.Qc3 Qc8 66.Nxc5+ dxc5 67.Qxe5 Qg8 68.Qf6 Ka7 69.Qa6+ Kb8 70.Qh6 Qd8 71.e5 Qe7 72.e6 Ka7 73.Qh8 Kb7 74.Qe5 Ka7 75.d6 cxd6 76.Qf5 Kb6 77.Qe4 d5 78.Qe5 Kb7 79.Qxd5+ Kc7 80.Qc6+ Kb8 81.Qb6+ Kc8 82.Qa6+ Kd8 83.Qxa5+ Kc8 84.Qa8+ Kc7 85.Qa7+ Kd6 86.Qxe7+ Kxe7 87.b6 Kxe6 88.b7 Ke5 89.b8Q+ Kd4 90.a5 Kxc4 91.a6 Be1 92.Kxe1 Kc3 93.a7 c4 94.Qb5 Kd4 95.a8Q Kc3 96.Qa3+ Kc2 97.Qbb2#
+- (#33) Depth: 84/66 00:01:56 8751MN

So, as for my premoves, please take the ones of the SF- line, even if there could be still a few moves spared as for shortest DTM maybe, I'd say, there can't be much to much of them now that way till we're ready regards

[Milos]

Wouldn't it be somewhat easier to install a Crystal 4.0 of your own? I can really recommend it

I deinstalled it long time ago. It has way too many holes in its search and way too much instability. SF and Lc0 are more than enough.