NNs seem to solve the Gusev Qxe5!! easily.

[Jouni]

Crystal solves in seconds and goes fast over +10:

Analysis by Crystal 140121:


1.Qxe5
= (0.08 ++) Depth: 26/34 00:00:03 32246kN, tb=20
1.Qxe5
= (0.16 ++) Depth: 26/34 00:00:03 32732kN, tb=20
1.Qxe5
= (0.28 ++) Depth: 26/34 00:00:03 33188kN, tb=20
1.Qxe5
+/= (0.46 ++) Depth: 26/50 00:00:05 49228kN, tb=20
1.Qxe5
+/- (0.71 ++) Depth: 26/50 00:00:05 56832kN, tb=20
1.Qxe5
+/- (1.03 ++) Depth: 26/75 00:00:06 64407kN, tb=33
1.Qxe5
+/- (1.47 ++) Depth: 26/75 00:00:07 74159kN, tb=33
1.Qxe5
+- (2.03 ++) Depth: 26/75 00:00:08 81717kN, tb=33
1.Qxe5
+- (2.75 ++) Depth: 26/75 00:00:08 89005kN, tb=40
1.Qxe5
+- (3.68 ++) Depth: 26/75 00:00:09 98283kN, tb=41
1.Qxe5
+- (4.87 ++) Depth: 26/75 00:00:11 111mN, tb=41
..
1.Qxe5 fxe5 2.Rf1 Qe7 3.Bd1 Rc4 4.Bb3 b5 5.a4 a6 6.Kg2 Qg7 7.Bxg7 Kxg7 8.Rf7+ Kh6 9.Bxc4 bxc4 10.e7 Re8 11.Kg3 g5 12.Rf6+ Kg7 13.Rxa6 Rxe7 14.Kg4 Rb7 15.Kxg5 Kf7 16.Kf5 Rxb2 17.Ra7+ Kg8 18.h4 Rb3 19.Kf6 h5 20.Rg7+ Kh8 21.Rg3 Ra3 22.Kxe5 Kh7 23.Kd4 Rxa4 24.Rg5 Kh6 25.e5 Ra5
+- (7.92) Depth: 31/59 00:00:41 423mN, tb=8914
1.Qxe5 fxe5
+- (7.74 --) Depth: 32/62 00:00:50 500mN, tb=14148
1.Qxe5
+- (7.78 ++) Depth: 32/63 00:00:50 501mN, tb=14792
1.Qxe5
+- (7.85 ++) Depth: 32/63 00:00:50 502mN, tb=14793
1.Qxe5
+- (7.93 ++) Depth: 32/63 00:00:50 502mN, tb=14793
1.Qxe5
+- (8.06 ++) Depth: 32/63 00:00:50 502mN, tb=14797
1.Qxe5 fxe5 2.Rf1 Qe7 3.Bd1 Rc4 4.Bb3 b5 5.a4 a6 6.Kg2 Qg7 7.Bxg7 Kxg7 8.Rf7+ Kh6 9.Bxc4 bxc4 10.e7 Re8 11.Kg3 g5 12.Rf6+ Kg7 13.Rxa6 Rxe7 14.Kg4 Rb7 15.Kxg5 Rxb2 16.Ra7+ Kg8 17.h4 Rb3 18.Kf6 h5 19.Rg7+ Kh8 20.Rg3 Ra3 21.Kxe5 Rxa4 22.Kd4 Kh7 23.e5 Ra6 24.Ke4 Ra1 25.e6 Rc1 26.Kf5 Rf1+ 27.Ke5 Re1+ 28.Kd5 Re2
+- (8.08) Depth: 32/63 00:00:50 502mN, tb=14799
1.Qxe5
+- (8.41 ++) Depth: 33/58 00:00:50 506mN, tb=15310
1.Qxe5
+- (8.46 ++) Depth: 33/65 00:00:50 507mN, tb=15312
1.Qxe5
+- (8.52 ++) Depth: 33/65 00:00:50 507mN, tb=15312
1.Qxe5
+- (8.61 ++) Depth: 33/65 00:00:50 510mN, tb=15391
1.Qxe5
+- (8.73 ++) Depth: 33/67 00:00:56 564mN, tb=17767
1.Qxe5
+- (9.41 ++) Depth: 33/69 00:01:04 643mN, tb=30996
1.Qxe5 fxe5 2.Rf1 Qe7 3.Bd1 Rc4 4.Bb3 b5 5.a4 a6 6.Kg2 Qg7 7.Bxg7 Kxg7 8.Rf7+ Kh6 9.Bxc4 bxc4 10.e7 Re8 11.Kg3 g5 12.Rf6+ Kg7 13.Rxa6 Rxe7 14.Kg4 Rb7 15.Kxg5 Rxb2 16.Ra7+ Kg8 17.h4 Rc2 18.Kf6 h5 19.Rg7+ Kh8 20.Rg3 Kh7 21.Kxe5 Ra2 22.Kd4 Rxa4 23.e5 Ra6 24.Ke4 Ra2 25.e6 Rh2 26.Kd5 Re2 27.Rg5 Kh6 28.Kd6
+- (8.75) Depth: 33/71 00:01:04 643mN, tb=31001
1.Qxe5
+- (9.16 ++) Depth: 34/60 00:01:06 668mN, tb=32218
1.Qxe5
+- (9.21 ++) Depth: 34/60 00:01:07 674mN, tb=32589
1.Qxe5 fxe5
+- (9.08 --) Depth: 34/68 00:01:10 697mN, tb=32973
1.Qxe5 fxe5
+- (9.00 --) Depth: 34/68 00:01:10 700mN, tb=33112
1.Qxe5
+- (9.08 ++) Depth: 34/68 00:01:14 739mN, tb=33926
1.Qxe5
+- (9.25 ++) Depth: 34/68 00:01:14 740mN, tb=34272
1.Qxe5 fxe5
+- (8.85 --) Depth: 34/70 00:01:14 740mN, tb=34379
1.Qxe5
+- (9.17 ++) Depth: 34/70 00:01:14 740mN, tb=34452
1.Qxe5 fxe5 2.Rf1 Qe7 3.Bd1 Rc4 4.Bb3 b5 5.a4 a6 6.Kg2 Qg7 7.Bxg7 Kxg7 8.Rf7+ Kh6 9.Bxc4 bxc4 10.e7 Re8 11.Kg3 g5 12.Rf6+ Kg7 13.Rxa6 Rxe7 14.Kg4 Rb7 15.Kxg5 Rxb2 16.Ra7+ Kg8 17.h4 Rc2 18.Kf6 h5 19.Rg7+ Kh8 20.Rg3 Kh7 21.Kxe5 Ra2 22.Kd5 Rxa4 23.e5 Ra5+ 24.Ke4 Kh6 25.e6 Ra1 26.Kf5 Rf1+ 27.Ke5 Re1+ 28.Kf6 Rf1+ 29.Ke7 Rh1 30.Kf7 Rxh4 31.Rg6+ Kh7 32.Rf6
+- (9.32) Depth: 34/70 00:01:15 750mN, tb=34715
1.Qxe5
+- (9.56 ++) Depth: 35/70 00:01:16 757mN, tb=35025
1.Qxe5
+- (9.60 ++) Depth: 35/70 00:01:23 826mN, tb=41265
1.Qxe5
+- (9.66 ++) Depth: 35/70 00:01:26 855mN, tb=74807
1.Qxe5
+- (9.75 ++) Depth: 35/70 00:01:26 860mN, tb=82071
1.Qxe5 fxe5 2.Rf1 a6 3.Bd1 b5 4.Bb3 Rc4 5.Kg2 Qe7 6.a4 Qg7 7.Bxg7 Kxg7 8.Rf7+ Kh6 9.Bxc4 bxc4 10.e7 Re8 11.Kg3 g5 12.Rf6+ Kg7 13.Rxa6 Rxe7 14.Kg4 Rb7 15.Kxg5 Rxb2 16.Ra7+ Kg8 17.h4 Rc2 18.Kf6 h5 19.Rg7+ Kh8 20.Rg3 Kh7 21.Kxe5 Kh6 22.a5 Ra2 23.Kf6 Rf2+ 24.Ke6 Rh2 25.Kd5 Ra2 26.e5 Rxa5+ 27.Ke4 Ra6 28.Kd4 Kh7 29.Kd5 Ra5+ 30.Ke4 Ra6 31.Kf5 Ra2 32.e6 Re2 33.Kf6 Rf2+ 34.Ke5 Re2+ 35.Kd6 Ra2 36.Kd5 Ra5+ 37.Kd4 Ra6 38.e7
+- (9.86) Depth: 35/83 00:01:35 944mN, tb=89427
1.Qxe5 fxe5
+- (10.16 --) Depth: 36/81 00:01:51 1101mN, tb=95790
1.Qxe5 fxe5
+- (10.12 --) Depth: 36/81 00:01:55 1140mN, tb=96917
1.Qxe5
+- (10.16 ++) Depth: 36/81 00:01:55 1142mN, tb=96918
1.Qxe5 fxe5
+- (10.06 --) Depth: 36/81 00:01:56 1145mN, tb=97492
1.Qxe5
+- (10.16 ++) Depth: 36/81 00:01:56 1146mN, tb=97804
1.Qxe5 fxe5 2.Rf1 a6 3.Bd1 b5 4.a4 Rc4 5.Bb3 Qe7 6.Kg2 Qg7 7.Bxg7 Kxg7 8.Rf7+ Kh6 9.Bxc4 bxc4 10.e7 Re8 11.Kg3 g5 12.Rf6+ Kg7 13.Rxa6 Rxe7 14.Kg4 Rb7 15.Kxg5 Rxb2 16.Ra7+ Kg8 17.h4 Rc2 18.Kf6 h5 19.Rg7+ Kh8 20.Rg3 Kh7 21.Kxe5 Kh6 22.a5 Ra2 23.Kf6 Rf2+ 24.Ke6 Rh2 25.Kd5 Ra2 26.e5 Rxa5+ 27.Ke4 Ra6 28.Kd4 Ra1 29.e6 Rh1 30.e7 Re1 31.Re3
+- (10.11) Depth: 36/81 00:01:56 1147mN, tb=98022
1.Qxe5
+- (10.05 ++) Depth: 37/68 00:01:58 1168mN, tb=99325
1.Qxe5
+- (10.09 ++) Depth: 37/68 00:02:00 1189mN, tb=100568
1.Qxe5
+- (10.15 ++) Depth: 37/68 00:02:03 1219mN, tb=107240
1.Qxe5
+- (10.24 ++) Depth: 37/81 00:02:14 1320mN, tb=125359
1.Qxe5 fxe5 2.Rf1 Qe7 3.Bd1 Rc4 4.Bb3 b5 5.a4 a6 6.Kg2 Qg7 7.Bxc4 bxc4 8.Bxg7 Kxg7 9.Rf7+ Kh6 10.e7 Re8 11.Kg3 g5 12.Rf6+ Kg7 13.Rxa6 Rxe7 14.Kg4 Rb7 15.Kxg5 Rxb2 16.Ra7+ Kg8 17.h4 Rc2 18.Kf6 h5 19.Rg7+ Kh8 20.Rg3 Kh7 21.Kxe5 Kh6 22.a5 Ra2 23.Kf6 Rf2+ 24.Ke6 Rh2 25.a6 Ra2 26.e5 Rxa6+ 27.Kf7 Ra7+ 28.Kf6 Ra2 29.e6 Rf2+ 30.Ke7 Ra2 31.Kf7 Rf2+ 32.Ke8 Re2 33.e7 Re1 34.Rg5 Re3 35.Rd5 Rxc3
+- (10.21) Depth: 37/81 00:02:14 1328mN, tb=128362

[MMarco]

If You do insignificant change like Ra1->b1 are NN engines still as fast?

I reckon it would still see it. It is not storing exact positions like the *.lrn file of crafty but storing ideas with similar piece configuration on the board.

But as a side note, using SF12+,
I quickly fired up my available cpu on my 12-year-old dual core (2008), yes very old laptop, and this Qxe5!! move already pops-up at Top 4 with -1.50 in
just a mere 3 minutes of calculation of SF12+. This is amazing in my opinion because most engines put this move outside Top 10 possible moves on this
specific position. Which means this move can be found by correspondence players in less than 2 hours even with an old machine if checked using SF12+.

Code: Select all

22/35	  12.817.808	-1.50	Qxe5 fxe5 Rf1 Rc7 Bb5 Qe7 Ba4 Qc5+ Kg2 Rg7 Rd1 Qf8 Bb3 Re7 Bxf8 Kxf8 Rd6 Kg7 Bd5 Rc7 c4 Kf6 b4 Ke7 c5 Rb8 a4 bxc5 bxc5 Rxc5 Rd7+ Kf6

That is an amazing line of SF12+, considering this was crunched in less than 3 minutes in a 12-yr-old (2008) dual core.
Faster machine and a bit of optimization on search module of SF12 might solve this position in a tournament time control.

Earlier SV nets (around july 2020) and pure nnue get this one quickly. On my old phone (kirin 710) 4 theaded sfnnue (before sf12 and hybridization) took 30s with SV-1134:

Code: Select all

[40] 2.99 Dxe5 fxe5 Tf1 De7 Fd1 b5 Tf2 Tc4 Fb3 a6 a4 De8 Da8 Rg3 De8 Tf3 De7 axb5 axb5 Rg2 Dd8 h4 De8 h5 De7 Fa2 g5 Fb3 Dd8 Tf7 De8 Tf5 De7 Rh2 Dd8 Rg1
d:41/46 1:Dxe5 t:38.41 n:30653k nps:797k tb:84 h:19

[M ANSARI]

What an incredible position!!! Did this come from a real game? If so do you have the full game? Thanks for posting.

[Master Om]

What an incredible position!!! Did this come from a real game? If so do you have the full game? Thanks for posting.

Yuri Gusev vs Yuri Averbakh.

[Andrew]

But Stockfish dev(12 November 2019) cannot solve it even after 35 billion nodes searched and 67 ply. It cannot even find the Qxe5 move.

Stockfish is known for pruning. Stockfish doesn't "look for the best move", it instead "looks for a move that probably wasn't a bad move". Late move reductions fundamentally means that Stockfish will spend far, far less time looking at something like Qxe5 (since it immediately results in losing ~900 centipawns).

Not only are moves like Qxe5 checked less often, but they are also searched to a far shallower depth. Stockfish may have searched the PV down to 67 ply, but it may have only checked Qxe5 down to 5 ply.

----------

This is why I like MCTS-based searches in theory. They have an innately better heuristic for search IMO. Even ignoring the neural-net here, the MCTS-based search will exhaustively try all paths to a shallow depth. There seems to be a more sound theoretical basis for "shallow vs depth" (aka: Explore vs Exploit) for MCTS-based engines.

AB-engines always felt like a hack. (Lets ignore late moves: it increases Elo!!) Nothing wrong with that, but it means giving up moves like Qxe5 in this position.

AB engines have troubles with the initial position. However if you play out Qxe5 fxe5 Rf1 it looks like most of them still have problems!
With black to move many still show a draw score and have trouble seeing they are losing. So is there another sacrifice or zugwang position
deeper in the analysis?

Andrew

[Andrew]

Or is the big problem for AB engines the trapped rook on H8?

Andrew

[Geonerd]

AB engines have troubles with the initial position. However if you play out Qxe5 fxe5 Rf1 it looks like most of them still have problems!
With black to move many still show a draw score and have trouble seeing they are losing. So is there another sacrifice or zugwang position
deeper in the analysis?

Andrew

Yea, this is a quite complex problem. There are several 'threshold' limits that seem to stymie the engines. Play it out! There are subtle and somewhat 'slow' bishop maneuvers, and zugzwang themes, that the engines have a hard time seeing.

[MMarco]

But Stockfish dev(12 November 2019) cannot solve it even after 35 billion nodes searched and 67 ply. It cannot even find the Qxe5 move.

Stockfish is known for pruning. Stockfish doesn't "look for the best move", it instead "looks for a move that probably wasn't a bad move". Late move reductions fundamentally means that Stockfish will spend far, far less time looking at something like Qxe5 (since it immediately results in losing ~900 centipawns).

Not only are moves like Qxe5 checked less often, but they are also searched to a far shallower depth. Stockfish may have searched the PV down to 67 ply, but it may have only checked Qxe5 down to 5 ply.

----------

This is why I like MCTS-based searches in theory. They have an innately better heuristic for search IMO. Even ignoring the neural-net here, the MCTS-based search will exhaustively try all paths to a shallow depth. There seems to be a more sound theoretical basis for "shallow vs depth" (aka: Explore vs Exploit) for MCTS-based engines.

AB-engines always felt like a hack. (Lets ignore late moves: it increases Elo!!) Nothing wrong with that, but it means giving up moves like Qxe5 in this position.

AB engines have troubles with the initial position. However if you play out Qxe5 fxe5 Rf1 it looks like most of them still have problems!
With black to move many still show a draw score and have trouble seeing they are losing. So is there another sacrifice or zugwang position
deeper in the analysis?

Andrew

Here is a piece of analysis:
http://www.talkchess.com/forum3/viewtop ... 66#p110578

And later on, here: http://talkchess.com/forum3/viewtopic.p ... ev#p777361

The position generated a lot of interest over the years.

[Jouni]

After finding winning 24. Qxe5!! Gusev played drawing 28. Bxc4?? missing from most net analysis .

[M ANSARI]

Actually not only is the rook trapped and black basically playing without a rook, but the king is also in a choke hold and queen is totally immobilized due to the instant mate threat. The black rook is also easily tied down with the white bishop as there is discovered check on that diagonal. You would think that somehow black could be able to find one or two moves to get a freeing check on the white King, but that an easily be avoided. It is truly remarkable that such a position arose from actual game play and was not a built up study.