NNs seem to solve the Gusev Qxe5!! easily.

[Milton]

Stockfish does find it, but not fast.

New game Line
4q1kr/p6p/1prQPppB/4n3/4P3/2P5/PP2B2P/R5K1 w - - 0 1

Analysis by Stockfish 141119 64 BMI2:

1.Qxe5 fxe5 2.Rf1 Qd8 3.Bd1 Qxd1 4.Rxd1 Rxe6 5.Rd8+ Kf7 6.Rxh8 g5 7.Rxh7+ Kg6 8.Rh8 Rd6 9.Bf8 Rd2 10.Rh6+ Kf7 11.Ba3 Ke8 12.Re6+ Kf7 13.Rd6 Re2 14.Rd7+ Kg6 15.Rxa7 Kh5 16.Rb7 Kg4 17.Rxb6 Kf3 18.Rf6+ Kxe4 19.Rf2 Re3 20.Kg2 Kd5 21.Be7 Re4 22.Kf3 g4+ 23.Kg3 Ke6 24.Bg5 Kd5 25.a3 Rc4 26.Rf5 Ke6 27.Rf6+ Kd5 28.Rb6 Re4 29.Rb5+ Ke6 30.Kf2 g3+ 31.hxg3 Rc4 32.Be3 Kf5 33.Bd4 Ra4 34.Rxe5+ Kg6 35.Be3 Kh7 36.Rb5 Kg7 37.Bd4+ Kg8 38.Kf3 Kh7 39.Ke4 Kg6 40.Kf4 Ra6 41.Kf3 Kf7 42.Ke4 Kg6 43.Rb6+ Rxb6 44.Bxb6 Kg5 45.Kf3 Kh5 46.Kf2
White is clearly winning: +- (9.60 ++) Depth: 80/91 08:25:44 1306466MN, tb=2147483647

[peter]

Any AB engine that can solve this relatively fast?

Relatively fast is Sting only, as far as I've seen till now, it's been shown already in other postings.

Second best is Houdini Tactical 1 in 2MV- Mode, but only for getting solution- move as second best into output, here after about one minute, depth 30/72:

4q1kr/p6p/1prQPppB/4n3/4P3/2P5/PP2B2P/R5K1 w - - 0 1

Analysis by Houdini 6.03 Pro x64-popcT:

1. = (0.00): 24.Qa3 Rxe6 25.Qxa7 Qf7 26.Qb8+ Qe8 27.Qa7
2. = (0.00): 24.Qxe5 fxe5 25.Rf1 Qd8 26.Rf7 Rc7 27.Bc4 Rxc4 28.Rg7+ Kf8 29.Rxh7+ Kg8 30.Rg7+ Kf8 31.Rh7+

After 10 minutes in depth 38 eval still not rising, aborted.

CorChess takes its time too, but still within one hour, here started with empty hash again in single variant mode:

4q1kr/p6p/1prQPppB/4n3/4P3/2P5/PP2B2P/R5K1 w - - 0 1

Analysis by CorChess 5.0 121119 64 POPCNT:

24.Db4 Txe6 25.Lc4 Sxc4 26.Dxc4
-/+ (-1.06) Tiefe: 7/8 00:00:11 37kN
...
24.Da3 Txe6 25.Dxa7 De7 26.Da8+ De8 27.Db7 Df7 28.Db8+ De8
= (0.00) Tiefe: 52/15 00:43:46 53712MN
24.Dxe5
= (0.06 ++) Tiefe: 53/31 00:54:12 66608MN
...
24.Dxe5
+- (3.03 ++) Tiefe: 53/83 00:57:24 70862MN
24.Dxe5 fxe5 25.Tf1 De7 26.Ld1 b5 27.Lb3 Tc4 28.a4 a6 29.Kh1 g5 30.h4 g4 31.Kg2 Dg7 32.Lxg7 Kxg7 33.Tf7+ Kg6 34.Lxc4 bxc4 35.Tf5 Te8 36.h5+ Kg7 37.Txe5 Kf6 38.Tc5 Tb8 39.Kg3 Kxe6 40.Kxg4 Txb2 41.Kf4 Kd6 42.Td5+ Ke6 43.Ke3 Tb1 44.Ta5 Td1 45.Txa6+ Kd7 46.Ta7+ Kc6 47.Txh7 Td3+ 48.Kf4 Txc3 49.a5 Kb5 50.e5 Tc1 51.e6 c3 52.Tc7 c2 53.e7 Te1 54.h6 c1T 55.Txc1 Txe7
+- (3.25) Tiefe: 53/83 00:57:37 71155MN
...
24.Dxe5
+- (7.19 ++) Tiefe: 54/99 01:08:36 87003MN

All trials with 24 threads of 12x3GHz CPU, 32G hash.

[Jouni]

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

[dragontamer5788]

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.

[shrapnel]

NN engines seem to be able to solve the Gusev Qxe5!! incredible Queen sacrifice quite easily!
And with solve i mean not only to find the Qxe5 move, but to have a positive score and show correct continuation also!

Fat Fritz solves it in seconds, same as Leela's T40(20x256) nets, e.g T40B.4-160 but also many others.
New immature T60 Leela net e.g 61104 i tried, solves it also after 1.7 million nodes (around 1 minute).
Yet Leela's J13B.2-188 bigger(24x320) net as also the 136 one, cannot solve it after 5 million nodes searched.

40 seconds, 1 minute 51 seconds ??
TCHAH !
My lco using the Superb Sergio 1178 NN finds it in EIGHT SECONDS FLAT !
Of course, having twin ASUS 2080 Tis helps, a bit. heh heh
Maybe you guys would have been better off saving the money spent on Fat ( and slow) Fritz, a pale imitation of Lc0, to buy a couple of good ol' 2080 Tis instead.
Feast your eyes, boys...

[Dann Corbit]

On the other hand, I have positions where NN engines toil for an hour without finding a tactical solution that alpha-beta engines find in a few seconds.

When you only have a hammer in your tool kit, everything looks like a nail.

[MikeGL]

i have no access to a relevant analysis by any NN because i dont use a seperate grafics-card...imo its pretty easy to make (queen)-sacs without having "-9" in your head

i predicted a superiority of NN's in approximately 5 years or less and i still dont see why this will be not the case...

Speaking of prediction, more than 1 year ago, I have also predicted that NN would solve the above position easily, when I have no access to GPU and NN engines that time. see this link::

http://talkchess.com/forum3/viewtopic.p ... e5#p777441

I have seen so many Lc0 games and knew its tricks against Stockfish are long term positional maneuver involving trapped piece, be it Queen or King or minor pieces.

[Zenmastur]

On the other hand, I have positions where NN engines toil for an hour without finding a tactical solution that alpha-beta engines find in a few seconds.

When you only have a hammer in your tool kit, everything looks like a nail.

A point often purposely (or otherwise) overlooked by NN advocates. Neither AB nor NN engines are a chess panacea. Each has their strengths and weaknesses. For anyone to claim otherwise seems like pure folly lacking extraordinary evidence to the contrary.

Regards,

Zenmastur

[Ovyron]

A point often purposely (or otherwise) overlooked by NN advocates. Neither AB nor NN engines are a chess panacea. Each has their strengths and weaknesses. For anyone to claim otherwise seems like pure folly lacking extraordinary evidence to the contrary.

Not only that, sometimes that happens between A/B engines. Just because Stockfish is the best in most positions doesn't mean it's the best in all, and some engine (non-Stockfish related) would outperform it (and presumably NNs if the position is tactical). And sometimes only 1 Stockfish derivative finds the move, but it's not always the same one...

All I know is that despite all the apparent progress ELO-wise from chess engines, I still have as much problems finding my moves as in 2015. The nature of the weaknesses has changed, but not the amount.

[MikeGL]

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.