Engine Blind Spot

Discussion of anything and everything relating to chess playing software and machines.

Moderator: Ras

Martin Brown
Posts: 46
Joined: Sun Oct 18, 2009 12:07 pm

Engine Blind Spot

Post by Martin Brown »

In a game arising out of poisoned pawn Sicilian the following position arose. Black didn't seem to realise his attack had run out of steam and actually played 21. ... a5 punished by a mate in 5.

[d]r1b2rk1/5pp1/p2pp2P/1p6/1q2PP2/3B3Q/2P1K2P/3R3R b - - 0 21

I was expecting one of g6 or e6 but could not decide which of these was best for black. It still looked winable either way. Setting off the engines analysing this position revealed a curiosity. They see the right PV whch is that white then plays Qh4 but they cannot see into the g6 line well enough to reallise that it fails until the search reaches 14+ ply.

After playing 21. ... g6 22. Qh4 a set piece attack against this sort of kingside position they can almost all see quickly that black is in big trouble inside a minute but from the original position only Stockfish can find the right answer in anything like tournament time controls. Strong engines like Rybka 2.32 and Shredder10 take over an hour to get there.

There is something odd about this position that blinds engines to the peril that black is so obviously in. Any idea why the engines can't see into g6 until the leading two plys are explicitly played out?

This position seems to pose unexpected difficulties for even the strongest engines. Any ideas why? Thanks for any enlightenment.
Martin Brown
Top
Martin Brown
Posts: 46
Joined: Sun Oct 18, 2009 12:07 pm

Re: Engine Blind Spot

Post by Martin Brown »

Oops!!!

Obviously it was e5 or g6 as the putative moves for black.
Martin Brown
Top
Terry McCracken
Posts: 16465
Joined: Wed Aug 01, 2007 4:16 am
Location: Canada

Re: Engine Blind Spot

Post by Terry McCracken »

Rybka 3 seems to spot g6 immediately. I'm not sure why your engine didn't or why a mate in 5 was allowed by pushing the a-pawn?
Terry McCracken
Top
Martin Brown
Posts: 46
Joined: Sun Oct 18, 2009 12:07 pm

Re: Engine Blind Spot

Post by Martin Brown »

Terry McCracken wrote:Rybka 3 seems to spot g6 immediately. I'm not sure why your engine didn't or why a mate in 5 was allowed by pushing the a-pawn?
Sorry I wasn't being clear. g6 looks superficially OK for black but also loses fairly quickly as becomes obvious to the engines after g6, Qh4 are played out on the board. e5 is significantly better but still loses. Most of the engines I tried found the latter move from the original position only with extreme difficulty after a hour or so. A 12 ply search from the position after playing out g6 Qh4 shows black is in serious trouble, but it takes a much deeper search from the base position.

The best line for black I reckon is something like
21. ... e5 22. Qh4 f6 23. Rg1 Ra7 24. Rg6
still in big trouble but back in the game if white makes an error

I also saw some of the weirdest fail high codes out of Shredder on this in the overnight multiPV run including: Qxe4+ 285.83 and #389 Bb7

The game continued after 21. ... a5 with

22. hxg7 Kxg7 23. Rg1 Kf6 24. Qh4#

An engine could last a little bit longer with something like

22. ... Qxe5+ 23. BxQ f5 24. Qh8+ Kf7 25. gxf8=Q Kg6 26. Qe8#

But humans don't usually do queen sacs just to stay alive.
Martin Brown
Top
govert
Posts: 270
Joined: Thu Jan 15, 2009 12:52 pm

Re: Engine Blind Spot

Post by govert »

What most engines seem to miss is the following move:

21. ... g6 22. Qh4 Kh6 23. Qf6!

Most PVs i've seen stick to 21. ... g6 22. Qh4 Kh6 23. Qe7 for a long time.
Top
royb
Posts: 565
Joined: Thu Mar 09, 2006 12:53 am

Re: Engine Blind Spot

Post by royb »

govert wrote:What most engines seem to miss is the following move:

21. ... g6 22. Qh4 Kh6 23. Qf6!

Most PVs i've seen stick to 21. ... g6 22. Qh4 Kh6 23. Qe7 for a long time.
Stockfish 1.6.3 spots this line fairly quickly and moves on to ...e5:

1 -1.41 21.... Bb7 (1.22)
2 +2.26 21.... Bb7 22.hxg7 Kxg7 23.Rhg1+ Kf6 (1.22)
2 -0.60 21.... e5 22.h7+ Kh8 (1.22)
3 -0.16 21.... e5 22.Qg3 g6 23.fxe5 (1.22)
4 -0.16 21.... e5 22.Qg3 g6 23.fxe5 (1.22)
5 -0.08 21.... e5 22.Qg3 g6 23.fxe5 dxe5 24.Qxe5 (1.22)
6 -0.28 21.... e5 22.Qg3 g6 23.fxe5 dxe5 24.Qxe5 (1.23)
7 -0.04 21.... e5 22.f5 g6 23.h7+ Kh8 24.Rb1 Qc5 25.Rhg1 gxf5 (1.23)
7 -1.01 21.... g6 22.Rb1 Qd4 23.Qe3 Qh8 24.c4 bxc4 25.Bxc4 Qxh6 (1.25)
8 -1.05 21.... g6 22.Rb1 Qd4 23.Qe3 Qh8 24.Qh3 d5 25.Kf2 Bb7 26.Rhe1 Rac8 (1.29)
9 -0.44 21.... g6 22.Qh4 Kh7 23.Qf6 Kxh6 24.Rb1 (1.36)
10 +0.60 21.... g6 22.Qh4 f6 23.Qg3 Kh8 24.Qxg6 Ra7 25.Rhg1 f5 26.Rb1 Qd4 27.exf5 Qxf4 28.Kd1 (1.63)
11 +0.68 21.... g6 22.Qh4 f6 23.Qg3 Kh8 24.Qxg6 Ra7 25.Rhg1 f5 26.Rb1 Qd4 27.exf5 Qxf4 28.Kd1 Qf3+ 29.Kc1 (1.98)
12 +1.29 21.... g6 22.Qh4 f6 23.Qg3 Kh8 24.Qxg6 Ra7 25.Rhg1 f5 26.Rg5 Qd4 27.Rdg1 fxe4 28.Bxe4 Raf7 29.Qg8+ (2.72)
13 +1.41 21.... g6 22.Qh4 f6 23.Qg3 Kh8 24.Qxg6 Ra7 25.Rhg1 Qc5 26.Rg2 Qc7 27.Rdg1 Qh7 28.f5 exf5 29.exf5 Qxg6 30.fxg6 Re8+ 31.Kd2 Rc7 (4.30)
14 +2.62 21.... g6 22.Qh4 Qc5 23.e5 f6 24.Rhg1 Kh8 25.exf6 Bb7 26.Rxg6 Qd5 27.Rdg1 Qf3+ 28.Kd2 Be4 29.Bxe4 Qxe4 (7.61)
15 +4.84 21.... g6 22.Qh4 Kh7 23.Qf6 Rg8 24.Qxf7+ Kh8 25.Qf6+ Kh7 26.Rhg1 Bb7 27.Qe7+ Kh8 28.Qxb7 Qc5 29.Qe7 Qh5+ 30.Ke3 Raf8 31.Qxe6 Qc5+ 32.Kf3 Qh5+ 33.Qg4 Qxh6 34.h4 Rf7 (21.79)
15 +3.79 21.... e5 22.Qh4 f6 23.hxg7 Kxg7 24.fxe5 Kf7 25.Qxf6+ Ke8 26.Qg6+ Kd7 27.Rhf1 Rxf1 28.Rxf1 Kc6 29.e6 Kb6 30.e7 Bd7 31.Rf8 Kb7 32.Rxa8 Kxa8 33.e8=Q+ Bxe8 34.Qxe8+ Kb7 (28.47)

I've not enough time at the moment to let it continue running but it seems to see that ...g6 falls apart pretty qucikly.
Top
Terry McCracken
Posts: 16465
Joined: Wed Aug 01, 2007 4:16 am
Location: Canada

Re: Engine Blind Spot

Post by Terry McCracken »

Martin Brown wrote:
Terry McCracken wrote:Rybka 3 seems to spot g6 immediately. I'm not sure why your engine didn't or why a mate in 5 was allowed by pushing the a-pawn?
Sorry I wasn't being clear. g6 looks superficially OK for black but also loses fairly quickly as becomes obvious to the engines after g6, Qh4 are played out on the board. e5 is significantly better but still loses. Most of the engines I tried found the latter move from the original position only with extreme difficulty after a hour or so. A 12 ply search from the position after playing out g6 Qh4 shows black is in serious trouble, but it takes a much deeper search from the base position.

The best line for black I reckon is something like
21. ... e5 22. Qh4 f6 23. Rg1 Ra7 24. Rg6
still in big trouble but back in the game if white makes an error

I also saw some of the weirdest fail high codes out of Shredder on this in the overnight multiPV run including: Qxe4+ 285.83 and #389 Bb7

The game continued after 21. ... a5 with

22. hxg7 Kxg7 23. Rg1 Kf6 24. Qh4#

An engine could last a little bit longer with something like

22. ... Qxe5+ 23. BxQ f5 24. Qh8+ Kf7 25. gxf8=Q Kg6 26. Qe8#

But humans don't usually do queen sacs just to stay alive.
It's a simple position and it's clear White wins regardless. I don't know why an engine would have a problem with this and a5?? is mate in five and have no idea why an engine would play that?

Or was Black human? Still a horrible move by Black.
Terry McCracken
Top
User avatar
Dr.Wael Deeb
Posts: 9773
Joined: Wed Mar 08, 2006 8:44 pm
Location: Amman,Jordan

Re: Engine Blind Spot

Post by Dr.Wael Deeb »

Terry McCracken wrote:
Martin Brown wrote:
Terry McCracken wrote:Rybka 3 seems to spot g6 immediately. I'm not sure why your engine didn't or why a mate in 5 was allowed by pushing the a-pawn?
Sorry I wasn't being clear. g6 looks superficially OK for black but also loses fairly quickly as becomes obvious to the engines after g6, Qh4 are played out on the board. e5 is significantly better but still loses. Most of the engines I tried found the latter move from the original position only with extreme difficulty after a hour or so. A 12 ply search from the position after playing out g6 Qh4 shows black is in serious trouble, but it takes a much deeper search from the base position.

The best line for black I reckon is something like
21. ... e5 22. Qh4 f6 23. Rg1 Ra7 24. Rg6
still in big trouble but back in the game if white makes an error

I also saw some of the weirdest fail high codes out of Shredder on this in the overnight multiPV run including: Qxe4+ 285.83 and #389 Bb7

The game continued after 21. ... a5 with

22. hxg7 Kxg7 23. Rg1 Kf6 24. Qh4#

An engine could last a little bit longer with something like

22. ... Qxe5+ 23. BxQ f5 24. Qh8+ Kf7 25. gxf8=Q Kg6 26. Qe8#

But humans don't usually do queen sacs just to stay alive.
It's a simple position and it's clear White wins regardless. I don't know why an engine would have a problem with this and a5?? is mate in five and have no idea why an engine would play that?

Or was Black human? Still a horrible move by Black.
100% agreed with your statement here Terry....
Dr.D
_No one can hit as hard as life.But it ain’t about how hard you can hit.It’s about how hard you can get hit and keep moving forward.How much you can take and keep moving forward….
Top
Terry McCracken
Posts: 16465
Joined: Wed Aug 01, 2007 4:16 am
Location: Canada

Re: Engine Blind Spot

Post by Terry McCracken »

Dr.Wael Deeb wrote:
Terry McCracken wrote:
Martin Brown wrote:
Terry McCracken wrote:Rybka 3 seems to spot g6 immediately. I'm not sure why your engine didn't or why a mate in 5 was allowed by pushing the a-pawn?
Sorry I wasn't being clear. g6 looks superficially OK for black but also loses fairly quickly as becomes obvious to the engines after g6, Qh4 are played out on the board. e5 is significantly better but still loses. Most of the engines I tried found the latter move from the original position only with extreme difficulty after a hour or so. A 12 ply search from the position after playing out g6 Qh4 shows black is in serious trouble, but it takes a much deeper search from the base position.

The best line for black I reckon is something like
21. ... e5 22. Qh4 f6 23. Rg1 Ra7 24. Rg6
still in big trouble but back in the game if white makes an error

I also saw some of the weirdest fail high codes out of Shredder on this in the overnight multiPV run including: Qxe4+ 285.83 and #389 Bb7

The game continued after 21. ... a5 with

22. hxg7 Kxg7 23. Rg1 Kf6 24. Qh4#

An engine could last a little bit longer with something like

22. ... Qxe5+ 23. BxQ f5 24. Qh8+ Kf7 25. gxf8=Q Kg6 26. Qe8#

But humans don't usually do queen sacs just to stay alive.
It's a simple position and it's clear White wins regardless. I don't know why an engine would have a problem with this and a5?? is mate in five and have no idea why an engine would play that?

Or was Black human? Still a horrible move by Black.
100% agreed with your statement here Terry....
Dr.D
When I re-read it, it appears Black was indeed human. Still, a position that shouldn't cause unusual analysis.
Terry McCracken
Top
User avatar
michiguel
Posts: 6401
Joined: Thu Mar 09, 2006 8:30 pm
Location: Chicago, Illinois, USA

Re: Engine Blind Spot

Post by michiguel »

Martin Brown wrote:In a game arising out of poisoned pawn Sicilian the following position arose. Black didn't seem to realise his attack had run out of steam and actually played 21. ... a5 punished by a mate in 5.

[d]r1b2rk1/5pp1/p2pp2P/1p6/1q2PP2/3B3Q/2P1K2P/3R3R b - - 0 21

I was expecting one of g6 or e6 but could not decide which of these was best for black. It still looked winable either way. Setting off the engines analysing this position revealed a curiosity. They see the right PV whch is that white then plays Qh4 but they cannot see into the g6 line well enough to reallise that it fails until the search reaches 14+ ply.

After playing 21. ... g6 22. Qh4 a set piece attack against this sort of kingside position they can almost all see quickly that black is in big trouble inside a minute but from the original position only Stockfish can find the right answer in anything like tournament time controls. Strong engines like Rybka 2.32 and Shredder10 take over an hour to get there.

There is something odd about this position that blinds engines to the peril that black is so obviously in. Any idea why the engines can't see into g6 until the leading two plys are explicitly played out?

This position seems to pose unexpected difficulties for even the strongest engines. Any ideas why? Thanks for any enlightenment.
I left Gaviota calculating all day. It still chooses g6 despite the score is below -3.00.
That means e5 is even worst. I do not think then that there is a reason why an engine should choose g6 over e5.

Code: Select all

analyze
********* Starts iterative deepening, thread = 0
set timer to infinite
       146   1:      0.0    +2.71  e5
      1275   2       0.0      :-(  
      2262   2       0.0      :-(  
      2401   2:      0.0    +1.23  e5 12.h7+ Kh8 13.f5
      7775   3       0.0      :-(  
     11772   3:      0.0    +0.70  e5 12.Qg3 g6 13.fxe5
     27233   4:      0.1    +0.70  g6 12.h7+ Kh8 13.Ra1 Bb7
     95689   5:      0.3    +0.47  g6 12.h7+ Kh8 13.Rb1 Qd4 14.Qh4
    154808   6       0.5    +0.71  g6 12.Rb1 Qc3 13.h7+ Kh8 14.Rb3 Qd4
    212314   6:      0.7    +0.71  g6 12.Rb1 Qc3 13.h7+ Kh8 14.Rb3 Qd4
    395088   7       1.2    +0.71  g6 12.Qh4 Kh7 13.Rb1 Qd4 14.Qe7 Rh8
    572314   7:      1.8    +0.71  g6 12.Qh4 Kh7 13.Rb1 Qd4 14.Qe7 Rh8
    954295   8       3.0      :-(  g6
   2218130   8       7.0      :-(  
   2496876   8       7.9    +0.21  g6 12.Qh4 Kh7 13.Qe7 e5 14.f5 Bxf5
                                   15.exf5 Kxh6
   3451168   8:     10.8    +0.21  g6 12.Qh4 Kh7 13.Qe7 e5 14.f5 Bxf5
                                   15.exf5 Kxh6
   4693033   9      14.7    +0.21  g6 12.Qh4 Kh7 13.Qe7 e5 14.f5 Bxf5
                                   15.exf5 Kxh6
   6866772   9:     21.5    +0.21  g6 12.Qh4 Kh7 13.Qe7 e5 14.f5 Bxf5
                                   15.exf5 Kxh6
   9637197  10      30.3      :-(  g6
  37173283  10     118.1      :-(  
  39936180  10     126.8    -0.31  g6 12.Qh4 Kh7 13.Qe7 e5 14.f5 Bxf5
                                   15.exf5 Kxh6 16.Ra1
  56912644  10:    180.5    -0.31  g6 12.Qh4 Kh7 13.Qe7 e5 14.f5 Bxf5
                                   15.exf5 Kxh6 16.Ra1
  68431928  11     217.8      :-(  g6
 286684349  11     911.0      :-(  
 301830015  11     959.2    -0.70  g6 12.Qh4 f6 13.Rhg1 Kh8 14.Rxg6 Ra7
                                   15.Rdg1 Qc3 16.Rg7 Bd7
 303011166  11:    962.9    -0.70  g6 12.Qh4 f6 13.Rhg1 Kh8 14.Rxg6 Ra7
                                   15.Rdg1 Qc3 16.Rg7 Bd7
 346968558  12    1105.3      :-(  g6
 846892268  12    2696.2      :-(  
1012812517  12    3231.8    -1.66  g6 12.Qh4 f6 13.Rhg1 Kh8 14.Rxg6 Ra7
                                   15.Rxf6 Rg8 16.e5 dxe5 17.Qf2 Rc7
                                   18.fxe5
2396840284  12:   7635.7    -1.66  g6 12.Qh4 f6 13.Rhg1 Kh8 14.Rxg6 Ra7
                                   15.Rxf6 Rg8 16.e5 dxe5 17.Qf2 Rc7
                                   18.fxe5
2868504195  13    9152.7      :-(  g6
1577290103  13   18735.8      :-(  
2019075442  13   20172.5      :-(  g6
3484819945  13   38526.8      :-(  
3643267067  13   39030.4    -3.11  g6 12.Qh4 Qc5 13.e5 Kh7 14.Qf6 Rg8
                                   15.Qxf7+ Kh8 16.Bxg6 Rxg6 17.Qxg6 Qc4+
                                   18.Ke3 Qc3+ 19.Rd3 Qc5+ 20.Kd2 dxe5
3643549474  13:  39031.3    -3.11  g6 12.Qh4 Qc5 13.e5 Kh7 14.Qf6 Rg8
                                   15.Qxf7+ Kh8 16.Bxg6 Rxg6 17.Qxg6 Qc4+
                                   18.Ke3 Qc3+ 19.Rd3 Qc5+ 20.Kd2 dxe5
3829012572  14   39612.1    -3.11  g6 12.Qh4 Qc5 13.e5 Kh7 14.Qf6 Rg8
                                   15.Qxf7+ Kh8 16.Bxg6 Rxg6 17.Qxg6 Qc4+
                                   18.Ke3 Qc3+ 19.Rd3 Qc5+ 20.Kd2 dxe5
 111226366  14:  55213.1    -3.11  g6 12.Qh4 Qc5 13.e5 Kh7 14.Qf6 Rg8
                                   15.Qxf7+ Kh8 16.Bxg6 Rxg6 17.Qxg6 Qc4+
                                   18.Ke3 Qc3+ 19.Rd3 Qc5+ 20.Kd2 dxe5
Top

Return to “Computer Chess Club: General Topics”