PDA

View Full Version : Computer can't handle it



Jesper Norgaard
25-01-2010, 09:06 AM
I have stumbled over an interesting case of chess program blindness. I am trying to analyze my game from last Thursday, which I analyzed with Rybka (I won easily after a blunder by White). This is the critical position after 90...Rgf6:

8/8/5r1k/8/p4r1p/P3Q1pP/6P1/6K1 w - - 0 91

White to move, and Black's long term threat is Rf1+ and going into a winning pawn endgame. For the moment that is ruled out because of the bind of Rf4. But White needs to move and has no useful moves, indeed he must reckon on Rf1+ mate if not careful. Unfortunately with whatever chess program I analyze this, the pawn endgame is too far out in the horizon, perhaps the program sees it is advantageous, but thinks that keeping both rooks on the board is even more advantageous, and so will not defend against the threat of entering the pawn endgame.

Play with chess programs often goes 91.Qd2?? (which actually happened in the game) 91...Kg7! (I played the inferior 91...Kg6) 92.Qd7+,Rf7 and now only Qd3 or Qd1 or Qb5 is possible after which Rf1+ wins easily.

Going back to the original position, 91.Qc1 seems like a better try, but also 91.Qe2 seems OK since the pawn endgame will only win with the black King starting from the g-file, not from the f-file. But a myriad of variations occur and I have many difficulties figuring out if there is a forced win. For instance it is a "forced" win after the beautiful sequence 91.Qc1,Kh7 92.Qc2+,R6f5 93.Qc7+,Rf7 94.Qc2+,Kh8 95.Qc8+,Rf8 96.Qc3+,Kg8! and now there are no more checks and Rf1+ wins, but this is not really forced at all, for instance because of the alternative 91.Qe2.

How can I make Rybka (or other engine) understand that Rf1+ is a decisive threat, so that this can be analyzed confidently?

I was sincerely in doubt whether I should put this in the computer section of chess chat or in the Games and Analysis, but thought this was more like real-game analysis.

Capablanca-Fan
25-01-2010, 12:14 PM
I have stumbled over an interesting case of chess program blindness. I am trying to analyze my game from last Thursday, which I analyzed with Rybka (I won easily after a blunder by White). This is the critical position after 90...Rgf6:

8/8/5r1k/8/p4r1p/P3Q1pP/6P1/6K1 w - - 0 91

White to move, and Black's long term threat is Rf1+ and going into a winning pawn endgame. For the moment that is ruled out because of the bind of Rf4. But White needs to move and has no useful moves, indeed he must reckon on Rf1+ mate if not careful. Unfortunately with whatever chess program I analyze this, the pawn endgame is too far out in the horizon, perhaps the program sees it is advantageous, but thinks that keeping both rooks on the board is even more advantageous, and so will not defend against the threat of entering the pawn endgame.

Play with chess programs often goes 91.Qd2?? (which actually happened in the game) 91...Kg7! (I played the inferior 91...Kg6) 92.Qd7+,Rf7 and now only Qd3 or Qd1 or Qb5 is possible after which Rf1+ wins easily.

Going back to the original position, 91.Qc1 seems like a better try, but also 91.Qe2 seems OK since the pawn endgame will only win with the black King starting from the g-file, not from the f-file. But a myriad of variations occur and I have many difficulties figuring out if there is a forced win. For instance it is a "forced" win after the beautiful sequence 91.Qc1,Kh7 92.Qc2+,R6f5 93.Qc7+,Rf7 94.Qc2+,Kh8 95.Qc8+,Rf8 96.Qc3+,Kg8! and now there are no more checks and Rf1+ wins, but this is not really forced at all, for instance because of the alternative 91.Qe2.

How can I make Rybka (or other engine) understand that Rf1+ is a decisive threat, so that this can be analyzed confidently?

I was sincerely in doubt whether I should put this in the computer section of chess chat or in the Games and Analysis, but thought this was more like real-game analysis.
Right, after the exchange on f1, Black needs to be able to seize the distant opposition with ...Kf7 or Kf5.

GM Joel Benjamin, in his Sunshine Coast 1998/9 lecture on Kasparov v Deep Blue (he was part of the latter's team), argued that computers are weak at seeing perpetual check as well. This explains the missed draw in Game 2.

Jesper Norgaard
25-01-2010, 03:12 PM
Right, after the exchange on f1, Black needs to be able to seize the distant opposition with ...Kf7 or Kf5.

The distant opposition is very important in this pawn endgame, but I wouldn't call it that exactly with ...Kf7 or ...Kf5 after all since ...Kf8 or ...Kf6 works fine too (because White can't play Kf2) but 1...Kg7 doesn't work because of 2.Ke2!,Kf6 3.Kd3! and Black will lose the opposition.

The question remains, is it a win? Is there any box out there that will actually try to avoid Rf1+? Or correspondence player?

This is a nice example of a position where a box will always lose as a patzer with White, while it will always allow perpetual check like a patzer with Black. Can we always get 75% against the chess programs if we corner them up in this kind of position?

Desmond
25-01-2010, 06:33 PM
Why is 91.Qc1 so much better than 91.Qd2?

Jesper Norgaard
26-01-2010, 04:54 AM
Why is 91.Qc1 so much better than 91.Qd2?
Because of extra checks: 91.Qc1,Kg7?! 92.Qc7+,Rf7? 93.Qe5+ etc. from here perhaps already Black can't untangle from the perpetual checks. Remember these were the exact same moves that win against 91.Qd2??

Kevin Bonham
26-01-2010, 10:39 AM
Nice example.

I gave this position to Fritz 11 and at first it is saying -.2 to -.4 for all the candidates. Then I entered in your main 91.Qd2 line and once I got to Kxf1 it saw ...Kf6 winning in a few seconds. From that point on it quickly realised Qd2 loses and won't go near it anymore. (Fritz 11 works like this; when it is "prodded" it can often back-revise.)

Then I tried your 91.Qc1 line. Once it got some distance in the king and pawn ending it realised that ending was a loss and started back-revising, but would only back-revise quickly so far. It is very hard to get it to see that ...Rf1+ wins in this line. At the moment after c.20 minutes of prodding it has realised that 95.Qc8+ loses, but it is still working through all the alternatives.


I was sincerely in doubt whether I should put this in the computer section of chess chat or in the Games and Analysis, but thought this was more like real-game analysis.

Where a thread could just as easily go in either section we will usually just leave it where it is posted.

Spiny Norman
26-01-2010, 10:53 AM
8/8/5r1k/8/p4r1p/P3Q1pP/6P1/6K1 w - - 0 91
Interesting position! Given the positioning of the Rf4, and White presumably wanting to give perpetual if possible, I would analyse as follows if I was at the board (and assuming that I had enough time).

Assuming Black wants to hide his king from check and is trying to play for a win, I notice that the Rf4 controls some key squares:
g5 (by blocking the c1-h6 diagonal)
d4 (and also the rest of the 4th rank)

The Rf6 performs a similar function by controlling:
g7 (by blocking the a1-h8 diagonal)
If that rook later moved to Rf7 to block a 7th-rank cheque, then White has to factor in that Black will presumably keep his king mostly on dark squares to take advantage of the Rf4 control of d4 ... therefore White ought (IMO) play Qc1 because he needs to have the option of a later Qe5+, which will not be possible to play from Qd7, but is possible to play from Qc7.

I don't know whether that is sound thinking, but if I am White there I would almost immediately bet on Qc1 being better than Qd2.
EDIT: I have a "theory" that when trying to initiate a perpetual, one should start by developing a "colour complex" strategy ... e.g. "I will try to initiate a perpetual on the black squares" or similar ... can someone comment on that? ... so my perpetual check ideas would be based around the following black squares: c1, c7, e7, e5, g5 and so on

As for the computer analysis, my guess is that they are not good on patterns, but only good at calculation (correct me if I am wrong). In the first couple of seconds that I looked at that position I concluded that Black had good winning chances, based on swapping off the Q+RR and taking distant opposition. It doesn't even need "calculation" to do this; one "sees" it almost immediately, that White's only move by the king to get out of the corner will be Kxf1 and therefore as long as Black's king is one square from f7 or f5 he will have an advantage ... though whether it is enough to win, if a pawn race ensues, is another matter.

So I reckon most human players, whether White or Black, would quickly find the best chances, perhaps in just a few seconds, whereas a computer short of time would flounder in such a position.

Kevin Bonham
26-01-2010, 11:48 AM
Fritz has now been sitting on infinite analysis for an hour in the position after ...Kh8 in the line 91.Qc1,Kh7 92.Qc2+,R6f5 93.Qc7+,Rf7 94.Qc2+,Kh8. It has still yet to reject:

95.Qe2
95.Qd3
95.Qc3+
95.Qc1

Any suggestions on what I should prod it with in these lines?

Garvinator
26-01-2010, 12:13 PM
Does 91. Qe1 change anything?

Kevin Bonham
26-01-2010, 04:55 PM
Does 91. Qe1 change anything?

Very interesting you ask that because after being left on infinite analysis in the original position for about 5 hours, Fritz now thinks that everything loses from the start position except 91.Qe1.

Garvinator
26-01-2010, 05:21 PM
Very interesting you ask that because after being left on infinite analysis in the original position for about 5 hours, Fritz now thinks that everything loses from the start position except 91.Qe1.
91. Qe1 was the very first move that Rybka kept analysing with infinite analysis on my laptop. I tried prodding it a few times but it kept coming back to Qe1.

Spiny Norman
26-01-2010, 05:26 PM
Qe1 seems to be a different strategy ... I assume its considering perpetual up/down the e-file?

Jesper Norgaard
27-01-2010, 05:56 AM
So I reckon most human players, whether White or Black, would quickly find the best chances, perhaps in just a few seconds, whereas a computer short of time would flounder in such a position.
Aren't we getting a tad optimistic here? I agree that the human could quickly make an overview of general ideas (like forcing exchange of 2R against Q) but only Superman would find it in just a few seconds :)

It seems there still is some hope for the computers, because Rybka could set up a winning "strategy" while prodding it, by shuffling the rooks to the e-file instead of the f-file.

White: Kf1,Qb1,a3,g2,h3
Black: Kf5,Re4,Re5,a4,g3,h4

1...Ke6! now the main variation (for a human) runs 2.Qb6+,Kd5 3.Qb5+,Kd4 4.Qxa4+,Kc3 5.Qc6+,Kb2 6.Qb7+,Ka1 and now there is no good defence against Re1+ so instead Rybka played with resignation 2.Qd1 (to stop Kd5) 2...Re1+ 3.Qxe1,Rxe1+ 4.Kxe1,Ke5! and Rybka is dead. However, this was not my idea but Rybka's. The king is dead, long live the king!

I don't know if this is forced, but I have a feeling it might be. I will put up some more precise variations when I get the time. At least this proofs there are plenty of winning chances.

Spiny Norman
27-01-2010, 06:07 PM
Aren't we getting a tad optimistic here? I agree that the human could quickly make an overview of general ideas (like forcing exchange of 2R against Q) but only Superman would find it in just a few seconds :)
Well, I'm not Superman, that's for sure; however I did find the main ending theme and the key squares in just a few seconds (maybe 10?). I'm a mid-1550's rating, though am a lot stronger in endgames (maybe 1800-ish). I have spent most of my chess study time on endgame theory, so perhaps that's why. "key squares", "opposition", "rook endings (how to win/draw)" and so on. I often will deliberately steer games against similarly rated players into an endgame as a result. Still, maybe with this one I just lucked out, I don't know.

Capablanca-Fan
29-01-2010, 11:52 AM
Well, I'm not Superman, that's for sure; however I did find the main ending theme and the key squares in just a few seconds (maybe 10?). I'm a mid-1550's rating, though am a lot stronger in endgames (maybe 1800-ish). I have spent most of my chess study time on endgame theory, so perhaps that's why. "key squares", "opposition", "rook endings (how to win/draw)" and so on. I often will deliberately steer games against similarly rated players into an endgame as a result. Still, maybe with this one I just lucked out, I don't know.
Seems like a sensible assessment.

As Jesper pointed out, the KP endgame is even better for Black because White lacks f2 to take a distant opposition. Black's chances are also enhanced because his Ps are so advanced, so will win promotion races. E.g.

98.Kf3 Kd4! 99.Kg4 Ke3 100.Kxh4 Kf2 101.Kg4 Kxg2 102.h4 Kf2 103.h5 g2 104.h6 g1=Q+

Igor_Goldenberg
29-01-2010, 12:43 PM
The funny this is had pawns bene placed on a2 and a3( instead of a3 and a4) the pawn endgame would be drawn (unless black king was already at the queen side at the time of exchange).

Jesper Norgaard
30-01-2010, 04:18 PM
The funny about this is had pawns been placed on a2 and a3 (instead of a3 and a4) the pawn endgame would be drawn (unless black king was already at the queen side at the time of exchange).

I guess instructive, and also why you should never-never-ever-ever-ever just evaluate a pawn endgame with general terms, a single tempo calculated wrong will often be decisive, all kinds of minefields appear. I love pawn endgames, still have my fair share of lost games in Blitz on the account of not being able to "control the beast". Rook endgames is a completely different ballgame, routinely a 1-2 pawn deficit is drawn by skillful players because they know what they are heading for, and can use general considerations and specific knowledge to steer the boat.

From the current Q vs. 2R endgame we analyze, the pawn endgame wins not because it is decisive that the a3 pawn is captured, because the a4-pawn can't be promoted, but it works as a decoy to finally distract the white king from defending against Kxg2,Kxh3. With pawns on a2 and a3, black can't take on a2 because of 1...Kxa2 2.Kc2! Ka1 3.Kc1 and now 3...a2 will only stalemate black, while 3...Ka2 4.Kc2 is repetition. Look out for the greedy 3.Kb3??, Kb1! (3...a2?? 4.Kc2 stalemate) 4.Kxa3, Kc2-+

Davidflude
30-01-2010, 05:02 PM
I have been running multiprocessor rybka 3 960 32bit on a dual core
3.8 computer with about 1024 meg hash tables.

Initially I set it to look at nine candidate moves.In seconds it decided that there were only six viable moves. Everything else loses to checkmate in a maximum of 10 moves.

So I cut back to analyzing the best six moves.

Within about two minutes

it decided that Qd2 and Qd3 gave winning endings for black. I will not give the details but the machine played preparatory moves followed by the exchange on f1.

Qe2 was more tricky. The machine found Rd6 in seconds. After this white is busted.

So we are down to two moves Qe1 and Qc1.

Now I analyzed Qe1 on its own. The machine found Rd6 within a minute. Again I suspect that this is a move that many humans would miss. Again white was busted after this move.

So the only move is Qc1. as other people have pointed out this is whites best chance. I shall have to turn my machine off now as it is running hot. More analysis later.

Bill Gletsos
30-01-2010, 05:57 PM
I have been running multiprocessor rybka 3 960 32bit on a dual core
3.8 computer with about 1024 meg hash tables.It appears you are running your machine overclocked.

I shall have to turn my machine off now as it is running hot.You need better cooling.

Jesper Norgaard
31-01-2010, 03:46 PM
Now I analyzed Qe1 on its own. The machine found Rd6 within a minute. Again I suspect that this is a move that many humans would miss. Again white was busted after this move.

It may be that after 91.Qe1,Rd6 white is busted (as it may be in many other move orders). However, I am not convinced at all. After 92.Qc1 what would be a convincing line? I keep getting nowhere with 92...Kg6 93.Qc2+,Kf6 94.Qc3+ on 94...Rdd4 95.Qc6+ white keeps coming up with new checks, and if 94...Rfd4 95.Kf1!,Kf7 96.Ke2! suddenly the troll is out of the box (e.g. the white king).

Davidflude
31-01-2010, 11:26 PM
It appears you are running your machine overclocked.
You need better cooling.

Not overclocking just a very hot day in Melbourne and I am conservative when my machine temperature rises. I have extra fans on the machine.

Jesper Norgaard
02-02-2010, 12:35 PM
I am a little disappointed in the various chess programs ability of finding a solution here. I think black is probably winning by force, but that the variations are lengthy and out of the horizon of both humans and programs. Still it would be a great case for retrograde analysis because it is only a matter of defining certain positions as won for black, and work back from them, reshuffling the queen, 2 kings and 2 rooks on the board (white king just confined between e1 and h1 I guess).

Here is an example of Rybka winning against me:

91.Qe1 R4f5 92.Qd2+ Kg6 93.Qd3 Re6 94.Qd1 Rf4 95.Qc2+ Rfe4 96.Kf1 Re5 97.Qc6+ Kh5 98.Qc1 Rf5+ 99.Kg1 Rd5 100.Qb1 Kg5 101.Kf1 Rf5+ 102.Kg1 Rc5 103.Kf1 Rce5 104.Qc1+ Rf4+ 105.Kg1 Rd5 106.Qe1 Kf5 107.Qb1+ Re4 108.Kf1 Rd2 109.Qb5+ Kf4 110.Qb1 Ke3 111.Qe1+ Kd3 112.Qb1+ Rc2 113.Qb5+ Ke3 114.Qb1 Kd2 0-1

Here is an example of me winning against Rybka:

91.Qe1 R4f5 92.Qe2 Kg7 93.Qg4+ Kf7 94.Qc4+ Re6 95.Qc7+ Kg6 96.Qc3 Rd5 97.Qc2+ Kg5 98.Qc1+ Kh5 99.Qa1 Rde5 100.Qd1+ Kh6 101.Qc1+ Kh7 102.Qc7+ Re7 103.Qc2+ Re4 104.Kf1 Kg8 105.Qc8+ Re8 106.Qc1 Rf8+ 107.Kg1 Ref4 108.Qd1 Rf1+ 0-1

Kevin Bonham
03-02-2010, 11:57 PM
Trying to see how far back I can track in Jesper's lines above and still have Fritz smell the coffee.

In Jesper's first game, white is losing after 100.Qb1. At present Fritz thinks 100.Kf1 is viable.

In Jesper's second game, for a long time Fritz could not find a win after 102.Kf1 but it now seems that there is a win in the line 102...Rf6+ 103.Kg1 Rc6! (rook en prise for back rank mate drives queen off c1). On that basis, 101.Qc1+ appears to lose, but Fritz has not yet raised concerns about 101.Kf1.

My experience of this line using Fritz is that it sits on an evaluation like -0.41 or -0.42 while it is not seeing anything. As soon as that evaluation blows out to below -0.5 that means it has caught a whiff of something that, in my experience so far, will always turn out to be a forced win. In some lines the way it initially detects a win is that it works out black can force the capture of the a-pawn with his king! At this point the evaluation blows out to -2 or -3 or so.

I also think that this is probably a forced win, simply because there are so many positions where the computer eventually sees the win but it is very subtle. But it would be very satisfying to reach a point where we can say with confidence that the position is a win.