01 Mar 07
I have heard tell that engines don't play endgames that well. Somehow that led me to this question:
Is there a position such that:
a. According to a tablebase white will win without violating the 50 move rule.
b. Modern Engines (i.e. Fritz 10) without a tablebase access the situation as black winning (anlysis less than or equal to -1) for at least the first 10 seconds of analysis*
*clearly time is dependent on the quality of the machine, but assuming at least 512+ ram, 2ghz+ processor.
If not, simply maximize the time element of b.
01 Mar 07
Originally posted by vipiuThis is correct. The problem for engines in endgames is that there are copious numbers of positions where the evaluation is the same. This means that the tree pruning algorithm doesn't work as well. On top of that the lines of play tend to be very long, so where a human can tell that a king can reach a square in 8 moves, and starts calculating from the future position without having to work out what's happened on the way - an engine has to mechanically go through each move.
I think Fritz10 would not be able even to win against you a K+N+B vs K in a blitz time....
not sure about a standard time.
K + N + B v K is a good example of this, the checkmate it takes upto 33 moves and your engine won't manage to calculate that deeply. So it'll just make pointless moves while assessing the position as +6.00.
Some of them can have problems with zugzwang positions at the end of their search trees (due to what is called the null move heuristic) - engines with that in their algorithm may actually completely misassess the position (mistake a loss for a win). You may be able to find an example of an engine unfriendly position of that type on the internet.
01 Mar 07
Originally posted by zebanoIt's weakness is that it consistently fails to recognise drawn positions (for itself or it's opponent). If the first two moves of a three move repetition starts after 6 moves (ie, you make six moves, then start repeating moves...) the draw is out of the engines horizon. Here's an interesting one which i've posted before (yes it's me, but i buggered if i'm going to look for another example...)
I have heard tell that engines don't play endgames that well. Somehow that led me to this question:
Is there a position such that:
a. According to a tablebase white will win without violating the 50 move rule.
b. Modern Engines (i.e. Fritz 10) without a tablebase access the situation as black winning (anlysis less than or equal to -1) for at least the fi ...[text shortened]... assuming at least 512+ ram, 2ghz+ processor.
If not, simply maximize the time element of b.
Game 2434914
...when the combinations started the engine see's a nice advantage in winning the pawn (it is probably the highest result the engine could find...) so it takes it, but it doesn't see the three move repetition as that happens towards the edge of it's horizon. Once the pawn is 'won' it soon becomes apparent that the best it can achieve is a draw, as all other lines are inferior for it...
There is a Mary Ann game as well where she resigned in a flat drawn position. I'll see if i can dig it up...
01 Mar 07
1 edit
Originally posted by DeepThoughtThat's pretty cool. How long does it take?
Crafty gets it even when I disable EGTB, although it does take it longer to find the Qxd6+ idea.
I left mine on for a couple of minutes but I don't think it found it.
Nevermind. It found it instantly this time, within seconds.
I must have have been using to much power the last time, Had internet open, fritz, Adobe etc...
But I know there are a couple I had before which the program could not find.
01 Mar 07
Originally posted by MarinkatombThat is the sort of position you would expect a 2200 player to spot easily and not get his queen trapped so would be highly indicitive of engine use.
It's weakness is that it consistently fails to recognise drawn positions (for itself or it's opponent). If the first two moves of a three move repetition starts after 6 moves (ie, you make six moves, then start repeating moves...) the draw is out of the engines horizon. Here's an interesting one which i've posted before (yes it's me, but i buggered if i' ...[text shortened]... as well where she resigned in a flat drawn position. I'll see if i can dig it up...
I say indicitive because sometimes even strong players make mistakes and if they can miss a mate in 1 they can miss this sort of combination. But if it is missed on a number of occasions in conjunction with the line being played being assessed as best by an engine that becomes more than indicitive.
01 Mar 07
Originally posted by RahimKI always do a quick engine check once a game is over to see if I missed anything tactical. In this game Game 1448038 the end position is assessed at +5.76 by Crafty at 22 ply. Which came as a suprise, because I was sure it was drawn when I was playing it. So I tryed playing it out against the engine (I can't remember how much time I gave it) and sure enough it moved it's king around pointlessly for about 30 moves...
That's pretty cool. How long does it take?
I left mine on for a couple of minutes but I don't think it found it.
Nevermind. It found it instantly this time, within seconds.
I must have have been using to much power the last time, Had internet open, fritz, Adobe etc...
But I know there are a couple I had before which the program could not find.
01 Mar 07
Originally posted by DeepThoughtNice.
I always do a quick engine check once a game is over to see if I missed anything tactical. In this game Game 1448038 the end position is assessed at +5.76 by Crafty at 22 ply. Which came as a suprise, because I was sure it was drawn when I was playing it. So I tryed playing it out against the engine (I can't remember how much time I gave it) and sure enough it moved it's king around pointlessly for about 30 moves...
I get annoyed when some programs don't allow a draw even when it's drawn.
That's why I set my chess program to draw early 🙂
02 Mar 07
Originally posted by DeepThoughtIn K + B + N v. K, there's a clear way to make progress: force the enemy king into a corner of the bishop's color. Any decent engine will have code to handle this endgame that evaluates the position not only on material but on how far away the king is from the corner, so that it will force the king toward the corner until it gets close enough to see the mate.
K + N + B v K is a good example of this, the checkmate it takes upto 33 moves and your engine won't manage to calculate that deeply. So it'll just make pointless moves while assessing the position as +6.00.
02 Mar 07
Originally posted by GregMWhy would a modern engine have such code when a tablebase would provide all the information needed without requiring the engine to be more complex?
In K + B + N v. K, there's a clear way to make progress: force the enemy king into a corner of the bishop's color. Any decent engine will have code to handle this endgame that evaluates the position not only on material but on how far away the king is from the corner, so that it will force the king toward the corner until it gets close enough to see the mate.
02 Mar 07
Originally posted by XanthosNZPerhaps tablebases aren't installed (you don't want your engine to get killed when for some reason it can't find its tablebases), or perhaps the author didn't want to go to the trouble to implement tablebase access (basic endgame knowledge like KBNK takes only a few lines of code), or perhaps its a slightly more complex game, like an endgame where pawns can be advanced, but only slowly and laboriously. The point is that while an engine might not see a clear win, it doesn't always need find one -- just a way to make progress.
Why would a modern engine have such code when a tablebase would provide all the information needed without requiring the engine to be more complex?
However, I think it is true that engines are often worse than humans at understanding long-term factors in endgames.
02 Mar 07
Originally posted by DeepThoughtso, how did they make Deep Blue such a beast?
This is correct. The problem for engines in endgames is that there are copious numbers of positions where the evaluation is the same. This means that the tree pruning algorithm doesn't work as well. On top of that the lines of play tend to be very long, so where a human can tell that a king can reach a square in 8 moves, and starts calculating from th ...[text shortened]... ay be able to find an example of an engine unfriendly position of that type on the internet.