Author: Robert Hyatt
Date: 09:43:10 09/29/05
Go up one level in this thread
On September 28, 2005 at 18:20:42, Uri Blass wrote: >On September 28, 2005 at 17:54:13, Robert Hyatt wrote: > >>On September 28, 2005 at 14:38:39, K. Burcham wrote: >> >>> >>> >>>Do you know of a position in the Deep Blue vs Kasparov games, that todays >>>programs cannot find? I am trying to find a position that not even one program >>>today can find this position without very long search. I will conclude that this >>>position cannot be found with default knowledge. I am looking for a position >>>that I can see that no program of today, with default settings will find this >>>Deep Blue game position without leaving one of my programs running for a very >>>deep search. >>> >>>kburcham >> >> >>I am almost certain that, by definition, such a move is not going to be found. >>Because many moves (both good and bad ones) are made for the wrong reasons. And >>most likely, unless we find some wild tactical position that simply confuses >>current searches completely, but DB was able to work thru it, then most any >>other kind of position is not going to be useful. >> >>For example, how many times have we seen problem positions where it is "white to >>play move XXX and win". And one (or more) programs play move XXX with a + >>score, that is convincing. And one (or more) programs play move XXX with a - >>score, which means it is playing the right move for the wrong reasons. Ditto >>for programs that also play move XXX with a score of 0.0 thinking it is the only >>way to avoid a loss, when it is the right way to actually win. Most positions, >>other than those ending in mate, or with huge material wins, rely on both search >>and evaluation. And a program that is a bit top-heavy on king safety might find >>more of Tal's favorite moves than a program with a more conservative >>(Karpov-like) evaluation. Yet both are equally strong overall. >> >>Finding these kinds of positions is _very_ difficult, because there are so few >>DB 2 (last match) games available. In fact, I only know of 6. Which limits the >>number of total positions available to a very small set. The likelihood of >>finding such a position in that small set is _very_ small to say the least. For >>example, many criticized the h5 move by DB in either game one or two (I don't >>remember and don't have the logs handy). Kasparov said "that was the only move >>to play". Later he questioned the move as being un-computer-like, until someone >>pointed out that some version of Deep Junior would play the same move. So as >>you can see, finding a bad move by DB that other programs avoid, is just as hard >>as finding a good move by DB that other programs can't find. > >I think that finding a bad move by DB that other programs avoid is not so hard >and I remember that I tested it some time ago and all the programs that I tried >could avoid Kf1 in game 2 after some minutes. The question is, do they avoid Kf1 because they see a forced draw? Or do they avoid Kf1 because they probably want to keep the king in the corner? Just keeping it in the corner is not always right. > >You may say that they avoid it not for the right reason because they do not see >the draw but the fact is that they avoid a bad move of DB. I can avoid stepping on a land mine by just not walking down that street. But it doesn't mean I won't step on land mines on other streets, unless I know what to look for and what to do when I see one. That's the point here. With so many programs, many will avoid KF1 just because they don't like the king on f1 as opposed to in the corner. But it doesn't mean they know diddly-squat about the problems moving the king to one or the other squares might cause or avoid. Crafty won't play the h5 move. Yet it has beaten many GM players even at longer time controls (non-blitz). Junior will play h5 and it has beaten GM players at longer time controls. So what do we conclude from that??? that deep blue sucks? :) > >Note that programs can see Qe3 in the pv before finding that Kf1 is bad when DB >could not see it. wrong viewpoint. Qe3 leads to a draw. But if a program doesn't see this, then choosing Qe3 over any other move is just a random chance event that is immaterial. When a program can prove that Kf1 leads to a draw, starting at Kf1 and not later on and backing up, then we will be able to say "that program is better than deep blue in this kind of position." Until then, pick most any move in any game, and you will usually be able to find one or more programs that will play it, or one or more that will not. But unless they "see" what is happening, it is pure luck. > >It tells me that one of the following is correct in the relevant position: >1)programs of today saw deeper than deeper blue so they can see that Qe3 is not >so bad as DB evaluated. >2)programs have better evaluation than deep blue in the relevant position. > 3) that programs of today have an evaluation that simply likes the queen in the center on the opponent's side of the board. Whether the position warrants it or not. (3) is just as likely as any other explanation at this point. >Uri
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.