Author: Uri Blass
Date: 11:43:05 04/11/02
Go up one level in this thread
On April 11, 2002 at 14:10:53, Robert Hyatt wrote: >On April 11, 2002 at 13:03:27, Uri Blass wrote: > >>On April 11, 2002 at 12:43:34, Robert Hyatt wrote: >> >>>On April 11, 2002 at 12:32:09, Uri Blass wrote: >>> >>>>On April 11, 2002 at 09:34:34, Robert Hyatt wrote: >>>> >>>>>On April 10, 2002 at 23:18:42, K. Burcham wrote: >>>>> >>>>>> >>>>>>it is one thing to be able to take each move of game two and game six, and find >>>>>>at least one program that will play the Deep Blue move. >>>>>> >>>>>>it is another to find a bad move that Deep Blue played, and find a program today >>>>>>that will avoid this move and play a better move, and watch the eval climb. >>>>>> >>>>>>what if we all agreed that one certain position that Deep Blue played >>>>>>was a bad move because_______? >>>>>> >>>>>>what if we find a program today that does not play this same bad move? >>>>>> >>>>>>what if we find a program that will play a better move and we can watch eval >>>>>>climb after this move? >>>>>> >>>>>>I know Robert, that in this case you could answer "well if a frog had wings, >>>>>>etc". >>>>>> >>>>>>but i assure you these are honest questions. >>>>>>what if the above did happen, what could we conclude? >>>>>>kburcham >>>>> >>>>> >>>>>The problem is that a game is game... not a series of moves. If you could >>>>>prove that program "X" plays every move in a DB game except move N, and then >>>>>you could prove that move N was better beyond any doubt, then you just proved >>>>>that in _that_ position (only) program X appears to know a bit more. I don't >>>>>think you will ever find a tactical move that program X can find that DB didn't, >>>>>so you are going to be looking at positional stuff only. And proving that >>>>>one positional move is better than another is a _very_ non-trivial thing to >>>>>do... >>>> >>>>The question is how to define tactical move. >>>> >>>>I suspect that all the top programs can find >>>>Kh1 instead Kf1 in game 2 after enough time and it seems that >>>>Kh1 is the winning move when Kf1 is a draw. >>>> >>>>You can say that they find it for positional reasons >>>>but it does not change the fact that they play the better move. >>>> >>>>Uri >>> >>> >>>Can the find it by discovering that Kf1 is a draw? >>> >>>Not a chance... >>> >>>So it is based on serendipity... which is, of course, true for a lot of >>>moves played by computers... >> >>I do not understand what is serendipity and I have not a dictionary >>now but it is based on knowledge that deeper blue had not. > >Serendipity means "something found by chance". "Lucky" in other words. > >_no_ program has "knowledge" that can indicate that Kf1 leads to a draw while >Kh1 does not. This is a search issue, not a knowledge issue, period. From a >general heuristic point of view, Kf1 is better. It centralizes the king, while >Kh1 puts it on the worst possible square the king can stand on, all things being >equal... > >So a "smart" program will probably like Kf1 until search shows the flaw with >it... No A smart program will understand the chances to draw by evaluation even without seeing the forced draw so it is going to prefer Kh1(both moves have negative evalation but Kh1 has more negative evaluation). It is exactly the case with the top programs. They see the line Kf1 with Qe3 and change their mind from Kf1 to Kh1. > > > >> >>It can see enough problems with Kf1 to prefer Kh1 > >This is not knowledge. this is search... This is search and knowledge because I assume that it can see problems with Kf1 but not enough problems to see the repetition but only enough problems to evaluate the line with Kf1 and Qe3 as worse than the line of Kh1. > > >>(I mean seeing Qe3 in the main line before changing it's mind to Kh1 >>or seeing another move for white to prevent Qe3). > >Qe3 is not a problem. Positionally. It is a problem tactically. Knowledge >has nothing to do with finding any of this. A program either sees the draw >and discards Kf1 because of that, or else it plays Kh1 for some other reason >that serendipitously happens to be good... No it is possible to see by search that is not deep enough in order to see the draw a position after Qe3 when black has 2 pawns and some attack for the piece. If the program knows to evaluate that position as worse for white relative to the position after Kh1 then it is going to prefer Kh1. 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.