Author: Gerrit Reubold
Date: 10:45:24 01/23/01
Go up one level in this thread
On January 22, 2001 at 16:02:56, Robert Hyatt wrote: <big snip> >The problem is endgame knowledge. A program _ought_ to know that if you have >a passer, then trade pieces to reach a won ending. Only in this case, that >heuristic back-fires as it is black who ends up winning. This is a _tough_ >exception to handle... > >although a GM would tell you instantly "No I won't trade queens..." Hello all, I think stating that a GM would know instantly that trading queens lose is misguiding. One tempo decides whether the queen exchange loses or wins. Both a GM and a chess engine should calculate here and not simply trust their evaluation. Of course, "tell you instantly" could mean that the GM did the calculations unconsciuosly and super-fast, but I don't think it was meant this way. Consider the following changes to the position after blacks 45.th move: (a) if the black a-pawn were on a6 instead of a5: 46. Qe6+ wins, but it is not obvious. (b) if the white King were on g2 instead on h1: 46. Qe6+ wins, even less obvious. I hope my analysis is OK, I have no chess board or chess engine at hand to check it, but the variations should be simple (if I remember my analysis of yesterday night correctly). Of course it would be nice when the evaluation "sees" that the black b-pawn is dangerous, but IMO the search *must* verify that. Too much guessing might be harmful :-) Maybe the evaluation should guide the search, e.g. when the evaluation sees that there is a dangerous passer in a pawn endgame, it could tell the search to extend moves of this passer. It would be easy to extend this to candidate passers, or to the candidate passer's neighbour pawn, or ... (BTW: in the middlegame: when the eval detects king safety problems, it could guide the search to exploit them... nice!) The "eval-guides-search"-idea is certainly not new and more likely than not just one of the ideas which *don't* work, but I think I will try it in my engine. Now it becomes slightly off-topic: That "eval-guides-search"-thing seems to be the way humans solve this kind of position, they know that the b-pawn is dangerous and don't care about black moves other than moving the dangerous pawn and its helper (the a-pawn). You have surely seen humans "analyse" these positions: they count 1 (a4), 2 (b3), 3 (b2), 4 (b1=Q) and since white needs more than 4 moves to promote a pawn, black wins. Hope this makes sense, Greetings, Gerrit
This page took 0.01 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.