Author: Robert Hyatt
Date: 05:29:26 08/27/98
Go up one level in this thread
On August 26, 1998 at 13:19:01, Thorsten Czub wrote: >On August 26, 1998 at 12:31:13, Robert Hyatt wrote: > >>Here I disagree. I have versions of Crafty that won't touch Qxg7... the ones >>playing in Paris are good examples... but they don't play that well either, >>because they get over-concerned about king safety... >> >>My interest in such positions would be for programs that don't play Qg7 when >>the attack on the g-file will work, but which do play it when the attack won't >>work. I'd bet that you can move a piece or two around, so that the g-file ends >>up being useless, and these same programs will still avoid taking the pawn. Now >>for the *wrong* reasons... > >In CSTal e.g. part of the evaluation-functions are connected to other parts. >It does not play a move BECAUSE THE G-FILE IS OPEN alone. The evaluation parts >are connected with each other. There must be few more criteria to be >fullfilled... and in this position THESE criteria are realized. > >1.The g-file is open >2.white is developed weak >3.black has bishop pair >4.white has (after g3) a finachetto king-position without having a >fianchetto-bishop, meanwhile BLACK has the bishop AND the queen !! >5.both bishops x-ray on the king. >6.the more pawns black would eat (f7/g7/h7), the more would this > HELP black. CSTal knows this. > >I am sure chris could tell you better how many of those criteria are fullfilled. >It is not ONE alone. > > >>So it isn't a breakthrough to not take such pawns, *unless* it is not taken for >>the right reasons. > >How do you know that cstal is taking it for wrong reason ?! Do you want to see >the main-line and evaluation of cstal win95 after Qxg7 to be sure ?! > I didn't mean to imply that CSTal avoids this move for the wrong reason. My point was that it is *very likely* that any program that avoids this move *is* avoiding it for the wrong reasons, however. IE it is easy to pick a single position where something works well. An example (for me) is win at chess position 2. I solve that instantly because I have specific code to evaluate a pair of connected passed pawns on the 6th rank. This code considers material, how far away the king is, blockading, etc. It took quite a while to get this tuned to an acceptable level, where it worked in nearly all situations, rather than just in one or two. My question would be if we move a bishop, or add a pawn, so that one of the bishops can *never* attack the king field, what happens? IE if you find 10 such positions, and in 9 of then Qxg2 wins, while in only one Qxg2 loses, what then? I'd question the evaluation if it says that in most cases Qxg2 loses due to king safety. > >> Here, I bet that is not the case... And I'm not saying the >>reasons for not taking are all wrong... but when such "intuition" is embedded >>into a program, it often leads to problems... > >Ah. thats what we are working on. We are trying to implement complex evaluation >terms that are related to each other. Like experience in the mind of a human. >And these connections make sure not ONE reason will cause ONE reaction. Only if >some reasons appear, cstal begins to sac or to increase tal or whatever. >But not ONE g-file. I can show you many positions where cstal eats this pawn, >because in those positions it is ok. >In this position it is NOT ok. that's the way it should work, of course... > >> because for every case it is >>wrong, another opponent will push you into such situations and let the error >>lead to a loss... > >Zugzwang lost with such a stupid queen move in paderborn against tiger. I have >quoted the game. Rainer Feldmann told me they have thrown out the queen-malus >for early opening-takes-on b7/g7 pawns. >So - at least for zugzwang a knowledge lack caused it... > >>If anyone is interested, we could take the position, and start subtly changing >>it and see how the programs fare then... > >Ok. If you want i am ready. > >Please PGN (with FEN)... that i can easily test with fritz... :-))
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.