Author: KarinsDad
Date: 08:21:15 08/21/99
Go up one level in this thread
On August 21, 1999 at 10:56:13, Robert Hyatt wrote: >On August 21, 1999 at 04:47:36, Rémi Coulom wrote: > >>Hello, >> >>I spent some time trying to write code to evaluate pawn races in TCB, and I have >>a few comments and questions for programmers. >> >>First, I took a look at the source code of Crafty. I was surprised to find that >>the code can fail in many cases and evaluate that a passed pawn can promote when >>it can not. My guess is that the gains can outweight the problems, but I tend to >>believe that "If you have something that can make a chess program play a stupid >>move, it will happen, and it will probably be during the World Championship". I >>made up a few positions that trap Crafty's evaluation. They are all cases when >>it underestimates the opponent's ability to promote first. It can happen : >> 1. If the opponent has a check that wins a tempo > > >I don't follow. I don't do the pawn race if the opponent has a piece left >(pawns only to trigger the test.) In that case, how can a check win a tempo? - = - = - = - = = - = - = p p - P = k = - = - p = - = - = - = - - = - P - = - = = - = - = - = - - = - = - P P P = - = - = - = K Simple, d5+ pulling the king out of the square wins. If a7, then black might draw. > > > >> 2. If the opponent has "hidden passed pawns", which are not taken into >>consideration > >I eval these, but I depend on the search to 'expose' the passer which >happens. I haven't seen this as a problem in any game so far. > > > >> 3. If the opponent has two passed pawns and the King is in both's square > >I don't worry about this one. again I depend on the search. With nothing >but pawns, 20 plies is easy to reach, and that is usually enough to resolve >things. > > >> 4. If the opponent can protect its passed pawn with his King >>I appended four analysis logs that show these problems. I am sure there are many >>other kinds of problems that can be found. > > >You have to make a conscious decision about what you are going to do in the >eval, and what you are going to do in the search.. If my eval says "white >wins" it is generally correct. If it says black wins, ditto. Yes, there are >cases where it says neither, and is wrong. But to resolve those, I depend on >the search to find positions where the eval will be right. > Quite right. If we had perfect evals, we wouldn't need to search (and if we were smart enough to code near perfect evals, we could probably play fairly good chess ourselves). KarinsDad :) [snip]
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.