Author: Robert Hyatt
Date: 17:22:59 08/18/05
Go up one level in this thread
On August 18, 2005 at 18:56:57, Uri Blass wrote: >On August 18, 2005 at 17:44:03, Robert Hyatt wrote: > >>On August 18, 2005 at 16:03:31, Vincent Lejeune wrote: >> >>>On August 18, 2005 at 14:09:08, Robert Hyatt wrote: >>>... >>>> The Sjeng game was simply something overlooked >>>>months ago that was easy to fix. This game had a book problem, but even after >>>>the g5, I wonder if Crafty could have played it better. I'm looking now... >>> >>>Could you give more informations about this "old bug", please ? >> >> >>Yes. Somewhere way back in the 19's, I removed the last bit of "root >>pre-processing" code. One of the pieces was a bit of code that initialized the >>pawn piece/square table from one of three different sources, depending on >>whether the games was in the opening, middle-game or endgame phases. Somehow I >>removed the preprocessing piece, but never pre-initialized the piece square >>table to slightly encourage pawn advances in general. As a result, there was >>_zero_ incentive unless it was pushing a passed pawn, or making a passed pawn. >>If you look at the Sjeng ending, it was a dead drawn position, but crafty just >>sat back and let Sjeng advance the pawns until you reach one of those positions >>where black can then make a passed pawn and it is closer to promotion than >>white's resulting passer. After initializing (by hand) the pawn piece-square >>table to values similar to 19.0, I replayed the game and Crafty varied so >>significantly I decided to play a lot of test games to see if it was "better" or >>"worse" before the changes. It was worse. :) So I left the changes in since >>they had effectively been tested for several years prior to their unintentional >>removal... >> >>Crafty version 20.0 is going to see the start of a major revision of the eval >>code to get it cleaned up. Too much overlap, and interaction. For example, >>blockading a passed pawn. I do it in one place, then give less of a bonus for a >>rook behind a blockaded passer somewhere else, which is duplicated work that >>could be avoided. Probably lots of unintended side-effects scattered around all >>over so it is about time to re-do this and clean things up completely... > > >If I understand you correctly you say that Crafty after the game against Sjeng >is without the bug of not advancing pawns and probably better than Crafty before >the game against Sjeng. > >Do you think that removing the last bit of the root pre-processing" code may be >the reason that Crafty19.17 seems to be weaker than older versions? > >Uri that will take some "diff'ing" to see where it happened.. I'll look..
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.