Author: Robert Hyatt
Date: 20:32:05 03/16/99
Go up one level in this thread
On March 16, 1999 at 17:19:47, Christophe Theron wrote: >On March 16, 1999 at 16:30:59, Pat King wrote: > >>A number of my evaluation terms get updated at every node... Force, Space, >>CtrCtl, Material. So all of them are available at the end of QSearch if I want >>them. My question is, are the terms other than Force worth anything after >>QSearch? If so, how might one weight them differently than in the full blown >>evaluation? >> >>Pat > >Every positional term is important during QSearch. If you don't evaluate them in >QSearch you can miss very important positional exchanges, especially queen >exchanges or important pawn structure changes. > >You can maybe use lazy eval to avoid evaluating everything. But this is risky >anyway. > > > Christophe Lazy eval doesn't _have_ to be risky. IE in Cray Blitz I did this and there was _no_ risk at all. Because I knew _exactly_ the max error that a positional score might have. I just incrementally kept up with an 'error term' that said, for each piece on the board, how big or small the eval could be. And as I made moves, I updated the total error with the value for a moved piece. Then I knew _exactly_ how big the lazy eval window was because I _knew_ how big the positional score could be, at max, for a given position, because I knew the max score for each piece. That lets you have lazy eval with zero error, so long as you don't have positional scores so large that the error epsilon gets too large, so that your lazy eval has to be executed fully most of the time. I don't do this in crafty because I have some _big_ positional scores. But I do have a 'sorta-lazy' eval, where the big terms are done first, and then I may bail out before I do the smaller (but way more costly) terms.
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.