Author: Robert Hyatt
Date: 18:52:49 12/15/05
Go up one level in this thread
On December 15, 2005 at 21:01:00, Brian Richardson wrote: >On December 15, 2005 at 20:03:47, Sergei S. Markoff wrote: > >>>Really? Just the line from fruit:eval.c: >>> >>> eval = ((opening * (256 - phase)) + (endgame * phase)) / 256; >> >>No. Not just this line, but the concept and well tuning. >> >>>I find that VERY hard to believe. That concept has been around a >>>very long-time. >> >>Really? Where it was introduced by the first time? > >Don't know about the first time, but GNUChess (v4) uses the same approach, >albeit only for some king eval terms, and many programs scale things by >material, although not necessarily as nicely as Fruit does. Most all programs use something similar. You can't afford to have discontinuities in the evaluation or when you are in positions that are right around the discontinuity, you will get some really bizarre results as the search will find ways to creatively cross over the discontinuity when it seems favorable, to produce a big score change over nothing. Crafty has lots of terms that "scale" by material. Kingsafety drops off as material goes away. Endgame features (pawn majorities, outside passed pawns, etc) scale up as material goes away. This is not new... I think Berliner first mentioned this "problem" 25 years ago in some paper or another... > >> >>There was a lot of material-dependent evaluatation factors, but its was tuned >>independently. The result for the most engines that used this approach was bad >>consolidation ability. >> >>>We need specifics. I see nothing special about Fruit 2.1 evaluation >>>to explain the strength. >> >>Lets experiment with it. It's easy to remake Fruit search to make it same to the >>Crafty's.
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.