Author: Robert Hyatt
Date: 11:22:14 11/20/99
Go up one level in this thread
On November 20, 1999 at 12:31:11, José de Jesús García Ruvalcaba wrote: >On November 20, 1999 at 10:24:46, Robert Hyatt wrote: > >[big snip] >> >>This can cause serious problems. There is a well-documented search/evaluation >>problem called "the discontinuity problem". If you suddenly switch from one >>engine to another, then at positions around this 'switch' you can make gross >>errors. Or you do like Fritz and have an eval of +1 before trading queens, >>and 0.00 after trading. The evaluation/search has to perform in a continuous >>way, slowly transitioning from opening to middlegame, and from middlegame to >>endgame. I used to have three distinct 'phases'. I am slowly eliminating >>those and making the program perform better. >> >>I don't believe it will be possible to beat GM players consistently with big >>discontinuities in the evaluation, or in the search space. They will notice >>the 'transition points' and take advantage of this. In ways detrimental to >>the program. It has happened to me many times. >> >>An example is king safety. You want to slowly phase it out as you approach >>the endgame, not just <wham> it is off. Because you might reach a position >>where your king shelter is gone, and you give up a pawn to trade away the >>piece that drops the king safety term to zero. And convert a likely win into >>a sure draw/loss. >> >>IE "think smooth" not "jerky"... > >Hi Bob, > strictly speaking, the evaluation function is always continous, as its domain >is a discrete space. Of course I understand what you mean, wanting a "more >continous" evaluation function, i.e., one with less variation. >José. No it isn't... In programs that 'break' on opening-to-middlegame and middle-game-to-endgame, usually based on material, you find large jumps in the eval, with no way to generate intermediate scores. Think about a program with a big king safety term, but this term is turned off when the opponent has either total pieces <=14 (q=9,r=5, not counting pawns). If your kingside pawns are 'gone' your kingside safety score will be way negative. If you can do whatever is necessary to get the piece count <= 14, that scoring term suddenly disappears. that is the wrong way to do evaluations... ie if your king safety score is -2.5, does trading a knight to get the piece count <= 14 _really_ gain 2.5 points of positional score? That is the wrong way to turn things off. IE I slowly phase king safety out as material comes off the board, so that by the time we get to 13 points (not counting pawns) the king safety has slowly phased itself out smoothly, not in one big jump... Berliner wrote a paper about this problem, explaining that around the 'edges' of those 'gaps' bad things can (and usually do) happen.
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.