Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Too Large Positional Contributions to Score

Author: Robert Hyatt

Date: 06:55:21 03/08/99

Go up one level in this thread


On March 07, 1999 at 23:29:31, Stuart Cracraft wrote:

>Crafty seems to have a Scale_Up/Scale_Down factor and I'm curious
>if Bob would talk a little about this. Seems costly to add a memory
>reference and a division (?).
>
>For example, seems like Bob has chosen to ScaleUp/ScaleDown fairly
>large contributions such as
>
>   weak back rank
>   bishop pair
>   certain rooks on open files
>   rook on 7th
>   protected passed pawns
>
>based on the side's material.
>
>Are there alternatives to controlling the "bloating" of the positional
>part of the evaluation that aren't so computationally expensive? I know
>there are very few of these compared to all the terms in Crafty so the
>cost isn't a big deal (probably.)


these are not attempts to reduce the positional scores... they are an attempt
to map them to where they belong.  IE the bishop pair is _far_ more dangerous
in simple endgames than it is in the middlegame.  Outside passed pawns (you
missed that one) are far more dangerous in the endgame than in the middlegame.

Other things (king safety things like weak back rank) are scaled down as
material comes off because they become less important, until king safety is
finally scaled all the way to zero at some magic 'endgame' threshold.

But you overlook the key:  these things simply teach crafty whether or not to
trade material, without using the very 'coarse' trade when ahead in material,
don't trade when behind.  It now knows that once it creates an outside passed
pawn, that every trade makes that pawn more valuable.  Or that if it has a
weakened kingside, every trade makes it less 'weak'.

>
>One I considered was going from centipawns to millipawns and then working
>backward looking at just those things that should be big (king safety)
>and sizing them accordingly letting them bloat to a certain degree beyond
>a pawn.
>
>--Stuart

What I'm doing is not 'limiting' the scores at all.  It just shows that simple
scores are not enough... that there are interactions.  IE the new king safety
stuff I am still working on includes piece interactions rather than being just
a simple 'get a piece closer to the king.')  But this is one of the things that
gives crafty a fairly decent endgame ability... knowing when to head for the
endgame and when to stay in the middlegame...  That can be important..



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.