Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: developing Junior (and other pro programs)

Author: Robert Hyatt

Date: 16:32:58 09/02/02

Go up one level in this thread


On September 02, 2002 at 15:53:20, Gian-Carlo Pascutto wrote:

>On September 02, 2002 at 13:12:33, Robert Hyatt wrote:
>
>>Yes.  but suppose the positional scores are +/= three pawns, as I have seen
>>out of fritz?
>
>So...what's your point? What do the positional scores have to do with this?

Where have you been during the discussion?

To back up and repeat again...

mtd(f) does poorly when an evaluation can produce a signficant "swing" in
the score after only one more move is made on the PV.  IE a big variance
due to king safety, pawn majorities and potentially won endgames, or whatever.

You _want_ the N+1 iteration score to be very close to the iteration N
score, unless material is being won or lost.  That produces the fewest
researches.  If the scores change drastically from one iteration to the
next, mtd(f) is going to be a loser overall.

And _that_ was the point about fritz.  It doesn't matter about the absolute
value of a pawn...  when a program can produce positional scores of +/- 3.00
in lots of positions.  That is enough to cause mtd(f) to produce very large
trees...

_that_ is what I am talking about.  It is _far_ less common to win or lose
material in a search than it is to win or lose a positional edge.  And if the
score for that edge is significant, then mtd(f) has problems.




>
>>And of course nobody would do 250 searches, there are better algorithms since
>>fail-soft gives you a better bound for the next attempt, rather than simply
>>adding one if you fail high...
>
>The effect still plays.
>
>--
>GCP



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.