Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Position-based learning glitch

Author: Stuart Cracraft

Date: 22:24:50 06/13/98

Go up one level in this thread



>the above should work.  However, you said one thing that sounds wrong,
>but you probably wrote it in an unclear way.  You said you write out a
>position to the learn database when the score drops between the last
>iteration and the final iteration.  I assume you mean from one *search*
>to another search?  That's the way I do it.

I suspect a matter of semantics here.

Slate's article in ICCA '87 indicated, as I read it, that if the final
iteration's score is a little more than 3/4 of a pawn (0.78) worse
than the best previous iteration in that same search, then write
out the position being searched, its score, etc. to the learn
file and mark it non-erasable permanent in memory.

For example:

  iteration 1  -100
  iteration 2  -105
  iteration 3  -107
  iteration 4  -200

since iteration 4 is more than 3/4 of a pawn (actually one full pawn)
worse than the best previous iteration (iteration 1), then the position
qualifies to be "remembered".

--Stuart



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.