Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Rebel book learning

Author: Stuart Cracraft

Date: 06:43:39 06/17/98

Go up one level in this thread



On June 16, 1998 at 12:40:29, Enrique Irazoqui wrote:

>On June 16, 1998 at 10:48:59, Moritz Berger wrote:
>
>>On June 15, 1998 at 18:56:12, Enrique Irazoqui wrote:
>>
>>>Rebel was unable to save its games and therefore to
>>>"learn", and it fell over and over on the same (not cooked) lines.
>>
>>Is it true that Rebel 9 has to save its games in order for any book learning to
>>take place?
>
>Yup.
>
>Enrique

Ugh!

The canonical implementation is to note which position had a very bad
final iteration as compared with the best previous iteration in that search.
If the difference is greater than about 3/4 pawn (program-dependent),
the position, score, alpha and beta bounds, and depth (and in my case
a visual representation of the position) is written out to a plain text file,
and possibly even a game record of the game so far.

Then, it is easy to see what the program is learning, why it is getting there,
and to adjust earlier positions or add them as well to move the program
further away from the bad position earlier.

--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.