Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: What's the difference between "book learning" and "autolearning"?

Author: Peter Fendrich

Date: 02:10:02 06/19/01

Go up one level in this thread


On June 19, 2001 at 02:15:48, Dana Turnmire wrote:

>  I was reading this review of Genius 5.
>1) Genius is now a learning program. You can turn on the autolearn feature in
>the options menu and Genius will save the critical position, from every game it
>loses, in a file called G5AUTOL.dat. Genius will only learn if it had thought
>more than 30 seconds on that position during the game.
>24) Genius has no book learning feature.
>
>  Is one learning feature superior to another?  Do the current top programs use
>the same type of learning features?

I don't know any details about Genius but I suppose that the "autolearning" is a
variation of what I in this post call "outside book" learning.
Both types of learning that you mention are of the the same type, called Root
learning. The basic principle is to save information about positions (or moves
if you like) for future use. There are many ways to implement this but one
obvious method is to use the outcome of the game and abandon positions that lead
to a loss and maybe promote postions that lead to a win. The main difference
between book learning and "outside book" learning in this perspective is that
the position is generated from the book in the book case and generated by the
programs search function in the other case. In both cases you store information
about the generated positition. Of course you have to addapt the implemantion to
suite the representation of the book and book moves when using book learning.

A standard implementation for the "outside book" learning is to save the
position in hash table form. For instance store the search score 10 moves later
in the game. Next time that position will be loaded in the hash table 10 moves
earlier with the new search score.

I have myself tried to back up values from critical positions backwards in the
list of game moves. I don't know it's better or worse than the standard
implementation but it is certainly more complicated.

Other types of learning (in contrast to root learning) can be of higher order
and can't be stored in positions as used in root learning. For instance to learn
that a knight should, as a rule of thumb, be developed before the bishop.

I hardly believe that any top program uses higher order of learning but many (or
most?) of them uses different variants of both book and "outside book" learning.

//Peter








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.