Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Book learning?

Author: Robert Hyatt

Date: 07:42:35 12/31/98

Go up one level in this thread


On December 31, 1998 at 07:38:05, Inmann Werner wrote:

>Hello
>
>Can anybody tell me how book learning works?
>The principle should be, mark good moves as good and bad moves as bad.
>But how can a program determine, what was a good or bad move?
>
>Werner


There are two answers, and I use both in crafty:

(1) if you lose a game, find the last book move where you had a choice of
moves to make and mark the one you played as "never play again."  This is
the easy case and is called "result learning" in crafty.

(2) take results from the first few searches after you leave book (I use 10
moves in Crafty) to get a feel for how the program likes the position.  From
these 10 search results, plus the rating of the opponent and the search depth
of the moves, I compute a "learned value" that represents whether this position
is good or bad, while factoring in the strength of the opponent and how long the
search had, and then I apply this score along the book path working backward.
At the first point where I had a choice, the move there gets the entire score
added in.  This score is then divided by the number of possible book choices at
that position and this modified value is backed up further, getting divided
every time there is more than one choice...

Works pretty well, based on several hundred thousand games...



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.