Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: offline opening book learning

Author: Dave Gomboc

Date: 11:32:26 06/22/99

Go up one level in this thread


On June 22, 1999 at 13:17:35, Jay Scott wrote:

>
>On June 22, 1999 at 03:45:51, Dave Gomboc wrote:
>
>>Keep track of the computed score for a position separately from the backed-up
>>score for a position.  If they differ significantly, an important alternative
>>was overlooked.
>
>You're suggesting the obvious algorithm, almost exactly what I suggested
>some months ago.

It predates "some months ago" by a long time.  I am sure people thought of it at
least 10 years ago.

>This bit is not quite right. The scores might differ solely because of
>the effective search horizon. For example, after 1. e4 e5 2. f4, suppose a
>normal search evaluates 2... exf4 as +100 for black, or close. Backing up deep
>values from the book might give the move an evaluation around 0--and yet
>still identify 2... exf4 as the best move. There's a big score difference,
>but it doesn't have anything to do with overlooking an alternative.

The alternative overlooked can be later.  Alternatively, the computation ended
before it should have.  It's a heuristic, nothing more.

>The most powerful way seems to be this:
>
>0. Having backed up all scores below a node,
>1. load these scores into the hash table, marking them as exact values
>so that the current search won't try to look beyond them. (They're
>based on a deeper search, so this search won't be able to see anything new.)
>2. Search from the node.
>3. If the best move is not already in the book, the program has found
>a novelty. This must be corrected immediately: add it to the book.

Overspecified, but it will work fine.

>A more intensive novelty-search would find all moves within, say,
>half a pawn of the best move. Just use the "Next Best" feature until
>you've got 'em all. For the Kasparov effect, repeat the novelty search
>on the novelties until you have a whole subtree of analysis ready for
>your next opponent. The new scores float up the tree and may change
>the best move at the original node--the ideal is to discover that the
>novelty that seemed at first to be second-best is actually quite strong.

Even the second-best ones are fine.  As a rule, you only get one game with them
before they're not novelties anymore.  As long as it gives the opponent
something to think about and a chance to go wrong, you're ahead.

>Offline book learning is *way* more powerful than the online book
>learning that chess programs use now. Here's a paper about a simple
>method (used in an othello program):
>
>"Toward opening book learning" by Michael Buro
>compressed postscript, five pages
>http://www.neci.nj.nec.com/homepages/mic/ps/book.ps.gz

I've read it.

>  Jay

Dave



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.