Computer Chess Club Archives


Search

Terms

Messages

Subject: learned book

Author: James Long

Date: 09:45:19 02/24/99



I've started implementing a "learned book," one that
the computer learns entirely on its own.  I'd be
interested to hear the results from others that have
tried similiar experiments...

Basically, the program has three "book modes;"
learning mode, tournament mode, and random mode.

On ICC, the default will be "learning mode."
In learning mode, the program probes the book file.
The moves that have a book entry are _removed_ from
the root move list, since they've been learned before.
A normal search is done with the resulting moves.
This process repeats itself until the move counter
reaches <book_depth> moves, probably 40-60.

At the end of the game, the program replays the first
<book_depth> moves, doing a fixed depth search on each.
The resulting score, depth, and program version are
stored in the book file....

In the (rare) event the program doesn't find any
unlearned moves at the root, it plays the move with
the oldest entry, or searched by the oldest version of
the program.

The drawback of this method is the obvious blunders the
program will play at times just to get a learned value.


I think that repeatedly learning from the same opponent
could be counter productive, so I'd probably put a switch
in that says "if you've played this opponent XX times already,
switch to another mode (tournament or random)."

In "tournament mode," the program would play lines that have
yielded the best learned values, with some degree of
randomness to avoid repeating games.

I hope this isn't too confusing... I'm still thinking over
the details.

Again, I'm interested in any comments, or results of
similiar experiments.

--
James



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.