Computer Chess Club Archives


Search

Terms

Messages

Subject: Book test (inspired by book discussions)

Author: Will Singleton

Date: 20:34:15 08/26/02


It's common wisdom that a good book will increase a program's strength, and I
have no doubt about it.  But, what about the typical amateur book, which is
usually created from pgn files, without a great deal of hand tuning?

I don't know much about chess, but I have observed that most programs don't have
a clue to what's happening in the opening as it pertains to the final outcome of
the game.  Chess is just too complex for that.  So, could a crude book, created
from pgn files, significantly help the typical amateur program?

I felt I should conduct a small test involving my program, Amateur.  Its book is
a compilation of generally available large pgn collections, and relies on
won-loss percentages to select moves.  A random factor is applied to choose
reasonable moves.  Against typical competition, it will leave book in 10-20
moves, with a good percentage of viable positions.  In short, it's an OK amateur
book.

I first played a match between Amateur 2.1 with book vs the identical program
without book.  Using a short time-control of 3/1, I was surprised to see that
the results of a 20 game match were about even.  I attribute that to the fact
that the book version was soon out of book after the non-book version played
some non-book move.

Then I played two matches using another program.  Amateur played Beowulf with
and without book.  This was even more surprising to me.  The non-book version of
Amateur lost to Beowulf 2-7-1.  The book version beat Beowulf 6-2-6.

I think one needs to have a larger sample of games to really have an opinion
about the issue.  I am not sure that the book plays a critical role if, upon,
leaving book, the programs have a more or less even eval.  But perhaps that's
deceiving.  Perhaps the won-loss percentages are the better indicator of success
than the program's eval, even after 15-20 moves.  If that's the case, then how
can the eval be adjusted to account for this?  Or should it?

I don't really know how to improve my book, or if the results I got are
significant.  I think my next step is to make a huge book using some of the
commercial pgn collections, then test that against my present book.  But I think
one thing is clear: a really good book is extremely difficult to make, and if
the commercial gui's have very good books, then programs that use them will have
a huge advantage.

Questions for other researchers:  in book testing, how soon is the book program
out of book?  If it is early, and without a significant advantage at that point,
what does that mean for the test?

How does one avoid duplicate games?  I avoided that by having the "non-book"
version play a maximum of 3 moves from book.  This usually results in different
openings, especially given a small test sample.

Will



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.