Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: The End of the Learning Debate

Author: Robert Hyatt

Date: 13:49:36 03/22/98

Go up one level in this thread


On March 22, 1998 at 16:01:00, Komputer Korner wrote:

>What you are also forgetting is How is your big varied book made? If you
>make it by dumping in PGN games and cutting it off at say 60 ply, then
>you are also including a lot of bad lines. How do you weed these bad
>lines out? Crafty has had this problem for the last couple of years, and
>only now is finally resolving it because of it's learning feature.

there are several answers, and I use 'em all.  First, you can cull
moves that were only played once or twice, as those are nothing but
moves from a single game and are not worth considering without a search
to verify them.  I use the "minplay" option set to 3 when I build a book
for Crafty, meaning that in the 330,000 game file I use, if a move was
not played at least 3 times, it is tossed out.

I also pay attention to several things about a book move, and give the
user a chance to indicate the "weight" each factor should influence the
move choice decision...  the weights are (1) frequency, which is nothing
but a count of how many times each book move was played;  (2) win/lose
ratio, (3) static evaluation (which would tend to avoid some book lines
like Bxf3 gxf3 where gxf3 was played as a speculative attacking move but
which shreds the kingside)  (4) learned results.  You can also add a
normal search on top of these if you want.

But the bottom line is, even if you take games between GM players, the
games are filled with tactical mistakes.  And I mean *filled*.  I have
yet to run thru a game that didn't have at least one, and often more,
such mistakes, which means you can't even play GM moves "safely."



>Testing a book at blitz is not the same as testing at longer time
>controls. So the bottom line is that making a huge book is no panacea
>because your program will lose a lot of games from all the bad lines in
>the book. making and adjusting a good opening book is somewhat the same
>for a human.

While I'd partially agree, blitz games *do* eliminate a ton of bad
openings.  Standard time controls will eliminate more, but I'd bet that
blitz will take care of 95% of the truly bad openings and at that speed,
it doesn't take years...



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.