Author: Robert Hyatt
Date: 18:00:53 10/28/97
Go up one level in this thread
On October 28, 1997 at 16:35:44, Randolph S. Baker wrote:
>
>On October 28, 1997 at 15:55:16, Chris Whittington wrote:
>
>(stuff pruned)
>
>>>>
>>>>Also Franz Morsch is not happy with his opening today - it was some of
>>>>repetition with a knight on c6-a5 repeating. Apparently the program came
>>>>with a little minus out of book, so Fritz repeated, this despite wanting
>>>>and expecting to win against Anmon. Big opening books don't always help.
>>>>
>>>>
>>>>>Chris Whittington
>>>
>>>
>>>What is happening with Fritz anyway!? Four draws and a loss, while
>>>everybody
>>>expected Fritz to be a contender for first place..... Does Thorsten
>>>have more
>>>information about this?
>>
>>I'll try and find out tomorrow morning. Or else Ingo (who has kindly
>>volunteered to take on my role as of tomorrow) can get the phone passed
>>direct to Frans and ask him straight ...
>>
>>Chris
>>
>>
>>
>>>
>>>Large books are very bad in tournaments like this, IMO. The variation
>>>you mention
>>>could be a Two Knight's Defence, in which White plays Qb3, Black ... Na5
>>>and then
>>>Qa4+, Nc6 Qb3, Na5 and so on, with a repetition of moves. I am curious
>>>to see
>>>this game.
>>>
>
>How do other programs (e.g. Crafty) have opening books generated for
>them? I don't know if Fritz5 is using the new Powerbooks at WMCCC, but
>the new Powerbooks are essentially automatically generated from a large
>set of high-quality games as opposed to being specifically tuned for a
>machine. (Although Fritz5 can adjust the weights for various opening
>lines based on its win/loss experience).
>
>It strikes me that what works for world-class players need not
>necessarily work for computers. If I understand the Fritz tree, however,
>Fritz only plays lines which resulted in a significant number of wins
>and were played in a statistically significant (whatever that means 8)
>number of games. This would seem to be a sensible approach.
>
>Ramdy
>>>Jeroen Noomen
I can only answer for Crafty, but here's what I did/do: I took about
1/4 million games from the various servers, ran 'em thru a database to
cull duplicates and make sure at least one player was a GM (this is an
imperfect test as I used last names, but there are computers with GM
last
names, unfortunately)
I then take this and run it thru crafty's book create code which was
told
to cull any position that wasn't reached at least 3 times. This left us
with about 130,000 positions from the original 16.6 million positions in
the
raw book data.
Then we do several things in selecting lines:
1. If a line has "zero" wins, it is culled, period.
2. If a line is played a small fraction of the time, it is culled,
period (I use < 20% of the number of games for the most popular move
as a cutoff, at present).
3. If, during a game, we reach a position where the least popular move
we
consider was played < 20 times, we do a fairly short search on the
set
of book moves to be sure we aren't walking into some tactical trap
that
the opponent overlooked in the real game.)
4. For selected opponents, we search *all* book moves for the one that
produces the best score. This led to the famous Ruy Lopez - Crafty
variation last year (1. e4 e5 2. Nf3 Nc6 3. Bb5 a6 4. Bc4!! :) )
This turned out well as it took everyone out of book, yet the moves
for the opponent are fairly obvious letting crafty stay in book for
25 moves or more. Worked well, although it was a total accident.
the "search" idea is somewhat dangerous as Crafty will obviously take
most sacrifices, which is not a good idea, although against a computer
it
is safer than against a human.
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.