The Halloween-Attack in the Four Knight Game ---- Fritz5 vs. crafty match ---- (C)opyright 1996-1997 by Steffen A. Jakob

Re: ‡‡‡‡‡‡‡‡ ---- Fritz5 vs. crafty match ---- ‡‡‡‡‡‡‡‡

From           hyatt@crafty.cis.uab.edu (Robert Hyatt)
Organization   CIS, University of Alabama at Birmingham
Date           14 Aug 1997 12:59:53 GMT
Newsgroups     rec.games.chess.computer
Message-ID     <5suvg9$as5@juniper.cis.uab.edu>

brucemo (brucemo@seanet.com) wrote:
: Lonnie wrote:

: > Being the impetuous fool that I am I decided to play some of if not all of the
: > games today (Wed 4:30PM EST).

: When I saw this I logged onto FICS and watched.  I got there in time to see the
: end of the 4th game, which brought the score to +2 -0 =2 from Crafty's point of
: view.

: In the 5th game I saw Crafty having some problems, but it managed to get a couple
: of advanced passers and win.  Crafty has become quite speculative regarding pawn
: structure.

: After the 5th game I had to leave.

: When I came back, I saw that Crafty had won four out of the last five games,
: bringing the score to +7 -1 =2 from Crafty's point of view.

: I spoke with Lonnie, and apparently it wasn't all Fritz 5.  The game Lonnie won
: was with Hiarcs, and I think he did a couple with Nimzo as well.  Crafty won all
: five games where it had white.

: So Crafty got four or five wins and one or two draws against Fritz 5.

: Some speculation was heard as to whether Bob had gotten some Cray time :-)

: bruce

I apparently caused a bit of confusion for this match.  I was logged on FICS
when Lonnie popped in and got the first game underway (roughly a 30 second
per move game).  I turned Crafty's "whispering" on so everyone could see
the game and crafty's analysis at the same time.  I then left to go to
class.

When I got back, there was quite a discussion going on, with a dozen or so
watching the match.  Lonnie was convinced that "something was fishy" for
several reasons.  The most pronounced one was that crafty mysteriously stopped
whispering after game one and didn't resume.  For anyone with the source to
crafty, if you check out "option.c", in the "level" command, you will notice
that Crafty disables whispering if either (a) base time is > 5 minutes, or
(b) increment is > 3 seconds.  I do this to avoid cheating, because I have
caught several humans playing crafty and logging in as guests in a second
window to observe and see what Crafty is thinking about doing.  For faster
games, this is hard to do and not lose on time.  In any case, lonnie was
playing 2 12 and the "12" caused whispering to be turned off.  The confusion
came because I manually turned it on in game 1, but then left to go to class
and didn't return until game 8 had started.

The results were interesting, and I won't try to explain them because I
can't.  It ended up with crafty winning 7, losing 1, and drawing 2.  The
first game was a draw (the one I had whispering on in) and in that game,
Fritz appeared to be winning (crafty was at -2 for a good while, although
it was one real pawn down.)  However, it reached a KQPPP vs KQPP ending
where it was a pawn down, but had an outside passer.  With queens on, this
looked like a draw, but fritz had chances with the extra pawn (it had 3
connected pawns on f,g and h files, crafty had pawns on g and c files.)
However, Fritz decided to trade queens, and crafty's score popped to almost
zero, even though it was a pawn down, because it knew the outside passer was
going to cost white time.  It did.  It appeared to be a victory for knowledge
to end in a draw.

The games Bruce watched I didn't see.

The next point that seemed to excite Lonnie was Fritz playing white and it
hit Crafty with the Halloween opening.  the first few moves were: e4, e5,
Nf3, Nc6, Nc3, Nf6, Nxe5

Here Crafty normally responds Nxe5 from the book.  However, "brause" on
ICC plays this opening all the time (it is a crafty clone) and the operator
tries it against Crafty all the time.  This caused two things to happen.
(1) crafty's book learning decided that 4. ... Nxe5 is not so good to play;
(2) crafty's position learning has learned a *lot* about this opening after
the position above.  In this game, then, crafty rejected Nxe5 as a book move
because the learned result was bad.  It did a search, and still found that
Ne5 was the best move, after wasting 21 seconds to reject the book move and
finding the same move by searching.  Lonnie wasn't happy as he concluded this
was a different book (and, therefore, a different crafty).  In reality, if you
log on to ICC and chat with "yobes" you will find the real guilty party here.
:)  He's simply "trained" crafty quite well, and lonnie forcing fritz to follow
this line simply backfired, and crafty quickly beheaded Fritz quickly and
painfully.

For anyone interested in the games, I have both the original game scores,
and the complete log files from Crafty.  This was played by version 12.7,
which is *exactly* the same as the one released yesterday.  It should be
easy to take the log files from the games and convince yourself that this
was a P6/200 running the same crafty that's on my ftp machine.

I had not planned on discussing this, but the discussion on FICS was so
loud and long, that every time I logged on last night, either on FICS *or*
ICC, I was immediately asked about what was going on.  I can't give you an
impression of Fritz 5 yet because I didn't watch enough of the games, and
apparently Lonnie dropped Fritz near the end of the match anyway, although
I didn't know this until Bruce said something.  If you'd like to see the
logs, let me know.  I can post 'em or put 'em on the ftp machine.  Then you
can decide.  Note that you will have trouble producing the same moves in the
halloween without my position.bin file, but I'll be happy to put the position.dat
file on the ftp machine, which should solve that easily using the import command.

The games are still on ICC in Lonnie's history if you'd like to log on and
have it email you pgn game scores.  If you have any questions, feel free to
ask.  I can tell you that there is not at present a "Cray Crafty" unfortunately,
as I have not yet gotten around to the parallel search.  Soon.  But not yet...