AEGON Tournament - Final Round
Fabian Maeser
==========================
Zarkov 3.0 - Nunn 1/2
Christiansen - Mephisto Berlin 1-0
Hitech - Vaganjan 1/2
Tasc R30 - Cifuentes 1-0
Hort - M-Chess Pro 1-0
Compl. Chess Sys. - Ligterink 0-1
Van Voorthuijzen - Gideon 2 0-1
Bronstein - Tasc R40 1-0
Joop Piket - Hiarcs 0-1
Final Standings:
----------------
Men: 114 - Computers: 114
1.-2. Christiansen, Nunn 5.5
3.-4. Bronstein, Van Voorthuijzen 5.0
5.-9. Gideon 2, Hort, Ligterink,
Vaganjan, N.Kuijf 4.5
...
Source: Dutch Teletext
Unfortunately, they don't give the complete standings...
Anyone got more information on the "Gideon" program ?
febi
Robert Hyatt
This is AKA "the chess machine" which is the ARM risc processor
board that plugs into a PC. The board includes a very fast processor
and software that the pc program "downloads" into the special-purpose
board. Runs like the blazes and plays well.
--
-------------------------------------------------------------------
Robert Hyatt Computer and Information Sciences
hy...@cis.uab.edu University of Alabama at Birmingham
(205) 934-2213 115A Campbell Hall, UAB Station
Jean-Christophe Weill
>>
>>Unfortunately, they don't give the complete standings...
>>Anyone got more information on the "Gideon" program ?
>>
>>febi
>This is AKA "the chess machine" which is the ARM risc processor
>board that plugs into a PC. The board includes a very fast processor
>and software that the pc program "downloads" into the special-purpose
>board. Runs like the blazes and plays well.
Not more true, now Gideon is a PC program. I think it was running
on a pentium at this tournament...
-- Jean-Christophe Weill
co-author of Joker and Ecume at the AEGON tournament.
Robert Hyatt
Sorry. Last time I encountered Gideon (ACM computer chess tournament
in 1993) Gideon was running on a laptop with the ARM processor board
installed therein. Haven't seen/heard anything since then so I assumed
it had stayed roughly the same....
Mark Mittelstaedt
-> that plugs into a PC. The board includes a very fast processor and
-> software that the pc program "downloads" into the special-purpose
-> board. Runs like the blazes and plays well.
Bob when is the last time CB went out to play? I am interested in
seeing how it will do with its new found strength.
Mark Mittelstaedt
Robert Hyatt
The last "real" event we played in was the 1993 ACM event in Indianapolis.
We finished in 2nd place. We really should have won the event, but we
managed to draw an easily won game when our "prototype C90" had one
processor go South and forget how to do integer multiply operations
and produce the correct sign. As a result, we watched the chess
board go from a normal position where each side had two bishops to
a position where one side had three and the other side had one (with
no captures or promotions!)
We had to stop the program and re-start after each and every move
to "limp" through that game. Once we found that it was a real machine
failure (this is HIGHLY unusual, so we didn't even consider it and
thought that we simply had another of our parallel processing bugs that
always seem to crop up at unexpected times...) we started using only
one processor, but still had problems as the UniCos operating system
would still bounce us from cpu to cpu, and when it bounced us onto the
"bad" one, zappo! When you think about it, we drew a game where we
were a piece up. We could easily have lost, just by putting our queen
on a square defended by "our" bishop that had changed it's color...
In any case, we are planning on attending the upcoming ACM event in
June if we can get machine time arranged. Cray is working on it for
us as we speak. As to it's "new found strength," I am keeping an open
mind. I recall that for every major leap forward, first there were
many small steps backward. We hope we are better, but this event
will give much more insight...
Mark Mittelstaedt
-> we managed to draw an easily won game when our "prototype C90" had
-> one processor go South and forget how to do integer multiply
-> operations and produce the correct sign. As a result, we watched the
-> chess board go from a normal position where each side had two bishops
-> to a position where one side had three and the other side had one
-> (with no captures or promotions!)
I have confidence that with your Cray's Power and your programming you
will finish #1 this time around!
Robert Hyatt
We wish. Just remember, if you consider the cray a "powerful" computing
platform, then Deep Thought could be thought of as a Cray on steroids....
My analysis (biased, of course):
Tactics: I believe that we are tactically stronger than any other single
program with the sole exception of Deep [*]. Tactics is the "easy" part of
computer chess, with machine speed being one large constraint. Running at
around 4 billion instructions per second on a C90, we can afford to extend
lines of play that a micro-computer simply doesn't have the horsepower to
look at without drastically reducing their overall search depth. For example,
we have probably given up a full ply of search over the past two years, in
order to search "interesting" lines deeper. Even if these extensions are
completely wasted in certain positions, we can still search deep enough that
we avoid most trouble.
Positional: Here, I suspect that we are not at the top of the heap, although
we are significantly better than in years past. The reason: it takes many hours
of testing over a wide variety of positions to properly tune an evaluation term.
It is even harder to add additional knowledge (it is also hard to simply
determine what additional knowledge *might* help.) In short, the more time
you have, the more games you can play against a variety of opponents with
a variety of playing styles, the better your evaluation becomes. After playing
lots of "fudged" games against Genius 2.0, I am still impressed with the
knowledge that Richard has managed to incorporate into his program. I have
played several games where C-B finds wild tactical things to do, but, with
a little "forcing" by me, the games were guided along a path where the tactics
were somehow eliminated. In certain types of positions, Genius is either very
luck or else very good, as it (and Cray Blitz) both play what I would consider
reasonable moves, yet Genius has won a couple of these games. Again, I'm not
sure I have been fair, as I have tried to control the depth of search for both
programs so that neither are able to see anything tactically important, which
forces them to rely on their evaluation functions (for the most part.) This
is not an exact science, but, clearly, genius 2.0 has been well-tuned while
C-B is only "on its way" to becoming well-tuned.
Opening: A very important part of the game. I have learned that, for a
program, the goal is not to choose the opening line where the program evaluates
itself as having the best score of all lines in the book; rather, we have to
force the program to follow lines where it "accepts" certain weaknesses in
return for "tactical/positional compensation" in the form of significantly
better winning changes. Sounds odd, indeed. An example is playing white
against 1. e4 c5, where white soon plays d4 exchanging a center pawn for a
wing pawn. C-B isn't particularly happy about this line, but we have found
that it gives better winning (and sometimes losing) chances than playing
the lines where the score is highest (which are often lines that include a
forced queen trade and lots of piece trades and end up looking drawish.)
Another problem with this part of the "system" is that we have to be very
careful to avoid playing the same line a second time. Giving your opponents
a year (or two) to "cook your book" is an easy way to lose a game or two
before ever getting your chess program "up and searching away." In short,
this takes a lot of time. With a part-time effort like ours is, time spent
here takes away from time spent on positional tuning or search extension
testing. At the skill level of today's best programs, a weakness in any
area can become a critical weakness.
As to our chances of winning, consider this: With a five round tournament,
assume that we are rated 200 points above every other entry with the exception
of Deep [*] (and maybe HiTech if they play.) Suppose that a rating of +200
has a win/loss average of 4/1 (I'm not sure that this is the exact USCF number,
but it is probably close). This means that for four rounds, each game has
an 80% probability of coming out in our favor. Lets also assume that Deep [*]
is rated 200 points above us. The predicted outcome for this game is .2 (20%
chance of a C-B win.) The probability of ending up 5-0 is .8x.8x.8x.8x.2 or
roughly 2% (2 out of 100 chances.) Let's analyze Deep [*]'s chances of ending
up 5-0. worst case is that Deep [*] is 200 points better than any other program
which gives a probability of going 5-0 of (.8)^5 which is not an impressive
number at all (about 1/3 chances!) Of course, 1/3 is *much* better than
1/50. After factoring in Murphy (as in law) anything can (and probably will)
happen...
Bob
David Gertler
: As to our chances of winning, consider this: With a five round tournament,
: assume that we are rated 200 points above every other entry with the exception
: of Deep [*] (and maybe HiTech if they play.) Suppose that a rating of +200
: has a win/loss average of 4/1 (I'm not sure that this is the exact USCF number,
: but it is probably close). This means that for four rounds, each game has
: an 80% probability of coming out in our favor. Lets also assume that Deep [*]
: is rated 200 points above us. The predicted outcome for this game is .2 (20%
: chance of a C-B win.) The probability of ending up 5-0 is .8x.8x.8x.8x.2 or
: roughly 2% (2 out of 100 chances.)
It's late and I'm sleepy, but isn't .8x.8x.8x.8x.2 .08192, or
about 8%? It should, after all, be 1/4 of (.8)^5, which you
point out is about 1/3.
: Let's analyze Deep [*]'s chances of ending
: up 5-0. worst case is that Deep [*] is 200 points better than any other program
: which gives a probability of going 5-0 of (.8)^5 which is not an impressive
: number at all (about 1/3 chances!) Of course, 1/3 is *much* better than
: 1/50.
It's also better than 1/12, but not by as much...
Dave Gertler, number cruncher
Robert Hyatt
You are right. Seems that my local electronic calculator was set to round
rather significantly. re-doing it on my *real* calculator does get 8%. Of
course, *neither* number looks particularly great... :^)
Mark Mittelstaedt
-> although we are significantly better than in years past. The reason:
-> it takes many hours
-> of testing over a wide variety of positions to properly tune an
-> evaluation term It is even harder to add additional knowledge (it is
-> also hard to simply determine what additional knowledge *might*
-> help.) In short, the more time
Who do you feel is at the "top of the heap"? I don't think that deep
[*] is, because it is just a tactical powerhouse and if and when Garry
plays it, he will try to use every positional trick left in the book to
beat it.
-> roughly 2% (2 out of 100 chances.) Let's analyze Deep [*]'s chances
-> of ending up 5-0. worst case is that Deep [*] is 200 points better
-> than any other progra which gives a probability of going 5-0 of
-> (.8)^5 which is not an impressive number at all (about 1/3 chances!)
-> Of course, 1/3 is *much* better than 1/50. After factoring in Murphy
-> (as in law) anything can (and probably will)
I think there is a flaw in this calculation somewhere... heehe after 6
years of eng school you get these intuitive ideas when looking at
problems... Anyway I think that deep [*] is at least 300pts above some
of those programs promising a win nearly 100% of the time, the rest of
the calculations should be followed from there.
mark