The Dortmund 2000 tournament
was an historic event, as the first-ever tournament with regulation
time controls that included both a computer program and several of
the world‘s elite Grandmasters. As it turned out, we also saw
another historical milestone, as Vladimir Kramnik‘s unbeaten streak
finally ended at an astounding 82 games. Deep Junior was never in
serious contention for first place, as Kramnik and Viswanathan Anand
traded the lead a couple of times and finally tied for first place
at +3, with three players tied at +1, followed by Deep Junior, alone
in sixth place with an even score. However, Deep Junior was
obviously a very important factor in determining the tournament
winner.
In fact, if we were to consider only the games played
between humans, discounting the Deep Junior games for the moment,
the tournament would have had quite a different outcome. Anand would
have won clear first place at +3, with Kramnik and Peter Leko tied
for second place at +2. This result was indeed what I predicted in
my pre-tournament calcuations for the "humans-only" subset of the
tournament, as my statistical model had named Anand as the
pre-tournament favorite, with a 26% chance of clear first place. I
did not include Deep Junior‘s chances in my pre-tournament
calculations, because of the lack of statistical evidence regarding
its strength.
As I mentioned in my Round One statistical
outlook, there were several reasons why Anand was given better
chances than the higher-rated Kramnik. Michael Adams and Alexander
Khalifman had the black pieces against Anand and the white pieces
against Kramnik, and the matchups of opening repertoires indicated
that Anand should have gained at least a full point on Kramnik out
of those two matchups. In fact, Anand actually gained 1.5 points, as
he beat both Adams and Khalifman, and Kramnik lost to Adams and drew
against Khalifman. On the other hand, although Evgeny Bareev and
Robert Huebner had the black pieces against Kramnik and the white
pieces against Anand, my calculations did not give Kramnik much of a
chance to gain back that ground against Anand. The reasoning was
that Bareev is quite strong with Black and correspondingly weaker
with White, and Huebner had such a low rating that he was probably
going to lose to both Anand and Kramnik. That was also correct, as
Huebner lost to both Anand and Kramnik, while Bareev drew against
each player. So, out of those four matchups, Anand was 1.5 points
ahead of Kramnik. They also each won one game with White against the
lowest-rated players (other than Huebner), with Anand beating Jeroen
Piket and Kramnik defeating Vladimir Akopian.
So, where did
Kramnik make up those 1.5 points? Well, he defeated Anand in Round
Six to catch up one full point. The remaining half-point was gained
by Kramnik in their games against Deep Junior, where Kramnik
defeated the computer while Anand could only draw against it. This
allowed Kramnik to finish even with Anand.
There were very
few upsets between human players at the Dortmund 2000 tournament.
The victories by Adams over Kramnik, Kramnik over Anand, and Anand
over Adams, were all won by the "favorites" (as determined by my
statistical model); the only two "upsets" (not counting Deep
Junior‘s games) were Evgeny Bareev‘s two losses, in games against
Akopian and Leko where he was actually a slight favorite. However,
the games involving Deep Junior were much harder to predict; its
loss to Piket with the white pieces and its victory over Leko with
the black pieces were two huge upsets. This was the big reason why
my estimate of Deep Junior‘s strength kept varying so much; it kept
doing the unexpected! By the way, my final estimate of Deep Junior‘s
strength, based on its performance at Dortmund, is 2731. This was a
little lower than my original estimate of 2740, which would have
predicted a finish of +1, but considerably higher than the estimate
of 2691 that it had going into the final round with a minus score!
Note that this estimate, calculated using a statistical technique
known as the Method of Maximum Likelihood, is more accurate than the
official tournament "performance rating"of 2702 which Deep Junior
received, because my method takes color into account. Deep Junior
had to play five out of nine games with the black pieces, making its
even score somewhat more impressive than if it had played five games
out of nine with the white pieces. My method also takes the
color-related abilities of opponents into consideration; it is much
more impressive to score 1.5/2 with White against Morozevich and
with Black against Khalifman, than it is to score 1.5/2 with Black
against Morozevich and White against Khalifman, because Khalifman is
so much better with White than with Black, whereas Morozevich scores
equally with White and with Black.
In any event, I said at
the start that my predictions of Deep Junior‘s chances in the
tournament were somewhat uncertain. One very important feature of my
statistical model is its ability to compare opening repertoires of
the two players in a game, and that was not possible to do for Deep
Junior. Allow me to digress briefly, to give an example. Garry
Kasparov, as is well known, has an incredible mastery of the
Sicilian Defense, with White or Black. If he is facing an opponent
against whom the Sicilian is very likely to occur, Kasparov is going
to have an even larger advantage than usual. On the other hand, if
the Sicilian is not very likely to occur, then one piece of his
mastery is going to be missing, and his chances against that
opponent might not be as good as we might otherwise think.
Judit Polgar, for instance, almost invariably plays 1.e4,
and also plays the Sicilian with Black almost 100% of the time.
Against Kasparov, a Sicilian Defense is very likely, and so Polgar
would be predicted to have even more trouble than their rating
difference would suggest. That prediction is certainly confirmed by
their lifetime results against each other. Conversely, Jeroen Piket
opens with 1.e4 less than 1% of the time, and he hardly ever plays
the Sicilian with Black against 1.e4. When Piket and Kasparov meet,
a Sicilian is very unlikely, and so Kasparov might not do as well as
their rating difference would suggest. Their results earlier this
year provide some confirmation of that prediction, as
well.
With Deep Junior, however, there are hardly any
regulation games (that we know about) against highly-rated
opponents. We don‘t really know yet whether it is particularly
dominant with White, or does incredibly well in open positions, or
has trouble with certain openings or pawn positions. In fact, we
don‘t even know its true rating yet. The top players are still
learning about the best approach against Deep Junior, and only time
will tell whether technology has caught up to humanity yet. However,
the only way for us to learn more about the strengths and weaknesses
of computer programs, is for more events of this type to be held in
the future. Over time, we will learn more, and it will be possible
to make even more accurate predictions.
As Vladimir Kramnik
recently pointed out, there is a serious cost that goes along with
the introduction of computers into tournament chess; every
tournament spot granted to a computer is one fewer opportunity for a
human grandmaster to participate. Of course, if the publicity leads
to more money and therefore more tournaments spots, those same human
grandmasters would definitely benefit. It is up to the chess world
to decide whether this experiment is worth the cost. However, the
Dortmund 2000 tournament was certainly a big success, and we can all
be grateful to the tournament organizers and participants for a
piece of chess history that was both entertaining and
educational. |
Free Email
Chat
Message Boards
KC Newsletter
