Pal/CSS
Freestyle Tournament
March
18 & 19, 2006
report
by Vasik Rajlich
On
the weekend of March 18 & 19, the PAL/CSS Freestyle Tournament was held on
the Playchess server. Under the freestyle rules, each participant can use any
means he likes to choose his chess moves – this includes analyzing with a chess
engine, accessing any chess theory, and even consulting other chess experts.
Computer chess experts as well as human chess masters have traditionally agreed
on one thing: a human chess master, assisted by a top chess engine, is capable
of producing the highest level of chess known to mankind.
With
a $16,000 prize fund attracting an eclectic combination of 148 participants,
including more than 15 over-the-board grandmasters as well as most of the
leading computer chess experts, we were bound to put this theory to the test,
and produce some exciting chess along the way.
Early
Rounds
The
early rounds of a typical swiss tournament tend to consist of some combination
of mismatched wipeouts and interesting upsets. Not so here – with (nearly)
everyone assisted by their favorite chess engine, the level of even the bottom
seeds was very high, and interesting chess could be seen right off the
bat.
Swedish
computer chess expert Bjorn Osterman (ie. King Crusher) got off to a slow start,
drawing his first three games. In his second-round game, the British grandmaster
Tony Kosten gave a good demonstration of the power that a human master brings to
the table as part of a centaur combination:
Rd
2: King Crusher – Kosten, Black to Move
Here,
Kosten played 28. .. Rf4!, understanding that he can draw without any trouble
the endgame after 29. fxg4 Nxg5 30. gxh5. King Crusher tried in vain to win
until move 102. This type of drawing pawn sacrifice remains elusive for modern
engines, although Rybka considers 28. .. Rf4 to be only very slightly worse than
28. .. Bf5. The example highlights one of the strengths which humans still have
over chess engines: humans better understand when a nominally worse endgame is
drawn enough to be practically equal.
In
some cases, though, the human input can backfire. This point is illustrated here
by last year’s champion, the American team ZackS, playing against another
unaided Rybka 1.1:
Rd
2: ZackS – Octapus, White to Move
White,
as any red-blooded human, smells the chance to launch an attack which could well
prove to be too deep for an unaided chess engine. The game continued 19. g5?!
hxg5 20. f4 gxf4 21. Qxf4 bxc3 22. e6 Ra7 23. Bxd5 and now black coolly repelled
the attack with 23. .. Qxd5 24. e7 Rxe7 and went on to
win.
Moving
on to the fourth round, a more successful human idea comes courtesy of the
up-and-coming American grandmaster Hikaru Nakamura (Star Wars), playing against
another unattended Rybka 1.1:
Rd
4: Star Wars – Hoshad, White to Move
16.
Bg5! hxg5 17. hxg5 Ng4 (Rybka already admits that white is slightly better) 18.
f4! Qd7 (18. .. Nf2 is busted by 19. Qf3) 19. Qh3 Nh6 and white has turned the
position in his favor, although the game was eventually drawn in 136
moves.
The
following fourth-round game gave us the first clear leader of the tournament,
Vigi Varkey from Britain, another Rybka 1.1 user who moved to 4/4 after a wild
game against Boguslaw Latas from Poland (Dykta).
Rd
4: VVarkey – Dykta, Black to Move
Rybka
thinks that black is in trouble and should grovel for a draw with 23. .. Qxc7
24. Nxe5 Qb6+ 25. Bd4 dxe5, but black trusts in his position and throws
everything at the white king: 23. .. g3! 24. Nxa6 Ne8 25. Rc2 Qh4 26. h3 Bxh3!
27. gxh3 Qxh3 28. Bd3 Nh4 and finally Rybka’s scores start to plummet. A roughly
equal endgame was reached in which Varkey’s passed b-pawn overcame Dykta’s
central pawn mass.
Round
5
In
Round 5, Robert Mraovich (Pawnstriker1978) from the US moved to the top board,
but Vigi kept rolling right along:
Rd
5: Pawnstriker1978 – VVarkey, Black to Move
Black’s
attack already looks faster than white’s. The game ended quickly after 21. .. b3
22. Bc4?! bxc2+ 23. Qxc2 Rab8 24. hxg6 Rxc4 and black went on to
win.
On the second board, the UAE chess playing
hardware Hydra (Zor_champ) moved to 4.5/5 with a fine victory over Sean Eaton
(revengeska) of the US:
Rd
5: Zor_champ vs revengeska, White to Move
White
is down a pawn but has a slight initiative, which he pursues with impressive
persistence: 13. a4 Rb8 14. Ra3 Bf6 15. axb5 axb5 16. Bxc6 Bxc6 17. Rd3 Qc8 18.
Qc2 Bd7 19. Re1 Be6 20. Qc6+ Kf8 and now white’s compensation is obvious and he
went on to win.
Round
6
Round
6 brought a clash of the leaders, with VVarkey at 5/5 taking on Zor_champ at
4.5/5. A sharp opening led to the following position:
Rd
6: VVarkey vs Zor_champ, Black to Move
Black
is slightly cramped and should probably defend with something like 17. .. Qc7 or
17. .. g6. Instead, Zor_champ lashed out with 17. .. f5 18. exf6 Nxf6 19. Nxf6+
Rxf6 20. Rd1 Bd5 21. Be3 Rf7 22. Rd2 Bd6 23. g3 a5 24. Ba6 Rc6 25. Bb5 Rc7 26.
Bb6
and it’s clear that the gamble did not pay off. Varkey went on to convert his
advantage.
Round
7
The
first place in the tournament was settled in this round. The Czech team
Equidistance took their 5/6 score to the first board and suffered the same
result as the others. Not much is known about this Czech team, except that they
seem to play some very interesting opening variations. Of course, being Czech,
they can’t be that bad. J
Rd
7: Equidistance vs VVarkey, White to Move
White
has already sacrificed one pawn somewhat speculatively, and now throws a piece
into the fire with 11. Nxe6 Qc8 12. Nd5 exf4. There were moments when it looked
like white might be doing fine, but in the end Varkey won again in a long
endgame, clinching first place with a round to spare.
With
the battle for first place decided, the remaining players turned their attention
to placing in the first 8 and therefore qualifying for the final on April 8
& 9. It looked like +4 (ie. 6/8) would be needed for
this.
Zor_champ
returned to +4 (ie. 5.5/7) with a patient manoevering victory over Michael
Babigian of the USA (Farseer).
Rd
7: Farseer vs Zor_champ, Black to Move
Black’s
advantage looks minimal, but watch how he manoevers around white’s only weakness
(the b2 pawn): 38. .. Bg7 39. Nfe3 Bd4 40. Na5 Qa6 41. Nac4 Kh7 42. Rc1 Qb5 43.
Nd6 Qb8 44. Ndc4 Nf4 and here white finally cracked with 45. Qe4 (apparently an
operator error, but the position was already very difficult) Bxb2 and black went
on to win.
Bjorn
Osterman (King Crusher) had by this point recovered from his slow start and also
reached the same score by refuting a rather unusual gambit by Dar Paris from
Brazil (Bychamp_II):
Rd
7: Bychamp_II vs King Crusher, White to Move
White
invariably plays 8. Nf5 here, but here Bychamp_II chose instead 8. Nb3 Nxg4 9. Bd2 Nc6 10.
Rg1 Qh4 11. Rg3 f5. I won’t claim to understand this completely, but black looks
a bit better and did go on to win the game.
The
final participant to reach +4 was yours truly, playing together with Iweta
Radziewicz and 4 Rybkas, who beat fellow international master Konstantin Maslak
of Russia (MASLAKKOSTIA) in an interesting game:
Rd
7: Rajlich vs MASLAKKOSTIA, White to Move
This
is a rather common opening position, but it looks like white has a brutal attack
at his disposal: 12. e6! Ndb8 13. Nd5 Qd8 14. Qb3 b6 (14. .. Na6 is also bad)
15. f5! gxf5 16. Nf4 Qd6 17. 0-0-0 Ne5 18. Bc3 Qc7 19. Kb1 and black is in
trouble (and could not hold the position).
Round
8
The
matchup on first board was anti-climactic. Vvarkey had already clinched clear
first place, while King Crusher needed just a draw to qualify and thus was not
disappointed that Vvarkey could easily hold a slightly worse
endgame:
Rd
8: King Crusher – Vvarkey, White to Move
Since
it was two automatic Rybkas doing battle, they were not aware of the tournament
situation, but if they were humans we would say that both were happy with a
draw: 15. Qxa5 Nxa5 16. Kd2 Bd7 17. h5 Rac8 18. Bd3 and white played until move
36 trying to make something out of the position.
Board
2 was an interesting matchup of centaurs: Hydra team vs Rybka team. However, the
tournament situation made a draw acceptable for both sides and the game
developed peacefully. In fact, the most exciting moment in the game came on a
technicality:
Rd
8: Zor_champ vs Rajlich, Black to Move
The
position is well on its way to being drawn, but I switched to the wrong window
and rather than playing .. Rc4 in an analysis board, I played it here in the
actual game! Those with Playchess experience know that there are some who
forgive mouse slips and some (many) who do not, and here the Hydra team was
given a golden chance to drop us into the tie-break. To their credit, they did
not – the game continued 21. .. Rc4 22. h3 Rc7 and was agreed drawn on move 59.
The Rybka-Hydra matchup will resume on April 8 & 9.
On
the remaining top boards, the participants had to win, as a draw would only be
good enough to land a place in a big tie-break for the last
spots.
The
first to succeed was the German international master Dennis Breder
(Klosterfrau), who managed to slowly outplay Tebat Cupulanu (Pulanu) from
Romania in a Berlin Spanish:
Rd
8: Pulanu vs Klosterfrau, Black to Play
White
has floundered somewhat in the opening – oddly enough, in this line it is
usually a mistake for white to allow black to trade his dark-squared bishop for
one of the knights – and now black puts his finger on white’s problems with 23.
.. g5! 24. Rd1+ Kc6 25. Ng2 Ng6 26. f4 gxf4+ 27. Bxf4 h5! and white could not
hold the position.
The
second to qualify for the final was another predictable face: correspondence
grandmaster and well-known computer chess expert Arno Nickel from Germany
(Ciron), who overcame the resistance of Sasha Belezky of Ukraine
(Schurick).
Rd
8: Ciron vs Schurick, White to Play
Black
had decided early in the game to house his king in the center, and so far white
has patiently operated against this. Now, more drastic operations begin: 28.
Qg7! Bf8 29. Qg8 Rxb4 30. Ng5 Qe7 31. Re1 and black could find nothing better
than fleeing with 31. .. Kd7 32. Rxd5+ Kc6 with a lost
position.
The
last participant to succeed was the mysterious Czech team Equidistance, which
overcame the resistance of EveRest from Turkey after yet another speculative
opening:
Rd
8: Equidistance vs EveRest, White to Play
Yes
– as vs Zor_champ, the Equidistance team has already sacrificed a pawn
immediately in the opening, and now they again throw a piece into the mix as
well. It’s not clear if it is correct, but we can see that sometimes fortune
favors the brave even when chess engines are involved: 12. Nxg5!? Hxg5 13. Bxg5
Re8 (it’s not clear how white continues in the case of 13. .. Ne5) 14. f4 Rxe4
15. Bd5 Re3 16. Qd2 Re8 17. Rf3 Bf5 18. Rg3 Kh7 19. Rxb7 and black could not
find anything better than 19. .. Rb8.
Three
other games could send the winner into the final, but all three were drawn. The
six participants of these games will participate in a tiebreak on March 24 &
25.
Sean
Eaton from the US (Revengeska) could not convert a slightly better position
against D.K. from Germany (Teutates), sending both participants into the
tiebreak. The most interesting part of the game may be the position where they
agreed to a draw:
Rd
8: Revengeska vs Teutates, White to Play
Agreed
drawn, 1/2 – 1/2. A bit early for a human observer like me, but those analyzing
with an engine will quickly see 32. Rf6 Nxf6 33. Qxf6+ Kg8 34. Nf5 gxf5 35. Qg5+
with a perpetual.
One
board down, Terrai Brauford from Korea (Ultima) could not squeeze Marek Baron
from Germany (Tatar) in an endgame, also sending both participants into the
tiebreak.
Rd
8: Ultima-Tatar, Black to Play
Black
has been suffering for a long time and now enters full grovel mode with 32. ..
Rxf7+ 33. exf7+ Kxf7, trusting that white’s two pawns would not be enough to win
– which they weren’t.
In
the last game between participants with a chance to qualify directly for the
final, Don Kassandder from South Africa (Teutates) could not overcome
Octapus.
Rd
8: Donkasend vs Octapus, Black to Move
White
has a chance for a nagging advantage, but black is able to simplify the position
with 26. .. Nf4 (already giving 0.00 scores) 27. Bxf4 exf4 28. Qxf4 Bb2 and the
position is dead drawn.
In
the end, 10 participants made it to +3 (ie. 5.5/8), yielding a 10-way playoff
for one spot in the final. Those who drew the games above were joined by the
following group of players who won their last-round games to also end the
tournament with +3:
Darren
DiAlfonso, USA (Relic)
Joe
Soney, USA (Jazzled)
Abel
Davalos, Mexico (Abeljusto)
Akhtar
Hashmi, UAE (Akhtar)
Post-Mortem
Before
the tournament started, I would have guessed that it would be dominated by the
competent centaurs. An intelligent chess expert should be able to guide his
engine down more productive search paths and correct any holes in its strategic
understanding; and an unaided Rybka should struggle to score 50% in such a tough
field.
As
it turned out, unaided Rybkas did far better than that - one unaided Rybka took
clear first place, another tied for 2nd through 7th, and a
whopping 6(!) more tied for 8th through
17th.
What
happened here? Did I just overestimate myself and what my human mind could
contribute? What does this tell us about the modern interaction between the
human chess expert and his engine?
First
of all, if you’re willing to look past the first place finish of Vvarkey as just
plain lucky (which I am J),
then you will find some pretty decent evidence in the results for the benefits
of an intelligent centaur. In the group which finished 2nd through
7th, 5 out of the 6 spots were taken by participants who fit this
bill – expert computer chess users playing as centaurs who know how to get the
most out of their chess engine.
Still,
the extent to which an intelligent centaur is better than a pure engine may be a
bit less than I had previously thought. I have run into this effect in a
slightly different form many times as a computer chess programmer: I cannot even
begin to count how many times I have “improved” Rybka’s evaluation, only to
watch Rybka’s results drop. A number of chess themes – things that many humans
would even consider to be positional principles – have turned out to simply not
work. How much of what we think is good chess play is in fact good chess play,
and how much is just myth?
Anyway,
all of this is still in the early stages. We need more data, we need more
tournaments such as this one, we need engines which have gone further yet than
Rybka 1.1 along the path from mere “search tools” to truly knowledgeable
analysis tools, and we need to give the human analysts more time to get familiar
with them. This tournament has certainly been an interesting initial data
point.