Deep Fritz 8 vs Deep Junior 9 – the truth within
By Aryan Argandewal
On a hot Monday, June 27th 2005 as Michael Adams was trying his best against
the electronic nine-headed monster in round six of the Man
versus Machine match. I was sitting a few yards away, feeling a little
depressed like everyone else in London Wembley Conference Centre. After all,
the future of humanity was on the line. Sitting next to me was a cheerful Dr.
Donninger, the human brain behind Hydra. Earlier I asked him how deep was their
opening preparation and whether human Grandmasters did most of the work. "No,"
replied the Austrian laconically, "humans would only spoil Hydra’s
evaluation." Yeah, I agreed in a low voice. And the key move, 14.Rb1,
is an original idea from Hydra itself, insisted Dr. Donninger with a distinct
Austrian accent. It deliberately complicates the position so as to make it
impossible for humans to correctly evaluate the position.
Michael Adams vs the giant Hydra system
Instead of expressing my views on that I simply smiled sadly. Aside from being
an Adams’ fan I hated the idea that he was losing game after game completely
free of charge! Let’s face it: Kasparov was paid $400,000 in 1997 after
losing to Deep Blue. Mickey received $10,000 in 2005 for doing exactly the
same! The odds of winning the promised $150,000 were that of hitting the national
lottery jack pot!
Although the result was a foregone conclusion by round six, the very thought
that the human player would score at least one victory in six rounds was something
to look forward to. In game six Mickey had a very nice position, and Hydra
didn’t seem in a hurry to open it up. For a moment everybody felt that
a draw was the very least White could hope for. Even Jonathan Speelman, who
was pretending to be reading a book, suddenly began to look up and down the
display board. But it didn’t happen. Michael played 31.c5 – opening
up the position to the tactical monster – and it was all downhill from
there.
As I was looking at the rather depressing picture on the display board, thinking
about Mickey Adams, and then about what was sitting across the table –
an object capable of calculating up to 40 moves ahead, a machine that had no
concept of fear – I suddenly imagined how history could remember the
day. The PAL Group Funding Bill is passed. The system goes on-line June 26th
2005. Human decisions are removed from strategic defence. Computer codenamed
‘The One’ begins to learn at a geometric rate. It becomes
self-aware at 2:14 a.m. GMT, June 29th 2005 London, England…
What are they? Sentient programs. They can move in and out of any software,
still hardwired to their systems. Designed exclusively for one mission: search
and destroy…
Sentient programs
At a time when machines are likely to take over not just how we play the game,
but also how we learn the game, it is of paramount importance to distinguish
between good and best chess programs. Aside from the sporting parameters it
is also important to see how different programs have adapted to human-style
chess.
Analysts believe that of the commercially available programs at the moment
the top three are Fritz, Junior and Shredder, with the latter having a slight
edge in Elo terms. I decided to see it for myself in a three-stage experiment:
the electronic gladiators would play 30 blitz games (4 min + 2 sec), five rapid
games (25min) and three games with classical time control.
The most important factor for me was to see how different it is to play against
each program. To judge it from a human perspective: aggression, positional
pressure, tactical combinations, choice of openings, etc. What really matters
is how close a chess program can get to human-style chess. The latter was of
the highest priority.
In my chess library I have a variety of chess programs but I find myself playing
tournament games only against one of these programs. Whenever I want to play
a long game I never have trouble choosing one. I just reach for one of these
programs, almost on auto-pilot. Which one is it? Let us leave this until we
finish the experiment, then I will reveal the secret...
Operation 64
In the first part of the experiment we shall witness the battle between World’s
Favourite Deep Fritz 8 and the four-time World Computer Champion Deep Junior
9. I decided to put to the test the multi-processor versions of both programs,
as that ensures maximum performance. Both programs ran on a Pentium 4, 3.6
GHz VIO notebook (512 Mb RAM, 40 GB HDD) available at your local PC store.
The winner would have to face the current favourite Deep Shredder 9.
For now, let us crack on with stage one of the experiment.
Fritz vs Junior
From the very start it looked that Fritz had a clear edge. After 30 blitz
games Deep Fritz 8 came out a clear winner with an undisputed score of: 19-11.
Okay, time for rapid 25s. After five games Deep Fritz 8 once again proves
to be the invincible German Mercedes, beating Deep Junior 9 overwhelmingly
with a score of 5-2. Let’s move on. It is time for real chess with classical
time-control.
Game one
In this game Deep Junior 9 finds itself in trouble in a Richter Rauzer Sicilian.
Despite having played 20 moves from its openings book the program is viciously
stormed on the kingside, and just ten moves after the last book move Junior
faces a lost position. On move 30 Deep Junior 9 spent nearly 17 minutes before
moving the king to h7, showing first signs of desperation. Finally on move
34, a few moves away from a forced mate Junior resigned.
Deep Fritz 8 (3300) - Deep Junior 9 (3300) [B66]
Game 1 (1), 03.06.2005
1.e4 c5 2.Nc3 Nc6 3.Nf3 d6 4.d4 cxd4 5.Nxd4 Nf6 6.Bg5 e6 7.Qd2 a6 8.0-0-0
h6 9.Be3 Be7 10.f4 Bd7 11.h3 b5 12.Bd3 Nxd4 13.Bxd4 b4 14.Ne2 e5 15.Be3 Qa5
16.Kb1 0-0 17.g4 exf4 18.Bxf4 Be6 19.b3 d5 20.e5 Ne4 21.Bxe4 dxe4 22.Nd4.
White's last book move 22...Rad8 23.Qe2 Bd5.
24.g5 Bxg5. 24...hxg5 25.Nf5 Qc7= (25...gxf4?? will give
the opponent a chance to mate in two 26.Nxe7+ Kh7 27.Qh5#). 25.Bxg5
hxg5 26.Rhg1 g6. 26...f6 27.exf6 gxf6 28.h4 is better for White. 27.Rxg5
Qb6. 27...Rfe8 was possibly the rescuing straw. 28.Nf5 Be6. 28.Nf5
Be6 29.Nd6 e3 30.h4 Kh7 [30...Rd7 31.h5 Qc5+–] 31.h5
Kg7 32.hxg6 f5 33.Rh5 Rh8 34.Rdh1 1-0.
Game two
In this game Deep Fritz 8 resorts to the super solid Petroff, which ends in
a dead 25 move grandmaster draw. I proved myself to be wrong that computers
play the position until it is completely exhausted. It turns out that when
the situation necessitates computers can be just as pragmatic and boring as
human grandmasters.
Deep Junior 9 (3300) - Deep Fritz 8 (3300) [C42]
Game 2 (1), 04.06.2005
1.e4 e5 2.Nf3 Nf6 3.Nxe5 d6 4.Nf3 Nxe4 5.d4 d5 6.Bd3 Nc6 7.0-0 Be7
8.c4 Nb4 9.Be2 0-0 10.a3 Nc6 11.cxd5 Qxd5 12.Nc3 Nxc3 13.bxc3 b6 14.Re1 Bb7
15.Bd3 Rae8 16.c4 Qh5 17.Be3 Na5 18.Be2 c5 19.d5 Bd6 20.h3 Bc8 21.Rc1 Re7 22.Nd4
Qe5 23.Nf3 Qh5 24.Nd4 Qe5 25.Nf3 Qh5 ½-½.
Game three
This game was probably the most conclusive proof of Deep Fritz 8's superiority.
After the opening, which was a solid Slav Defence, the game appeared on equal
terms – until Fritz somehow miraculously found something to build on.
Gradually this little plus turned into a full (+–) and once that happened
it was impossible to stop Fritz from demolishing opponent’s defensive
resources. Just like a human grandmaster who takes advantage of his weaker
opponent Deep Fritz 8 won game three in a 75-move marathon. See it for yourself.
Deep Fritz 8 (3300) - Deep Junior 9 (3300) [D14]
Game 3 (1), 04.06.2005
1.d4 d5 2.c4 c6 3.Nf3 Nf6 4.cxd5 cxd5 5.Nc3 Nc6 6.Bf4 Bf5 7.e3 e6 8.Bd3
Bxd3 9.Qxd3 Bd6 10.Bxd6 Qxd6 11.0-0 0-0 12.Rfc1 Rfc8 13.h3 h6 14.a3 Na5 15.Rc2
a6 16.Rac1 Rc7 17.Nd2 Rc6 18.Na4 Rxc2 19.Rxc2 b5 20.Nc5 Nc4 21.Nb1 Nb6 22.Qd1
Qd8 23.Qc1 a5 24.Qf1 Nc4 25.b3 Nd6 26.Nc3 Qe8 27.Qd3 b4 28.axb4 axb4 29.Na2
Ra3 30.Nxb4 Ra1+ 31.Kh2 Nde4 32.g3 h5 33.Nxe4 Nxe4 34.Nc6 f6 35.f3 Nd6 36.e4
dxe4 37.fxe4 Kh7 38.Qc3 Rf1 39.Re2 Qd7 40.Kg2 Rd1 41.d5 Nf7 42.Qc2 Ra1 43.b4
e5 44.Qc3 Ra4 45.Qf3 Kh6 46.b5 Nd6 47.b6 Qc8 48.Rb2 Ra1 49.g4 Nb7 50.Kh2 Qa8
51.Rg2 Ra2 52.gxh5 Qa3 53.Qf5 Rxg2+ 54.Kxg2 Qb2+ 55.Kg3 Qb3+ 56.Kh4 g5+ 57.hxg6
Qd1 58.Qf2 Nd6 59.Qe3+ Kxg6 60.Ne7+ Kf7 61.Nc8 Nc4 62.Qg3 Nxb6 63.Nd6+ Ke7
64.Qa3 Kd8 65.Qc5 Qd4 66.Nb7+ Ke8 67.Qc6+ Kf8 68.Qxf6+ Kg8 69.Nd6 Qg1 70.Qd8+
Kh7 71.Nf5 1-0.
Aggregate score: 2.5:0.5.
Mental projection of one’s digital-self
The recent Freestyle
Chess Tournament on the Playchess server
gathered no fewer spectators than heavy-weight Grandmaster bouts. What turned
out to be even more shocking was that the $10,000 prize was scooped off not
by participating heavily-armed Grandmasters or International Masters, but by
some kids with Elo ratings of about 1700 and 1400! And if that wasn’t
enough food for contemplation, they also managed to demolish the fearsome Hydra
along the way! The latter is going to have profound implications for chess
software industry.
Amateurs who defeated heavily-armed GMs in style were using not multi-million
dollar chess hardware but the very Fritz, Junior and Shredder we are all lucky
to have. It follows, therefore, that even for an amateur chess program’s
main engine analysis, i.e. evaluation of candidate lines is the vital
component.
Junior the Adventurous
Looking at how Deep Fritz 8 and Deep Junior 9 analyse a given position you
cannot help wondering how different the two really are. Junior’s evaluation
of a position swings wildly from plus zone to minus. In other words, given
a position to evaluate, it will come up with what it considers to be the best
line, and then a few moves later it will suddenly change its own evaluation
of the line, i.e. what was assumed a definite (+–) can easily change
to (–+).
This is a double-edged sword. Sometimes it wins a game in fine style, sometimes
it looses it outright. This former parameter "in style" will be the
most precious for the lovers of dynamic, flamboyant chess. Junior’s search
for a good move reminds us of the adventures of Emil Sutovsky, who last week
beat Vladimir Kramnik by sacrificing a piece. On a more cautious note one needs
to look at Sutovsky’s overall performance at Dortmund to realise how
important it is to treat this flamboyant strain with care. Junior's strength
is that no moves are excluded. The program is willing to take serious risks
and will produce lines that no other program will even consider to go into.
Agent Fritz
In sharp contrast to Junior, Deep Fritz 8 has a more mature approach. It slowly
builds up a profile, and once it has reached what it considers to be a winning
position, it almost never gives up. It gradually builds on that advantage until
the opponent is mercilessly crushed. If you see a slight advantage in Fritz’s
evaluation you’d be rest-assured that it will hang on to it. More importantly
it will do everything "computably possible" to increase the advantage.
Looking at Agent Fritz you are reminded of the human boa constrictor Anatoly
Karpov. The similarity stems from the fact that neither will give the opponent
the slightest counter play. And they will gradually increase the pressure,
while keeping tactics under control, unless the position objectively calls
for some fireworks.
Whilst typing these words I am watching a fascinating battle on the Playchess
server between the Boa constrictor himself and a Grandmaster named N. David.
It is a slow Caro-Kan but Karpov’s game seems a lot more interesting
than any of the Dortmund games transmitted live on the server. Despite some
terrible breakdowns in transmission (the game has restarted four times so far!)
it is obvious that Karpov is doing very well indeed: offering a pawn-sac here
and there, manoeuvring a piece from one diagonal to another. Despite being
offered several times, refusing to exchange pieces (unlike some modern Super
Grandmasters) Karpov is slowly improving his position, or at least keeping
his opponent's play to an absolute minimum. Fritz plays more or less a similar
game.
When you play against Deep Fritz you feel you are constantly under positional
pressure. The program is relentless and the pressure is there throughout the
game. Fritz does not forgive the slightest misjudgment whatsoever. With Junior
you feel less pressurised. You also feel at ease to roam around the board at
your will and come up with your own ideas. Not against Fritz. It forces you
to follow its own ideas from the very start. Junior can hit you with a surprise
tactical punch every now and then, but you are not sitting on the edge of your
chair as you do against Deep Fritz.
The name is Fritz
When I played Deep Junior 9 for the first time I thought it was a great program.
I still think it is a joy to play against it for fun. But to compare it to
the world’s favourite Dr. Fritz I am afraid it’s been slightly
disappointing. It is inferior in many categories. Deep Fritz 8 still feels
like the best program for deep positional analysis, as well as for human-computer
practice. In my humble opinion, whether you want a program for fun, analysis
or tournament preparation, Deep Fritz 8 is the clear winner of the two.
In the next part we will witness the battle between the current favourite
Deep Fritz8 and the strongest Elo monster – Deep Shredder9.
Aryan Argandewal studies Law at University of
Surrey, England. He speaks English, Russian and Persian, and is able to read
and write Japanese, Arabic and Pashto. He is a member of Guildford Chess Club,
Surrey, UK.
