Deep Blue Prototype Copenhagen matches (games and comments)
Feng-Hsiung Hsu
exhibition matches, possibly in conjunction with IBM Sweden. At some
point, there was talk about playing Swedish GMs as well, but the final
arrangement was to play only in Copenhagen and only against Danish players.
Two separate offical matches were scheduled, a 4-game match against GM
Bent Larsen and a 4-game match against the Danish national team, with the
last game against Larsen also counted as part of the team match. So,
there are actually only 7 official games. 5 additional exhibition one hour
sudden death games were also scheduled during the weekdays. The Copenhagen
Chess Union co-sponsored the matches along with IBM Denmark.
The games officially started on Feb. 24 (Wed). We arrived early on Friday
Feb. 19 in order to participate in a publicity event the next day with
the human World Chess Champion Garri Kasparov. We had dinner with
Kasparov and the organizers on Friday eve. During the dinner, we first
learned about how well prepared the Danes were. Jens Nielsen, who created
the Nielsen test set for chess computers, had made a book for the Danish
players containing the Deep Thought games, various relevant articles,
pointers about machines' weaknesses, and so on. We were concerned, but
the machine could not be set up for match preparation over the weekend
because of the prescheduled weekend events.
We started setting up the machine on the match site Monday Feb. 22. The
machine room wasn't adequately cooled (no air conditioning) and one of the
processors overheated and died after a few hours. The ventilation was
improved Tuesday and a test game was played without further mishap.
The machine was referred to as Nordic Deep Blue by the organizers. It is
actually a Deep Blue simulation running on Deep Thought II hardware. The
host computer was an IBM RS/6000 550 workstation, and 14 custom chess
processors were housed in a separate box communicating with the host via
a VXI cable. It searches 4-5 million nodes/sec at peak speed. The total
cost of all the chess specific processors is about $24,000.
Over the official 7 games of the matches, the machine scored 4 to 3. The
average rating of the opponents is around 2510-2530. The performance is
therefore around 2560-2590 FIDE. Given the well-done preparation by the
opponents and the winning chances it had in some of the drawn games, I have
to say that the overall result is very encouraging. With the new hardware
coming up this summer, things should get interesting.
--Hsu
-----------------------
Day 1 (Feb. 24, Wed.)
This was a disastrous day for us. It was also the first time we fully
realize how serious our own lack of preparation was and how well prepared
the opposition was. Frantic work followed after the games.
Larsen adopted a very simple strategy that worked surprisingly well in the
first match game. He traded off all the machine's knights, allowing the
machine to have the bishop pair but without an open position to realize
the potential of the bishop pair. This really should not have worked
if the machine had been told explicitly to trade off some pawns to increase
the scope of the bishop pair. This diagnosis, however, came a little bit
late, as you will see later. Black's opening preparation was also too
superficial.
Larsen match, game 1
GM Bent Larsen (FIDE 2565?) vs. Nordic Deep Blue
Copenhagen, Denmark, 02/24/93 [40/2,20/1,SD/30min]
1. e4 e5 2. Nf3 Nc6 3. Nc3 Nf6 4. Bb5 Bb4 5. O-O O-O 6. Bc6 dc6 7. d3 Qe7
8. Ne2 Bg4 9. Ng3 Nh5 10. h3 Ng3 11. fg3 Bc5 12. Kh2 Bc8 13. g4 Be6
14. Qe2 f6 15. Be3 Be3 16. Qe3 h6 17. a4 Qb4 18. b3 b6 19. Rf2 c5
20. Kg3 Qa5 21. h4 Qc3 22. Raf1 Rad8 23. g5 Bb3 24. cb3 Rd3 25. Qe2 hg5
26. hg5 fg5 27. Rd1 Re3 28. Qb2 Qb3 29. Qb3 Rb3 30. Rd5 Ra3 31. Re5 g4
32. Kg4 c4 33. Rd2 Ra4 34. Rd7 Rc8 35. Ng5 Ra2 36. Rc7 Ra8 37. g3 Rf2
38. Ree7 Kh8 39. Rg7 Rh2 40. e5 Rd8 41. Rh7 Rh7 42. Nh7 Rg8 43. Ng5 1-0
After the first Larsen game, there were two exhibition SD/1 games against
IM Bjarke Kristensen and IM Jens Kristiansen, both again ended in disasters.
The machine wrongly went after BR on a8, neglected its own king safety
and regretted it two plies later in the game against Bjarke. Jens played
an excellent anti-computer game, although our own opening preparation again
has something to do with the bad play by the computer.
-----------------------------
Day 2 (Feb. 25, Thu.)
The night before, a few modifications were made to the program. One of
them concerned removing the equivalent of surgical-knife-left-in-the-stomach
from the machine's search extensions code.
Three more SD/1 exhibition games were played during the day against Bjarke,
Jens and IM Svend Hamann respectively. It drew Bjarke, beat Jens and lost
to Svend from a completely winning position. The loss against Svend had
a position that is quite interesting. The machine needed 43 sec to see the
move that it played loses, but had only 37 sec allocated. We tried the
same position on various commercial programs, and none of them saw the problem
in reasonable time.
The Danish team was probably using the exhibition games to look for
additional weaknesses of the program. We were using them to catch up on
our preparations.
The second serious game and the first of the team match was played in the
evening. Danielsen was surprised by several sharp moves by the machine,
and after 15. Rab1?!, was greeted with the unpleasant Bf5. 16. h3? without
restraining the d5 pawn first was probably the losing move, as 16. ... d4 wins
material. Although had the machine been too greedy and grabbed the a2 pawn
with 18. ... Qa2, white would have adequate counter play. We slept a little
easier this night.
Danish team match, game 1
IM Henrik Danielsen (FIDE 2445?) vs. Nordic Deep Blue
Copenhagen, Denmark, 02/25/93 [40/2,SD/1]
1. Nf3 Nf6 2. b3 g6 3. g3 Bg7 4. Bb2 O-O 5. Bg2 d6 6. d4 c5 7. O-O cd4 8. Nd4
d5 9. Na3 e5 10. Nf3 e4 11. Nd4 Nc6 12. c4 Nd4 13. Qd4 Bg4 14. Qe3 Qe7
15. Rab1 Bf5 16. h3 d4 17. Bd4 Qa3 18. Bc5 Qa5 19. Bf8 Rf8 20. g4 Bd7 21. Rfd1
h6 22. a4 Bc6 23. Rd6 h5 24. g5 Ne8 25. Rdd1 f5 26. b4 f4 27. ba5 fe3 28. fe3
Be5 29. c5 Rf5 30. Rd8 Rg5 31. a6 ba6 32. h4 Rg4 33. Rc8 Bc7 34. Rb7
Bb7 35. Re8 Kf7 36. Rh8 Be5 0-1
-----------------------------
Day 3 (Feb. 26, Fri.)
A quiet day. Both match games ended in draws.
The machine had good chances to win the second Larsen match game, but
could not pull it off. This game, together with the first game, made
us realize that the machine needed to be told to open up the position
when it had the bishop pair. Something also needs to be done about
unlike bishops endings. The machine knows that one pawn up in bishops
of opposite color ending is drawish. Unfortunately, some time one extra
rook each does not change the result either.
Larsen match, game 2
Nordic Deep Blue vs. GM Bent Larsen (FIDE 2565?)
Copenhagen, Denmark, 02/26/93 [40/2,20/1,SD/30min]
1. e4 c5 2. Nf3 g6 3. c4 Bg7 4. d4 cd4 5. Nd4 Nc6 6. Be3 Nf6 7. Nc3 O-O
8. Be2 d6 9. O-O Bd7 10. Qd2 Nd4 11. Bd4 Bc6 12. f3 a5 13. b3 Nd7 14. Be3 Nc5
15. Rab1 Qb6 16. Rfc1 Rfc8 17. Rc2 h5 18. Nd5 Bd5 19. cd5 Qb4 20. Qb4 ab4
21. Bd2 Na6 22. Rbc1 Bd4 23. Kf1 Rc2 24. Rc2 Bc5 25. Bd3 Kf8 26. Bb5 Nc7
27. Bh6 Kg8 28. Bd3 Ne8 29. Bd2 Nf6 30. Ke2 Kg7 31. Bb5 h4 32. h3 Rd8 33. Bg5
Rh8 34. Kd3 Rh5 35. Bf4 e5 36. de6 fe6 37. Ba4 b6 38. Bc6 Kf7 39. Bd2 d5
40. ed5 ed5 41. Bf4 Rf5 42. Bc7 Nh5 43. Re2 Nf4 44. Bf4 Rf4 45. Bd5 Kg7
46. Be4 Rf7 47. Kc4 Rd7 48. Bd5 Re7 49. Rc2 Re1 50. Bb7 Kf6 51. Be4 Rd1
52. Kb5 g5 53. Re2 Ra1 54. Bd3 Rd1 55. Bh7 Ra1 56. Kc4 Rg1 57. Kd5 Rd1
58. Kc6 Rf1 59. Bd3 Rf2 1/2-1/2
Hoi got a good position in part because of an oversight in our opening
preparation. Book move is 17. Qd3 instead of 17. Qe4 as in the game. Black
is better after the queen trade. 20. h3?! is strange, but it is tricky.
After 22. Kd2, it looks like 22. ... Ng2 wins for black. In fact, in the
analysis room, people were wondering why Hoi took so long to take the pawn.
Hoi declined the pawn based on intuition. After the game, we checked with
the machine. On 22. ... Ng2, 23. f5 gf5 and then 24. Raf1 threatens to trap
the knight, and with Rf5, Ne4 and Rg1 coming, black might be losing.
Danish team match, game 2
Nordic Deep Blue vs. IM Carsten Hoi (FIDE 2445?)
Copenhagen, Denmark, 02/26/93 [40/2,SD/1]
1. e4 d6 2. d4 g6 3. Nc3 Bg7 4. f4 Nf6 5. Nf3 c5 6. Bb5 Bd7 7. e5 Ng4 8. e6
Bb5 9. ef7 Kd7 10. Nb5 Qa5 11. Nc3 cd4 12. Nd4 Bd4 13. Qd4 Nc6 14. Qc4
Qb6 15. Qe2 h5 16. Bd2 Nd4 17. Qe4 Nf5 18. Qa4 Qc6 19. Qc6 bc6 20. h3 Nge3
21. Be3 Ne3 22. Kd2 Nc4 23. Ke2 Raf8 24. b3 Nb6 25. Rhf1 Rf7 26. Rf3 h4
27. Rd1 Rhf8 28. Ke3 e6 29. a3 a5 30. Rdf1 c5 31. Ke2 Kc6 32. Kd1 Nd5
33. Nd5 ed5 34. a4 d4 35. R1f2 d5 36. g4 hg3 37. Rg3 Rf6 38. Rfg2 Rf4 39. Rg6
R8f6 40. R2g5 Rf1 41. Ke2 Rf2 42. Ke1 Rf1 43. Ke2 Rf2 44. Ke1 1/2-1/2
---------------------------------
Day 4 (Feb. 27, Sat.)
We did not choose the Scandinavian Defence, Danish variation to honor our
host. The 1. e4 e5 defence was busted (well, sort of), and nothing else was
ready. 8. ... o-o-o?? is a horrible mistake by the machine. Probably the
only chess player at the match site that did not think that Larsen was
winning was the machine. 13. ... Bd6! is critical. If 13. ... ed5, then
Bc7! is annihilating. Larsen could have won a pawn by 14. Bd6, but might
have to give up the attack. The attack petered out into a better ending for
black, but the QN vs QB ending would probably require the like of Karpov to
win it.
Larsen match, game 3
GM Bent Larsen (FIDE 2565?) vs. Nordic Deep Blue
Copenhagen, Denmark, 02/24/93 [40/2,20/1,SD/30min]
1. e4 d5 2. ed5 Qd5 3. Nc3 Qa5 4. Nf3 Nf6 5. d4 Bf5 6. Be2 e6 7. O-O Nbd7
8. a3 O-O-O 9. Bf4 Qb6 10. Nb5 Nd5 11. Bg3 a6 12. c4 ab5 13. cd5 Bd6 14. Qb3
Bg3 15. hg3 Nf8 16. a4 Rd5 17. ab5 Kd7 18. Bc4 Rd6 19. d5 Ke7 20. Rfe1 Nd7
21. Qc3 Nf6 22. Nh4 Bg4 23. Re3 Rhd8 24. de6 fe6 25. Rae1 Rd1 26. Qe5 Re1
27. Re1 Rd2 28. Nf5 Bf5 29. Qf5 Rd1 30. Qe5 Re1 31. Qe1 Qd6 32. Qe2 h6
33. b3 Qd7 34. Qf3 b6 35. Qe2 Kf7 36. g4 Qd6 37. g3 Nd5 38. Kg2 Kf6 39. Qf3
Ke7 40. Qe4 Nf6 41. Qg6 Kf8 42. g5 hg5 43. Qg5 Qd4 44. Qc1 Ke7 45. Qg5 Kf7
46. Qc1 Qe5 47. Qd2 Ke7 48. Qb4 Kd7 49. Qd2 Nd5 50. Qd3 g5 51. Qf3 Ke7 52.
Kg1 Qd4 53. Qh5 Qg7 54. Qg4 Kd6 55. Qe4 Qa1 56. Kg2 Qf6 57. Qc2 Qe5 58. Qc1
Kd7 59. Kg1 Ke7 60. Kf1 Kf6 61. Qa3 Kf7 62. Qc1 Ke7 1/2-1/2
The following is probably the best game of the match by the machine. One
commentator went as far as saying that the machine played like a Super
Grandmaster. Lars Bo had recent successes against the 10. ... Be7 line, but
said after the game that he might have to rethink the whole line. DT I used
to win games against GMs by tactics. This game was won by positional play.
The final coup de grace is a tactical combination, but the game was finished
before that. Lars Bo criticized his own 20. Qc3 as tempo losing. Machine
concurred.
Danish team match, game 3
GM Lars Bo Hansen (FIDE 2565?) vs. Nordic Deep Blue
Copenhagen, Denmark, 02/27/93 [40/2,SD/1]
1. Nf3 Nf6 2. c4 e6 3. Nc3 d5 4. d4 Be7 5. Bf4 O-O 6. e3 c5 7. dc5 Bc5 8. Qc2
Nc6 9. a3 Qa5 10. Nd2 Be7 11. Bg3 Bd7 12. Be2 Qb6 13. O-O d4 14. Na4 Qd8
15. b4 Rc8 16. Nb2 e5 17. Nb3 de3 18. fe3 Qe8 19. Rad1 Ng4 20. Qc3 Bg5
21. Bg4 Bg4 22. Rde1 f6 23. c5 Nd8 24. Nd2 Qc6 25. Qc2 b6 26. Ne4 Nf7 27. Nc4
bc5 28. Nc5 Bh5 29. e4 Nd6 30. Nd6 Qd6 31. Bf2 Qd2 32. Qb3 Bf7 33. Qh3 Rfd8
34. Be3 Be3 35. Re3 Bc4 36. Rfe1 Rb8 37. Qg3 a5 38. Rc3 Qd4 39. Qe3 ab4
40. Qd4 Rd4 41. ab4 Rb4 42. Kf2 Rb5 43. Na4 Ra5 44. Nb6 Ba6 45. Nd5 Bb7
46. Ne7 Kf8 47. Nf5 Ra2 48. Kf3 Rdd2 49. Rc7 Rf2 50. Kg4 Rg2 51. Ng3 Raf2
52. h4 g6 0-1
----------------------------------
Day 5 (Feb. 28, Sun.)
Larsen got into early trouble underestimating 14. f5. Machine probably had
an winning edge at move 25. But it willingly went into another unlike bishops
ending without proper assessment. The resultant queen and bishop ending might
be winnable had the queen been traded and the KN pawn (but not the KR pawn)
become passed. The machine, however, was trying hard to avoid trading the
queen. This time, it did not know that certain unlike bishops endings are
winnable.
Larsen match and Danish team match, game 4
Nordic Deep Blue vs. GM Bent Larsen (FIDE 2565?)
Copenhagen, Denmark, 02/28/93 [40/2,SD/1]
1. e4 c5 2. Nf3 d6 3. d4 cd4 4. Nd4 Nf6 5. Nc3 a6 6. a4 g6 7. Be2 Bg7 8. O-O
O-O 9. f4 Nc6 10. Be3 Bd7 11. Nb3 Be6 12. Ra3 Rc8 13. Kh1 Re8 14. f5 Bb3
15. Rb3 Qd7 16. fg6 hg6 17. Nd5 Nd5 18. ed5 Ne5 19. a5 Bf6 20. c3 Kg7 21. Rb4
Rh8 22. Qb3 Rc7 23. Bb6 Rcc8 24. Kg1 Rh4 25. Bd4 Rc7 26. Be5 de5 27. Rh4
Bh4 28. Rf7 Kf7 29. d6 Kg7 30. dc7 Qc7 31. Qb4 Bg5 32. Bf3 b5 33. ab6 Be3
34. Kh1 Bb6 35. Qe4 Qc5 36. Qb1 a5 37. Be4 g5 38. Bh7 Kf8 39. Bf5 Qf2
40. Bg6 Kg7 41. Bh7 Kh8 42. Bf5 Kg7 43. Bd3 Be3 44. Bh7 Kh8 45. b3 Bd2 46. c4
Be3 47. Bg6 Bd4 48. Bf5 Bc5 49. Be4 Be3 50. Qd1 Kg7 51. Qa1 Bd4 52. Qc1
1/2-1/2
-----------------------------------
Final result: Larsen 2.5, DB 1.5
DB 3, Team 1