Karpov-DT Match, Face to Screen
Feng-Hsiung Hsu
Summary: close but no cigar.
Both GM Ron Henley and IM Mike Valvo called it a moral victory for computers,
but, of course, the machine knows nothing about morals.
First, some personal impressions about the match and Karpov. It is sort of
amazing that Danny Edelman, the president of Harvard Chess Club, could pull
this off. While the setup was done by students, it was nonetheless quite
well executed. Karpov is a little bit more plump than I expected. During
the press conference, he seemed to be somewhat nervous--his hands were
twitching behind his back. Compared to Kasparov, Karpov appeared to be less
comfortable with the press. His command of English is not as good as
Kasparov's. Across the board, Karpov has a poker face, while Kasparov
constantly shifts his expressions. Mike Valvo seemed to be able to tell
Karpov's mood changes even from off the stage though.
Before the match, we were expecting to toss a coin to decide the color. GM
Ron Henley, who is going to be Karpov's second for the Timman match, had a
different idea. Suggesting that Karpov had not slept for 24 hours during the
trip, he would like Karpov to have white. The decision was made without our
consent, but given that we were fully expecting to lose with either color, we
had no serious objection.
Karpov's play may have been hampered by the jet lag, but it was also evident
that he was not as well prepared for the match as Kasparov. This might be
in part due to DT's selection of opening line. Over its entire career, it
played Caro-Kann 3 times and none of the games have been widely circulated.
Karpov might have been expecting Alekhine defense (surprisingly, the only
Alekhine defense game that DT lost was the postal game against Valvo...).
Valvo did tell Karpov the day before to expect Caro-Kann.
A slower but safer hardware configuration was used for the Karpov match.
The 6-processor version that played Kasparov had been found to contain
serious software bugs; some of them were fixed, but more bugs showed up
right before the Karpov match, and the old 2-processor version was
resurrected. The machine was operating from IBM T. J. Watson Lab.
Before the match, I predicted that the game would end in an clearly winning,
at least to the titled players, endgame for Karpov and with me resigning for
the machine somewhere between move 40 and 60. I was off by 5 moves and
wrong about the nature of the endgame.
The comments are what I remembered from the postmortem analysis. Valvo will
give his own analysis shortly on the net.
Karpov--DT, 1hr/SD.
1. e4 c6 2. d4 d5 3. Nd2 g6 4. c3 Bg7
5. e5, this move is not in ECO.
5. ... f6!?. DT out of book. According to GM Shamkovich, who saw the game
on USA Today LINC service, this could be a new move.
6. f4 Nh6
7. Ngf3, this move drew criticism from GM Ron Henley. Ndf3, with the idea
of overprotecting the pawn chain base with Ne2 instead of the latter
Nb3 might be better. DT, incidentally, was expecting Ndf3.
7. ... o-o 8. Be2 fe5 9. fe5 c5! a tactical positional move.
10. Nb3 cd4 11. cd4 Nc6 12. o-o Qb6 13. Kh1 a5 14. a4, IM Patrick Wolff
suggested that this move might not be necessary in light of what happened
later.
14. ... Bf5 15. Bg5 Be4 16. Nc5!?, a deep plan by Karpov to get into
a better endgame. In hindsight, DT did not have to go into it. DT was
expecting Bb5, Nf5 and then Nc5. Karpov started to think for long
periods of time for the next few moves.
16. ... Qb2!? Possibly not the best move. One of the commentators suggested
moving the rook on f file. DT saw Ne6 recovered the material for white,
but black standed better. Had the game being a regularly timed event,
it would have found the problem with white's next move and switched to
Nf5, which it assessed, after the game, to lead to a slight advantage
for black. GM Shamkovich also prefers Nf5.
17. Ne4! de4, DT thought it was down one third of a pawn, and predicted
basically the game continuation, without the sidetrack.
18. Rb1 Qa3 19. Bc1 Qc3 20. Bd2 Qa3 21. Bc1 Qc3 22. Rb3 Qa1 23. Bc4 Kh8
The next few moves were played without delays by both man and machine and
caused confusions on the demo board.
24. Bh6 Qd1 25. Bg7 Kg7 26. Rd1 ef3 27. gf3 Ra7!? an ugly move that had
some hidden value. Shamkovich prefers Rad8. On site commentators were
talking about Nb4 as a possibility.
28. Bd5, DT expected this and next move. Wolff thought this was a slip.
28. ... Rd8 29. Rb5 Ra6 30. Be4 Ra7 31. Bd5 Ra6 32. Rc5 Rd7 33. Kg2 Rb6
34. Bc6, forced. bc6 35. Kf2 Rd5, the commentators thought now black standed
slightly better, but still a theoretical draw.
36. Rd5 cd5 37. Rc1 Rb4 38. Ke3 Ra4!? Valvo suggested that Rb3 is an easy
draw. Machine is not aware of the problem with the center pawns yet.
The position is still a draw.
39. Rc5 e6 40. Rc7 Kg8 41. Re7 Ra3 42. Kf4 Rd3 43. Re6 Rd4 44. Kg5 Kf7
45. Ra6 a4 46. f4 h6 47. Kg4 Rc4?! g5 draws instantly, the machine still
thought it was ahead and refused to give up a pawn for safety.
48. h4 Rd4 49. Rf6 Kg7 50. Ra6 Kf7 51. h5! gh5?! machine thought it was
down for the first time in the endgame. g5!? may or may not draw
at this point. Right after the game, Karpov thought he would still
be winning. Afterwards, on the cab back to hotel, he changed the
assessment to slightly better for white. After the game move, black
is theoretically lost. According to Valvo, "now he (Karpov) is happy".
Karpov at this time was down to his last few minutes, but it was easy
for him from now on.
52. Kf5 Kg7 53. Ra7 Kf8 54. e6 Re4 55. Rd7 Rc4 56. Rd5 h4 57. Rd3 Ke7
58. Rd7 Kf8 59. Rh7 h5 60. Ke5 h3 61. f5 Kg8 62. Rh5 a3 63. Rh3 a2?
A human player would play Ra4 to prolong the struggle. 64. Ra3 Rc5
65. Kf6 resigns. DT for the first time saw itself down by at least 6
pawns.
Karpov was down to his last 40-50 seconds. DT had about 20 min left.
We are expecting to take a 6-month hiatus from active play to retool the
entire software. Afterwards, DT should never lose this ending again.
Murray Campbell
> In article <78...@pt.cs.cmu.edu> f...@vlsi.cs.cmu.edu (Feng-Hsiung Hsu) writes:
> >[.......]
>
> >63. ... a2?
> > A human player would play Ra4 to prolong the struggle.
>
>
> --Noam D. Elkies (elk...@zariski.harvard.edu)
DT saw 64.e7, and if Kf7 65.Rh7+ followed by f6 and Rh8+. In fact,
with some further search, DT sees that 65. Rh7+ is mate in 14 or less.
Murray