DT2 games in ACM 94
Feng-Hsiung Hsu
sleep.
DT2 - Zarkov
Round 1, ACM 94
1. e4 e5 2. f4 exf4 3. Nf3 g5 4. h4 g4 5. Ne5 Nf6
6. d4 d6 7. Nd3 Nxe4 8. Bxf4 Bg7 9. c3 O-O 10. Nd2 Re8
11. Nxe4 Rxe4 12. Be2 Qe8 13. Kd2 h5 14. Re1 c5 15. dxc5 dxc5
16. g3 Na6 17. Bf1 Bf5 18. Qb3 Qc6 19. Rxe4 Bxe4 20. Nf2 Rd8
21. Ke1 Bd5 22. Qb5 Qe6 23. Qe2 Qd7 24. Rd1 Re8 25. Ne4 Qa4
26. Rxd5 Rxe4 27. Rd8 Kh7 28. Be3 Bh6 29. b3 Qc6 30. Bg2 Rxe3
31. Bxc6 bxc6 32. Rd7 c4 33. b4 Kg7 34. Kf2 Rxe2 35. Kxe2 c5
36. b5 Nb4 37. cxb4 cxb4 38. Rxa7 b3 39. axb3 cxb3 40. Ra3 Kg6
41. b6 Bg7
1-0
A lucky escape. DT2 was out of book on move nine. 12. Be2? could be
a mistake and might have caused the subsequent trouble. Book move is Kf2.
The collapse of black's position was surprising.
MChess - DT2
Round 2, ACM 94
1-0
Heavy rain induced power outage at Hawthorne, NY. [4 inches in
one hour.] Possibility of rescheduling was discussed. Mchess
decided to take a time forfeit win.
DT2 - Wchess
Round 3, ACM 94
1. Nf3 Nf6 2. e3 Nc6 3. d4 e6 4. c4 d5 5. Be2 Be7
6. Nc3 O-O 7. o-o Bd7 8. Bd2 dxc4 9. Bxc4 Bd6 10. Rc1 e5
11. d5 Ne7 12. e4 b5 13. Bd3 b4 14. Na4 c6 15. dxc6 Bxc6
16. Bg5 Bxe4 17. Bxf6 Bxd3 18. Qxd3 gxf6 19. Rcd1 Bc7 20. Qa6 Qe8
21. Nc5 Qc6 22. Qxc6 Nxc6 23. Nd7 Rfd8 24. Nxf6 Kh8 25. Rc1 Rd6
26. Ne4 Rh6 27. Rc4 Rb8 28. Rd1 Kg8 29. Neg5 b3 30. a3 Rc8
31. Rd7 Rd6 32. Rxd6 Bxd6 33. Ne4 Bxa3 34. bxa3 b2 35. Nfd2 Rd8
36. g4 Rd4 37. Rxc6 Rxe4 38. Rc8 Kg7 39. Rb8 Rxg4 40. Kf1 b1Q
41. Rxb1 Ra4 42. Rb3 Kf6 43. Rh3 Ke6 44. Rxh7 Rxa3 45. Rh6 Ke7
46. Nc4 Ra1 47. Kg2 f6 48. Ne3 Kf7 49. h4 Kg7 50. Nf5 Kf7
51. h5 Ke6 52. Ng3 Ra4 53. Rh7 f5 54. Rh6 Kf7 55. Nxf5 Rg4
56. Kh3 Rf4 57. Nd6 Ke7 58. Kg3 a5 59. Nc8 Kf8 60. Ra6 Kg7
61. Rxa5 Rf5 62. Kg4 Rxf2 63. Rxe5 Rg2 64. Kh3 Ra2 65. Nd6 Kf6
66. Rc5 Ra6 67. Rf5 Ke6 68. Rg5 Ra3 69. Rg3 Ra1 70. Rg6 Ke7
71. h6 Rd1 72. Kg2 Rd5 73. h7 Rh5 74. Nf5 Kf7 75. Nh6 Kxg6
76. h8Q Rxh6 77. Qf8 Kg5 78. Kg3 Rf6 79. Qd8 Kh5 80. Qd5 Kg6
81. Kg4 Kg7 82. Qd7 Rf7 83. Qd8 Rf1 84. Qd4 Kh6 85. Qe3 Kg7
86. Kg5 Rf7 87. Qc3 Kf8 88. Kg6 Rd7 89. Qe5 Rd1 90. Kf6 Rf1
91. Ke6 Kg8 92. Qg3
1-0
WChess finally saw the mate.
StarSocrates - DT2
Round 4, ACM 94
1. e4 c5 2. Nc3 Nc6 3. Nge2 Nf6 4. d4 cxd4 5. Nxd4 d6
6. Bg5 e6 7. Qd2 a6 8. o-o-o h6 9. Bf4 Bd7 10. Nxc6 Bxc6
11. f3 d5 12. Qe1 Bb4 13. a3 Ba5 14. Bd2 O-O 15. exd5 exd5
16. Bd3 Re8 17. Qh4 d4 18. Na2 Bxd2 19. Rxd2 a5 20. Bc4 b5
21. Rxd4 Qe7 22. Bf1 Qe3 23. Rd2 b4 24. Qd4 bxa3 25. Qxe3 axb2
26. Kxb2 Rxe3 27. Rd6 Rb8 28. Kc1 Ra3 29. Rxc6 Rxa2 30. g3 Ra1
31. Kd2 a4 32. Bg2 Rd8 33. Ke2 Rxh1 34. Bxh1 Ra8 35. Rb6 Nd5
36. Rd6 Nc3 37. Kd3 a3 38. Kxc3 a2 39. Rd1 a1Q 40. Rxa1 Rxa1
41. Bg2 Rg1 42. Bh3 Rh1 43. Bc8 Rxh2 44. g4 Rf2 45. Bb7 g6
46. Kd3 h5 47. gxh5 gxh5 48. Be4 h4 49. Ke3 Rg2 50. Bf5 Rg5
51. Bh3 Rg3 52. Bf1 h3 53. Kf2 h2 54. Bg2 Rg7 55. f4 f5
56. Kf3 Kf7 57. Kf2 Rg4 58. Kf3 Ke7 59. Kf2 Rg8 60. Kf1 Kd6
61. Kf2
0-1
A pleasant win if you are not StarSocrates. They were outsearching
us in ply depth by 2 to 3 plies according to their printouts. We
were probably outsearching them by 10 plies in the critical lines.
MChess - DT2
Round 5, ACM 94
1. e4 c5 2. Nf3 Nc6 3. d4 cxd4 4. Nxd4 e5 5. Nb5 d6
6. N1c3 a6 7. Na3 b5 8. Nd5 Nce7 9. c4 Nxd5 10. exd5 bxc4
11. Nxc4 Nf6 12. Be3 Rb8 13. a4 Be7 14. Be2 O-O 15. o-o Bf5
16. f4 Be4 17. fxe5 Nxd5 18. Qd4 Nxe3 19. Nxe3 d5 20. Bxa6 Rb4
21. Qa7 Rxb2 22. Qd4 Qb8 23. Bb5 Rb4 24. Qd2 Qxe5 25. Ng4 Qh5
26. h3 Bc5 27. Kh2 Bd6 28. Kg1 Rb3 29. Be2 Bc5 30. Kh2 Qh4
31. Rf3 Bxf3 32. Bxf3 h5 33. Qxd5 hxg4 34. Qxb3 Bd6 35. Kg1 gxf3
36. Qb2
0-1
Since the round 2 game was never played, the committee decided that
MChess would be paired as if it was given a bye, and hence the
pairing.
Hal Bogner
>Just got back. I am getting too old for this. Time to get some
>sleep.
Me too! But there's one thing I want to elabortae on before I retreat into the
Sierra for a weekend of recovery.
[Rd 1 stuff deleted]
>MChess - DT2
>Round 2, ACM 94
>1-0
>Heavy rain induced power outage at Hawthorne, NY. [4 inches in
>one hour.] Possibility of rescheduling was discussed. Mchess
>decided to take a time forfeit win.
>
Not so, though this happened away from you, so you may not realize what
actually happened. The decision had nothing to do with MChess or it's
programmer, Marty Hirsch.
IM Mike Valvo, who has served as Tournament Director for these events for over
a decade, made the decision in consultation with ICCA President Tony Marsland
and Monty Newborn, who is responsible for the ACM's chess activity.
The round was scheduled to begin at 7 PM. The decision was taken at 10:45 PM,
when there still seemed no prospect of being able to play that night. The
fact that DT2 was the favorite was not considered, though it might have caused
the decision to be deferred a few hours later than it could have been made.
The issue was how the possible postponement of the game could affect the final
results of the tournament. Rounds three and four were both scheduled for the
next day, and would not have been possible to make up the game on that day.
Round five was scheduled Monday evening, with the day free. But playing the
round two game after round four would very likely have lead to a distortion of
the pairings, and possibly of the results. Maybe not for DT2, but certainly
for some competitors. And all of the competitors merit equal consideration
under the rules.
It was a painful decision to have to make, but I think it was the only one
possible under the circumstances. And the forfeit loss does not reflect badly
on you...these things happen sometimes. That you were able to recover and win
the tournament is to your credit. But had you not done so, it would still not
justify reconsidering the decision and judging it to have been wrong.
[more stuff deleted]
>
>MChess - DT2
>Round 5, ACM 94
>1. e4 c5 2. Nf3 Nc6 3. d4 cxd4 4. Nxd4 e5 5. Nb5 d6
>6. N1c3 a6 7. Na3 b5 8. Nd5 Nce7 9. c4 Nxd5 10. exd5 bxc4
>11. Nxc4 Nf6 12. Be3 Rb8 13. a4 Be7 14. Be2 O-O 15. o-o Bf5
>16. f4 Be4 17. fxe5 Nxd5 18. Qd4 Nxe3 19. Nxe3 d5 20. Bxa6 Rb4
>21. Qa7 Rxb2 22. Qd4 Qb8 23. Bb5 Rb4 24. Qd2 Qxe5 25. Ng4 Qh5
>26. h3 Bc5 27. Kh2 Bd6 28. Kg1 Rb3 29. Be2 Bc5 30. Kh2 Qh4
>31. Rf3 Bxf3 32. Bxf3 h5 33. Qxd5 hxg4 34. Qxb3 Bd6 35. Kg1 gxf3
>36. Qb2
>0-1
>Since the round 2 game was never played, the committee decided that
>MChess would be paired as if it was given a bye, and hence the
>pairing.
>
Not a bye, per se. It was a forfeit win. The most basic rule of the Swiss
Syatem is that no opponents should play each other twice (please take note,
Mr. Goichberg!). Since no game had been played in round two, there was no
reason to avoid this pairing in round five.
-hal
Peter Gillgasch
|> StarSocrates - DT2
|> Round 4, ACM 94
<game deleted>
|> A pleasant win if you are not StarSocrates. They were outsearching
|> us in ply depth by 2 to 3 plies according to their printouts. We
|> were probably outsearching them by 10 plies in the critical lines.
Does this mean that
a) the CM-5 w. StarSocrates had a larger brute-force look-ahead?
and
b) DTII won the game because of some clever selective deepening
heuristics (e.g. singular extensions)?
I think this is note-worthy because DT2 has still more horsepower than
a CM-5 (regarding chess of course ;-) and StarSocrates was derived
from a Heuristic Software product (I am not sure about how much of it
came from StarTech or Socrates, an infos?) and hence should be expected
to have a very effective selective search.
Any enlightening comments from the DT or the StarSocrates team?
Peter
W C Newell Jr
>sleep.
Ready to start reading all the analysis?
>DT2 - Zarkov
>Round 1, ACM 94
>1. e4 e5 2. f4 exf4 3. Nf3 g5 4. h4 g4 5. Ne5 Nf6
>6. d4 d6 7. Nd3 Nxe4 8. Bxf4 Bg7 9. c3 O-O 10. Nd2 Re8
>11. Nxe4 Rxe4 12. Be2 Qe8 13. Kd2 h5 14. Re1 c5 15. dxc5 dxc5
>16. g3 Na6 17. Bf1 Bf5 18. Qb3 Qc6 19. Rxe4 Bxe4 20. Nf2 Rd8
>21. Ke1 Bd5 22. Qb5 Qe6 23. Qe2 Qd7 24. Rd1 Re8 25. Ne4 Qa4
>26. Rxd5 Rxe4 27. Rd8 Kh7 28. Be3 Bh6 29. b3 Qc6 30. Bg2 Rxe3
>31. Bxc6 bxc6 32. Rd7 c4 33. b4 Kg7 34. Kf2 Rxe2 35. Kxe2 c5
>36. b5 Nb4 37. cxb4 cxb4 38. Rxa7 b3 39. axb3 cxb3 40. Ra3 Kg6
>41. b6 Bg7
>1-0
>A lucky escape. DT2 was out of book on move nine. 12. Be2? could be
>a mistake and might have caused the subsequent trouble. Book move is Kf2.
>The collapse of black's position was surprising.
A very interesting game. As a seasoned KGA player, esp. against CM4000, I
have a few comments which might be of interest.
12. Be2 is playable. In this position White wants to consolidate his
development while at the same time booting that rook out of e4. There are two
main alternatives available. The first plan involves attacking the rook with
the KB. After 12. Kf2, White intents Qd2, g3 and Bg2 in that order. This
claims the long diagonal and has the added advantage of minimizing the impact
of ...c5. Regardless of what Black does, White should get adequate positional
compensation for the material, but that's about all.
The second plan involves placing immediate pressure on g4, and using the knight
to force the rook back. After 12. Be2, White intends Qd2, Nf2 and either Kf1
or O-O depending on Black's responses. Black must choose between the lesser of
two evils in defending g4. If ...h5, then White threatens eventual Bh6 and/or
a possible sac at g4. If ...Bf5 then Nd1-Ne3.
13. Kd2? is a blunder. In most lines of the KGA, White eschews castling except
when it fits his plan as an offensive maneuver. The king becomes a mobile part
of the defense, moving to one side or the other depending on where Black
focuses his counterattack. In this line, the g-pawn is White's critical
weakness. The K must stay on his side of the board to defend it. If not,
Black can focus his attack there and get enough counterplay to maintain his
advantage. The game effectively demonstrates this, although Zarkov frittered
it away later on.
In playing the KGA against CM4000, I have found that its default parameters
leave it overly worried about back-line defense and thus it doesn't play with
a proper "feel" for the opening. It castles at inopportune moments. However,
after I set the "King Safety" parameter to 0, it gave me a very aggressive game
and chose its early moves consistent with book [ECO] lines. I have concluded
that there ought to be some preferred parameter adjustments for certain
openings, and that it might be worthwhile to store such information in the
program's book.
As an excercise, I turned CM4000 loose on the key position above. With default
settings, it played the same moves as DTII. With "King Safety" set to 0, it
vacillated between 12. Be2 and 12. Kf2 before converging on the former after
11 plies. After 12...Qe8 it looked at 13. Kf1 for a long time before switching
to Qd2 after 11 plies. I would also consider 12...Qe8?!, preferring either
...Nd7 or ...Bf5, but I can see how a program which assumes White needs to
castle would give added weight to ...Qe8.
Make any sense?
I also note that CM4000 can't handle most QP vs QP endings unless one sets the
King Safety to 0. I spent a lot of time studying Q4/c #826 in ECE after
playing a game which led to a virtually identical position. For those who
don't have the reference, here's the position, Black to move:
White: Kh2, Qd2, Pg3, Ph4.
Black: Kg7, Qc4, Pf7, Pg6, Ph5.
CM4000 finds the winning move 1...Kf6! iff King Safety = 0. But that's really
another thread; I shouldn't be digressing so much here.
--
Bill Newell
Douglas Lee Stewart
>>Since the round 2 game was never played, the committee decided that
>>MChess would be paired as if it was given a bye, and hence the
>>pairing.
>>
>
>Not a bye, per se. It was a forfeit win. The most basic rule of the Swiss
>Syatem is that no opponents should play each other twice (please take note,
>Mr. Goichberg!). Since no game had been played in round two, there was no
>reason to avoid this pairing in round five.
I believe the correct way to say this is that the game was paired with
round two being an unplayed game, rather than a forfeit or bye.
Douglas Stewart
President, Mississippi Chess Association
Douglas Lee Stewart
>In article <Cs64r...@hawnews.watson.ibm.com>, f...@watson.ibm.com (Feng-Hsiung Hsu) writes:
>
>|> StarSocrates - DT2
>|> Round 4, ACM 94
>
><game deleted>
>
>|> A pleasant win if you are not StarSocrates. They were outsearching
>|> us in ply depth by 2 to 3 plies according to their printouts. We
>|> were probably outsearching them by 10 plies in the critical lines.
>
>Does this mean that
>
>a) the CM-5 w. StarSocrates had a larger brute-force look-ahead?
>
>and
>
>b) DTII won the game because of some clever selective deepening
> heuristics (e.g. singular extensions)?
Or was he c) being sarcastic ("Oh you were playing better the whole way..
right up until you lost"). :)
Feng-Hsiung Hsu
>In article <Cs64r...@hawnews.watson.ibm.com> f...@watson.ibm.com (Feng-Hsiung Hsu) writes:
>>MChess - DT2
>>Round 2, ACM 94
>>1-0
>>Heavy rain induced power outage at Hawthorne, NY. [4 inches in
>>one hour.] Possibility of rescheduling was discussed. Mchess
>>decided to take a time forfeit win.
>
>Not so, though this happened away from you, so you may not realize what
>actually happened. The decision had nothing to do with MChess or it's
>programmer, Marty Hirsch.
Actually, you missed what really happened. I was under the impression
that Martin might be willing to reschedule the game, and asked the TD
whether rescheduling was possible if Martin agreed to it. Martin was then
asked to make the awkward decision to reschedule the game or not. I don't
know what I would have done in his position. After long deliberations,
he decided not to. I apologized to him that he had to make the decision, but
I had to ask.
>>MChess - DT2
>>Round 5, ACM 94
...
>>Since the round 2 game was never played, the committee decided that
>>MChess would be paired as if it was given a bye, and hence the
>>pairing.
>Not a bye, per se. It was a forfeit win. The most basic rule of the Swiss
>Syatem is that no opponents should play each other twice (please take note,
>Mr. Goichberg!). Since no game had been played in round two, there was no
>reason to avoid this pairing in round five.
Sorry about the mistake. I am not a NTD. That was the best I could make
out of your explanations during the tourney.
Feng-Hsiung Hsu
Peter Gillgasch <gil...@i41s19.ira.uka.de> wrote:
>Does this mean that
>
>a) the CM-5 w. StarSocrates had a larger brute-force look-ahead?
>
StarSocrates does not have search extensions, but it does use selective
pruning. (d-2 null move pruning). Their ply depth has a quite different
meaning from ours.
We don't do null move pruning. I prefer to be conservative.
>b) DTII won the game because of some clever selective deepening
> heuristics (e.g. singular extensions)?
>
>I think this is note-worthy because DT2 has still more horsepower than
>a CM-5 (regarding chess of course ;-) and StarSocrates was derived
>from a Heuristic Software product (I am not sure about how much of it
>came from StarTech or Socrates, an infos?) and hence should be expected
>to have a very effective selective search.
Just to show you that all those people that claim their programs would
be better than DT2 when ported to the same hardware are wrong:-).
StarSocrates' node rate was roughly in the same ballpark as DT2.
I don't know the real answer to your question. I know what we did, but I
don't know what the StarSocrates people did. They seem to be of the opinion
that selective pruning is much preferable to selective extensions.
I believe that both mechanisms are needed.
It should also be borne in mind that StarSocrates probably did not
reach its full potential at the time of the competition.
Robert Hyatt
>
>|> StarSocrates - DT2
>|> Round 4, ACM 94
>
><game deleted>
>
>|> A pleasant win if you are not StarSocrates. They were outsearching
>|> us in ply depth by 2 to 3 plies according to their printouts. We
>|> were probably outsearching them by 10 plies in the critical lines.
>
>Does this mean that
>
>a) the CM-5 w. StarSocrates had a larger brute-force look-ahead?
>
NO! StarSocrates uses the typical micro-processor-based-program
technique that is based on selective search. When they say they are
doing a 12 ply search, this is really some "n" ply exhaustive (n < 12)
and "m" ply selective (n+m=12). In short, when they are doing a 12
ply search, they are doing something like 7 (a guess) exhaustive ply
and the rest selective. When Deep Thought (or Cray Blitz) says a
10 ply search, we are talking about 10 plies exhaustive + selective
extensions. Sort of apples and oranges comparisons...
>and
>
>b) DTII won the game because of some clever selective deepening
> heuristics (e.g. singular extensions)?
No again. DT was doing a deep exhaustive search + singular extensions
+ who knows what else. Star Socrates simply wasn't looking along the
*right* lines and really "got a chess lesson" (Valvo's words...)
>
>I think this is note-worthy because DT2 has still more horsepower than
>a CM-5 (regarding chess of course ;-) and StarSocrates was derived
>from a Heuristic Software product (I am not sure about how much of it
>came from StarTech or Socrates, an infos?) and hence should be expected
>to have a very effective selective search.
D-T II didn't have a lot more "horsepower" since a 512 node CM5 represents
a *lot* of computing power. They were close in horsepower, but *not* in
search algorithms....
>
>Any enlightening comments from the DT or the StarSocrates team?
>
>Peter
--
Robert Hyatt Computer and Information Sciences
hy...@cis.uab.edu University of Alabama at Birmingham
(205) 934-2213 115A Campbell Hall, UAB Station
(205) 934-5473 FAX Birmingham, AL 35294-1170
Hal Bogner
>In article <hmbCs7...@netcom.com>, Hal Bogner <h...@netcom.com> wrote:
>>In article <Cs64r...@hawnews.watson.ibm.com> f...@watson.ibm.com (Feng-Hsiung Hsu) writes:
>>>MChess - DT2
>>>Round 2, ACM 94
>>>1-0
>>>Heavy rain induced power outage at Hawthorne, NY. [4 inches in
>>>one hour.] Possibility of rescheduling was discussed. Mchess
>>>decided to take a time forfeit win.
>>
>>Not so, though this happened away from you, so you may not realize what
>>actually happened. The decision had nothing to do with MChess or it's
>>programmer, Marty Hirsch.
>
>Actually, you missed what really happened. I was under the impression
>that Martin might be willing to reschedule the game, and asked the TD
>whether rescheduling was possible if Martin agreed to it. Martin was then
>asked to make the awkward decision to reschedule the game or not. I don't
>know what I would have done in his position. After long deliberations,
>he decided not to. I apologized to him that he had to make the decision, but
>I had to ask.
>
Hmmm. We both missed different parts of what happened, it turns out. I don't
know if Marty's point of view informed any of the discussion I was present
for; the discussion when decision time came (10:45 or so) between Mike Valvo,
Tony Marsland, and Monty Newborn didn't turn on MChess' interests at all.
Mike made the decision based on the fact that waiting until after rd. 4 would
very likely impact the pairings and results for an unknown number of the
participants, including the likely winners.
>>>MChess - DT2
>>>Round 5, ACM 94
>...
>>>Since the round 2 game was never played, the committee decided that
>>>MChess would be paired as if it was given a bye, and hence the
>>>pairing.
>>Not a bye, per se. It was a forfeit win. The most basic rule of the Swiss
>>Syatem is that no opponents should play each other twice (please take note,
>>Mr. Goichberg!). Since no game had been played in round two, there was no
>>reason to avoid this pairing in round five.
>
>Sorry about the mistake. I am not a NTD. That was the best I could make
>out of your explanations during the tourney.
>
>
Semantics. Hardly worth mentioning.
On a different subject: Did DT2 see a winning line for Zarkov in their game?
-hal
Feng-Hsiung Hsu
>On a different subject: Did DT2 see a winning line for Zarkov in their game?
It saw a few winning lines against some of the moves that Zarkov was expecting,
but not against its actual defence. It might have an easier game playing
some of the bad defences as Zarkov wasn't able to see the killer moves, but who
knows what would have happened then? BTW, John was excited at one point as
Zarkov did see a killer line against its predicted response, but DT2 never had
any inclination to play that particular prediction. My guess is that DT2 was
never practically lost during the game, despite the pawn deficit. Maybe it
would be lost against a decent Grandmaster, but not against Zarkov. Programs
are notoriously bad in executing attacks, all the claims of Grandmaster level
tactics notwithstanding. Even DT2 is lacking in this regard. Of the four
games from ACM, it only executed the attack in the last game well. Newborn
refered to the win over StarSocrates as one of the most brilliant games in
computer chess history, but it would have been a much better game had it played
b4 first before playing Qe3+, as then white would not have the chance to trade
the queens.
Mark Uniacke
Hi,
In article <CsHH4...@hawnews.watson.ibm.com>, f...@watson.ibm.com (Feng-Hsiung Hsu) writes:
> would be lost against a decent Grandmaster, but not against Zarkov. Programs
> are notoriously bad in executing attacks, all the claims of Grandmaster level
> tactics notwithstanding. Even DT2 is lacking in this regard. Of the four
> games from ACM, it only executed the attack in the last game well. Newborn
Yes, but when an attack does come off what a great feeling!
The following game is probably the best attacking game Hiarcs has ever played
especially considering the strong opponent and the importance of the game.
[Event "12th World Microcomputer Chess Championship"]
[Site "Munich"]
[Date "1993"]
[Round "8"]
[White "Hiarcs"]
[Black "Genius 2"]
[Result "1-0"]
1.d4 b6 2.e4 Bb7 3.Bd3 e6 4.Nf3 Nf6 5.e5 Ne4 6.O-O Be7 7.Re1 f5 8.ef6 Nxf6
9.Nc3 O-O 10.Ng5 h6 11.Nh3 Nc6 12.Be3 Nb4 13.Bg6 Nfd5 14.Bd2 Nxc3 15.bxc3
Nc6 16.Qh5 Ba3 17.Ng5 Qf6 18.Ne4 Qd8 19.Rab1 Ne7 20.Bxh6 gxh6 21.Qxh6 Nxg6
22.Qxg6+ Kh8 23.Qh6+ Kg8 24.Re3 Kf7 25.Rf3+ Ke8 26.Qg6+ Ke7 27.Qg7+ Ke8
28.Rxf8+ Bxf8 29.Nf6+ Qxf6 30.Qxf6 Bd5 31.a4 Bd6 32.f3 a5 33.Qg6+ Ke7
34.Qg7+ Ke8 35.Rb5 Bf8 36.Qg6+ Ke7 37.Rxd5 exd5 38.h4 c6 39.h5 Re8 40.h6
Bxh6 41.Qxh6 Kd8 42.Kf2 Kc7 43.f4 b5 1-0
20. Bxh6 was initially found in 1 second on purely positional grounds, although
deeper search showed it was winning tactically as well. Although it would have
been nice to end the sequence with a mate :), alas none was found :(. Perhaps
DTII could find a mating sequence somewhere between moves 21-29 ?
Cheers,
Mark
--
The opinions and comments expressed herein are my own and do not in anyway
represent those of BNR Europe or Northern Telecom.
ma...@bnr.co.uk
Bradley Kuszmaul
authors of *Socrates. This letter consists of a collection of separate
paragraphs, each of which addresses some minor point.
When *Socrates says it is doing a 10 ply search, it is 10 ply plus
quiescence. (That is, contrary to Bob Hyatt's assertion, we seem to
measure ply the same way that Cray Blitz does.)
Remember that selective search extensions cost time. Clearly, the DT2 team
has decided that they would rather have 10 additional ply along the
"critical variations" and have fewer ply of full-width search. The
*Socrates team didn't really make a conscious choice on this decision.
(See the discussion below regarding the state of *Socrates.) Thus we are
comparing 8 ply plus extensions plus quiescence for DT2 against 10 ply plus
quiescence and some selectivity for *Socrates. This is comparison between
search plys is difficult to make, but the overall advantage probably favors
DT2.
The problem with the move
18. Na2 ....
in our game with DT, had nothing to do
with searching down the right paths or the wrong paths. *Socrates simply
an error in judging the position. I may be misremembering, but I think
that DT2 suggested the same move at the same time. It was simply a case of
the computers not knowing how to play the position. We were unlucky enough
to actually play the bad move. This sort of thing explains why one game
makes it difficult to judge the strenght of a program. (However, there is
good evidence that DT2 is better than *Socrates: For example, *Socrates
managed to lose to Zarkov from a dead drawn position. I doubt DT2 would
have lost.)
As F.H. Hsu graciously points out, the *Socrates program was not up to
snuff. In fact, it had only played its first full game (without crashing)
on the day before the tournament. Some background: We didn't start porting
the Socrates program to the CM-5 until about May 27. We did the porting by
writing a new version of th eprogram, rather than porting the old one
directly. As a result, much of the chess knowledge of Socrates did not
appear in *Socrates. It is not clear that at the time of the tournament,
that *Socrates was any better than Socrates.
How much "horsepower" is there in these machines? Again it is diffucult to
compare 512 SPARC's to a dozen chess chips. The chess chips are capable of
very fast move generation and static evaluation. On the other hand, the
SPARC's can do anything we can program. The CM-5 has 60 gigaflops of peak
vector floating point performance, and we did not use the vector units at
all for this program. A more reasonable measure is that the 512 SPARC's
provide about 10,000 MIPS (peak). How many MIPS are there in DT2? The
question does not even make sense. Hsu says that DT2 consistently searched
at least 3 million nodes per second. For the serial version of Socrates to
search 3e6 nodes per second would require a 12,000 MIPS SPARC processor.
For chess, there is a lot of horsepower in DT2. For other things there is
a lot of horsepower in a CM-5. It will likely be very difficult to beat
special purpose chess hardware using a general parallel machine.
Overall, I was very happy with the performance of *Socrates at the
tournament. We need to make some improvements to the program, and it is
clear what many of those improvements should be. Happily, the search
engine is working quite effectively. There is much chess knowledge from
Socrates that should be put into the parallel version. I don't think there
is any lesson about whether "if you port your program to fast hardware it
might be better than DT". There is simply insufficient evidence. My
personal belief is that by the time we port all of Socrates to the parallel
program that it will be stiff competition for DT.
Congratulations to Deep Thought II, Zarkov, and NOW, for their excellent
performance in the tournament.
-Bradley
Robert Hyatt
>I would like to add a few points to this discussion. I am one of the
>authors of *Socrates. This letter consists of a collection of separate
>paragraphs, each of which addresses some minor point.
>
>When *Socrates says it is doing a 10 ply search, it is 10 ply plus
>quiescence. (That is, contrary to Bob Hyatt's assertion, we seem to
>measure ply the same way that Cray Blitz does.)
My error. Last year's Socrates operated just as I stated... that is, it
did something like (I believe in our game which ended in a draw) 5 plies
exhaustive and 4 plies selective or some such. They were reporting 9 ply
iterations based on this, even though it was comparing apples and oranges.
If the "selective" phase of Socrates has been "toned down" then that's a
pretty interesting idea since mainstream chess has been going the other way
for several years (ie, DT II has steadily searched shallower and shallower
despite hardware speeds going up, primarily due to trying to follow those
all-important lines as deep as necessary.)
However, back to a reality check. In general, with DT II using 12 processors
to search 3-4M nodes per second, and we assume *Socrates can do 3-4M nodes per
second on a 512 node machine, DT II is still going to have at least a 1 ply
advantage (I'm assuming equivalent programs here, just the hardware is
different -- same evaluation, search extensions, etc.) Why? because the
alpha/beta algorithm gets worse and worse as processors are added due to the
inherent sequential property of the search... Yes, I know that we all do
pretty well with a parallel search, but I (and others) get more out of a
4 pprocessor machine than we would out of an 8 processor machine with each
cpu only 1/2 as fast as the 4 processor box, even though total computing power
is equivalent. In short, the speed of a single processor is still an important
measure of performance. Granted, that given a machine with power X, then
two processors gives 2X, 4 give 4X, etc. However for chess, this begins to
tail off pretty rapidly. Hence my "edge" given to DT II based on speed.
Another important point to the discussion of parallel performance is just how
good is the serial code before we parallelize it? Tony Marsland ran a test
years ago with a really bad version of Belle (parabelle) that produced really
good speed-up curves. When they compared this with a program with proven good
move ordering, parabelle produced much better parallel search speedups. I have
personally seen this occur in Cray Blitz. For every enhancement to move
ordering, there is a corresponding drop-off in overall parallel performance.
Granted, my testing also showed that when we reach perfect move ordering,
speedup is again near-perfect, but this is pretty much hopeless to think
about.
I'll paraphrase Ken Thompson and say "give me one fast processor as opposed to
many slow ones any day." Yes, I'll take two (or 16 on a C90) fast processors
over 1 fast one, but I would never take 128 slow ones over my 1 fast one, even
if the 128 slow ones offer greater total throughput.
Also, don't forget, that by the time you "get Socrates completely ported to the
CM-5" that Hsu and company will be searching 3-5M nodes on a single processor.
Then things will really get interesting. I have no idea what 128 of those will
do, but I strongly suspect that if they can reach a speed-up of even 32 it will
be dynamite.
I hope no one took my comment as a put-down of *Socrates, I was simply
replaying a "common" discussion that recurs often enough to mention.
I have "been there", "done that", etc. myself with a program that has not
played a game. We won the 1983 world championship with a program that had
not played a single game. In fact, While we were driving to New York from
Hattiesburg, Mississippi, Harry Nelson was in Minneapolis at Cray trying to
track down some bizarre bug. We were working on the program, adding routines
that were temporarily "disabled", and otherwise debugging even after the first
round was over. Not fun, but with a little luck it worked out. You guys did
well to get things working. However, regardless of how well we play, the DT II
program is still in a class by itself and likely will remain there. However,
we can still create a title of something like "world computer chess champion
on a general purpose computer" and then do battle... Just as long as Hsu
doesn't try to convince me that his chip can add subtract multiply and divide
like a real machine. If it can run a C program, I'll buy that argument..
:^)
In any case, I look forward to seeing you guys next year at whenever. Hopefully
you can get a '5e and we can get a Triton... That'll improve things quite a
bit for both sides....
Bob
Feng-Hsiung Hsu
> Yes, but when an attack does come off what a great feeling!
>The following game is probably the best attacking game Hiarcs has ever played
>especially considering the strong opponent and the importance of the game.
Actually, you just proved my point about computers being poor attackers.
>
>[Event "12th World Microcomputer Chess Championship"]
>[Site "Munich"]
>[Date "1993"]
>[Round "8"]
>[White "Hiarcs"]
>[Black "Genius 2"]
>[Result "1-0"]
>
>1.d4 b6 2.e4 Bb7 3.Bd3 e6 4.Nf3 Nf6 5.e5 Ne4 6.O-O Be7 7.Re1 f5 8.ef6 Nxf6
>9.Nc3 O-O 10.Ng5 h6 11.Nh3 Nc6 12.Be3 Nb4 13.Bg6 Nfd5 14.Bd2 Nxc3 15.bxc3
>Nc6 16.Qh5 Ba3 17.Ng5 Qf6 18.Ne4 Qd8 19.Rab1 Ne7 20.Bxh6 gxh6 21.Qxh6 Nxg6
>22.Qxg6+ Kh8 23.Qh6+ Kg8 24.Re3 Kf7 25.Rf3+ Ke8 26.Qg6+ Ke7 27.Qg7+ Ke8
>28.Rxf8+ Bxf8 29.Nf6+ Qxf6 30.Qxf6 Bd5 31.a4 Bd6 32.f3 a5 33.Qg6+ Ke7
>34.Qg7+ Ke8 35.Rb5 Bf8 36.Qg6+ Ke7 37.Rxd5 exd5 38.h4 c6 39.h5 Re8 40.h6
>Bxh6 41.Qxh6 Kd8 42.Kf2 Kc7 43.f4 b5 1-0
I did not bother to check beyond move 20. But 17. ... Qf6 deserves a ? mark,
as does 18. Ne4. 18. Bh7+ wins the queen for two minor pieces with a continuing
attack. 19. Rab1 looks like a typical computer bad play. 19. ... Ne7? blocks
the queen's access to e7 square and aggravates the effect of Bh6 shot.
20. ... gh6?? is also bad, 20. ... Bxe4! first is needed, and black can play
on.
At move 17, black may need to play Rf6 and give up the exchange for the
bishop on g6 with some compensation. Of course, 16. ... Ba3? was just horrible.
16. ... Bf6, and it was still a game. Genius 2 is always a puzzle to me.
Sometime it seems so good, and yet sometime it is just bad.
Feng-Hsiung Hsu
>quiescence. (That is, contrary to Bob Hyatt's assertion, we seem to
>measure ply the same way that Cray Blitz does.)
Neither Cray Blitz nor *Socrates was really searching using what would
traditionally be called brute force. Cray Blitz was doing (Depth-1) null
move pruning, while *Socrates was doing (Depth-2) recursive? null move
pruning. What this means is that both Cray Blitz and *Socrates could have
blind spots that won't go away with deeper searches. On top of the
blind spots, certain types of combinations would take one extra ply for
Cray Blitz, and two extra plies for *Socrates to see. *Socrates does get
about one extra ply due to the pruning.
> 18. Na2 ....
>in our game with DT, had nothing to do
>with searching down the right paths or the wrong paths. *Socrates simply
>an error in judging the position. I may be misremembering, but I think
>that DT2 suggested the same move at the same time. It was simply a case of
>the computers not knowing how to play the position. We were unlucky enough
>to actually play the bad move. This sort of thing explains why one game
>makes it difficult to judge the strenght of a program. (However, there is
>good evidence that DT2 is better than *Socrates: For example, *Socrates
>managed to lose to Zarkov from a dead drawn position. I doubt DT2 would
>have lost.)
Knights on the rim are dim. No, DT2, with its excellent positional sense,
would never have expected *Socrates to play Na2. Just kidding:-).
What really happened was that *Socrates thought it was winning the d4 pawn
by playing Na2, while DT2 knew all along that the d4 pawn was untouchable.
It is not a matter of chess knowledge. It is a matter of calculating the
dynamics of the position. It was expecting Ne2, by the way, and calculated
that even after Ne2 it had a bind in the position. Why did *Socrates miss
the complications? I don't know. We were regularly getting 20+ plies deep
PVs, so it might be the search extensions. But it could also be the null
move pruning that *Socrates was doing. Of course, I can't rule out that it
might be our much maligned evaluation function:-).
>question does not even make sense. Hsu says that DT2 consistently searched
>at least 3 million nodes per second. For the serial version of Socrates to
Not consistently. Can get below 3 million in really hairy positions, and
much much lower under unusual conditions as it actually happened in some
of the games...
Bradley C. Kuszmaul
> 18. Na2 ....
>in our game with DT, had nothing to do
>with searching down the right paths or the wrong paths. *Socrates simply
>an error in judging the position. I may be misremembering, but I think
>that DT2 suggested the same move at the same time. ...
What really happened was that *Socrates thought it was winning the d4 pawn
by playing Na2, while DT2 knew all along that the d4 pawn was untouchable.
It is not a matter of chess knowledge. It is a matter of calculating the
dynamics of the position. It was expecting Ne2, by the way, and calculated
that even after Ne2 it had a bind in the position.
Thanks for setting me straight on this.
-Bradley
Mark Uniacke
> Actually, you just proved my point about computers being poor attackers.
I believe computers execute attacks well given the attacking position where
the tactics are flowing, but are poor in building up to such positions.
I just offered the game as one unusual example where a computer was able to
build a very successful attack, accepting that black had a casual reaction to
the threats to 'his' king.
> >
> >[Event "12th World Microcomputer Chess Championship"]
> >[Site "Munich"]
> >[Date "1993"]
> >[Round "8"]
> >[White "Hiarcs"]
> >[Black "Genius 2"]
> >[Result "1-0"]
> >
> >1.d4 b6 2.e4 Bb7 3.Bd3 e6 4.Nf3 Nf6 5.e5 Ne4 6.O-O Be7 7.Re1 f5 8.ef6 Nxf6
> >9.Nc3 O-O 10.Ng5 h6 11.Nh3 Nc6 12.Be3 Nb4 13.Bg6 Nfd5 14.Bd2 Nxc3 15.bxc3
> >Nc6 16.Qh5 Ba3 17.Ng5 Qf6 18.Ne4 Qd8 19.Rab1 Ne7 20.Bxh6 gxh6 21.Qxh6 Nxg6
> >22.Qxg6+ Kh8 23.Qh6+ Kg8 24.Re3 Kf7 25.Rf3+ Ke8 26.Qg6+ Ke7 27.Qg7+ Ke8
> >28.Rxf8+ Bxf8 29.Nf6+ Qxf6 30.Qxf6 Bd5 31.a4 Bd6 32.f3 a5 33.Qg6+ Ke7
> >34.Qg7+ Ke8 35.Rb5 Bf8 36.Qg6+ Ke7 37.Rxd5 exd5 38.h4 c6 39.h5 Re8 40.h6
> >Bxh6 41.Qxh6 Kd8 42.Kf2 Kc7 43.f4 b5 1-0
> I did not bother to check beyond move 20. But 17. ... Qf6 deserves a ? mark,
> as does 18. Ne4. 18. Bh7+ wins the queen for two minor pieces with a continuing
> attack. 19. Rab1 looks like a typical computer bad play. 19. ... Ne7? blocks
> the queen's access to e7 square and aggravates the effect of Bh6 shot.
> 20. ... gh6?? is also bad, 20. ... Bxe4! first is needed, and black can play
> on.
I agree Rab1 is not fully in keeping with the attack, although it can play
a part via b5 in some variations, but the other moves are coordinated even
if 18. Bh7+ was winning earlier (I have not checked this, could you post
the analysis - thanks). 20. ... Bxe4! was the reply Hiarcs was expecting when
it played Bxh6!, though black is still in serious difficulties.
> 16. ... Bf6, and it was still a game. Genius 2 is always a puzzle to me.
> Sometime it seems so good, and yet sometime it is just bad.
Aren't all programs !?
Feng-Hsiung Hsu
>> >[Event "12th World Microcomputer Chess Championship"]
>> >[Site "Munich"]
>> >[Date "1993"]
>> >[Round "8"]
>> >[White "Hiarcs"]
>> >[Black "Genius 2"]
>> >[Result "1-0"]
>> >
>> >1.d4 b6 2.e4 Bb7 3.Bd3 e6 4.Nf3 Nf6 5.e5 Ne4 6.O-O Be7 7.Re1 f5 8.ef6 Nxf6
>> >9.Nc3 O-O 10.Ng5 h6 11.Nh3 Nc6 12.Be3 Nb4 13.Bg6 Nfd5 14.Bd2 Nxc3 15.bxc3
>> >Nc6 16.Qh5 Ba3 17.Ng5 Qf6 18.Ne4 Qd8 19.Rab1 Ne7 20.Bxh6 gxh6 21.Qxh6 Nxg6
>> >22.Qxg6+ Kh8 23.Qh6+ Kg8 24.Re3 Kf7 25.Rf3+ Ke8 26.Qg6+ Ke7 27.Qg7+ Ke8
>> >28.Rxf8+ Bxf8 29.Nf6+ Qxf6 30.Qxf6 Bd5 31.a4 Bd6 32.f3 a5 33.Qg6+ Ke7
>> >34.Qg7+ Ke8 35.Rb5 Bf8 36.Qg6+ Ke7 37.Rxd5 exd5 38.h4 c6 39.h5 Re8 40.h6
>> >Bxh6 41.Qxh6 Kd8 42.Kf2 Kc7 43.f4 b5 1-0
> I agree Rab1 is not fully in keeping with the attack, although it can play
>a part via b5 in some variations, but the other moves are coordinated even
>if 18. Bh7+ was winning earlier (I have not checked this, could you post
>the analysis - thanks).
From memory, 18. Bh7+ Kh8 19. Ne4, and now the expected best response was
Kxh7 20. Nxf6. Moving the queen leads to disasters.