Deep Blue and ply depth...
Ed Schroder
Deep Blue and ply depth...
Computer Corner :) wrote..
18 ply full width still won't be enough. Kasparov thinks over 18 ply
on every move, not full width but he knows which lines to throw out.
The selective lines that Deep Blue prunes for special consideration
to 30-35 ply are not as useful as Gary's. Kasparov can be beaten,
but more knowledge is needed by Deep Blue. Games 5 and 6 of the last
match are only a small sample of the problems. Prediction:
4 draws for the computer and 2 wins for Kasparov.
-------------------------------------------------------------------
Ed Schroder wrote...
Where can I find this PLY info:
- Brute force (18 ply)
- Selective till 30-35 ply
I don't believe a word of it before the Deep Blue team will publish
this heavy numbers themselves.
- Ed Schroder -
---------------------------------------------------------------------
Bob Hyatt wrote...
I'd suspect speeds in the 13-15 ply range, based on what I see with
Cray Blitz with null move R=1, non-recursive. Since Hsu is and has
been anti null-move, this seems reasonable.
the selective to 30-35 plies seems pessimistic from what I've personally
seen. I've been on the wrong end of such PV's from a machine much slower
than DB. I wouldn't be surprised if they are averaging 20-25 plies deep
counting their multiple extensions... What that's going to produce is a
couple of months away from being seen of course.
It *is* intimidating...
-----------------------------------------------------------------------
Let's take the famous Deep Thought II - Fritz game, Hong Komg 1995...
1. e2-e4 c7-c5
2. Pg1-f3 Pb8-c6
3. d2-d4 c5xd4
4. Pf3xd4 Pg8-f6
5. Pb1-c3 e7-e5
6. Pd4-b5 d7-d6
7. Lc1-g5 a7-a6
8. Pb5-a3 b7-b5
9. Lg5xf6 g7xf6
10. Pc3-d5 f6-f5
11. Lf1-d3 Lc8-e6
12. Dd1-h5 f5-f4
13. O-O ???
After this move white is lost, hope we all agree on that??
My point is why DT-II did not saw 13. 0-0 is losing?
The real loss is certainly *NOT* 30-35 plies deep!!
Let's continue the game...
13... Th8-g8
14. Kg1-h1 Tg8-g6
15. Dh5-d1 Ta8-c8
16. c2-c4 Dd8-h4
17. g2-g3 Dh4-h3
18. Dd1-d2 f4-f3
19. Tf1-g1 Tg6-h6
20. Dd2xh6
And white loses the queen....
Losing main line of the game just *15* plies!!
Bottem line...
- No 18 ply brute force by DT-II;
- No 30-35 ply selective search by DT-II;
- Not even 15 plies.
The loss maybe can be postponed a few plies but I do hope I made my point.
Maybe just maybe DT / DB isn't such a deep thinker as we all *assume*
I also remember DT lost from Lang (on slow hardware) 3-4 years ago.
I am not against Deep Blue, I am a big fan of them. I will enjoy the
next match against Kasparov. Hsu and IBM made it all possible. My thanks
go to both of them. But I like to keep the balance... :)
- Ed Schroder -
brucemo
Ed Schroder wrote:
> Losing main line of the game just *15* plies!!
>
> Bottem line...
> - No 18 ply brute force by DT-II;
> - No 30-35 ply selective search by DT-II;
> - Not even 15 plies.
>
> The loss maybe can be postponed a few plies but I do hope I made my point.
>
> Maybe just maybe DT / DB isn't such a deep thinker as we all *assume*
I was looking over their shoulders and I saw them reaching depths that
looked to be similar to what everyone else was doing.
Maybe there's four plies of brute-force in every tip node though, I don't
know.
But that wasn't Deep Blue. Who knows what that was.
In any case, speculating gets us nowhere. I'd rather read a few paragraphs
of clarification from the DB team than reams of speculation from everyone
else.
bruce
Ed Schroder
Computer Corner :) wrote..
-------------------------------------------------------------------
Ed Schroder wrote...
- Ed Schroder -
---------------------------------------------------------------------
Bob Hyatt wrote...
It *is* intimidating...
-----------------------------------------------------------------------
1. e2-e4 c7-c5
2. Ng1-f3 Nb8-c6
3. d2-d4 c5xd4
4. Nf3xd4 Ng8-f6
5. Nb1-c3 e7-e5
6. Nd4-b5 d7-d6
7. Bc1-g5 a7-a6
8. Nb5-a3 b7-b5
9. Bg5xf6 g7xf6
10. Nc3-d5 f6-f5
11. Bf1-d3 Bc8-e6
12. Qd1-h5 f5-f4
13. O-O ???
After this move white is lost, hope we all agree on that??
My point is why DT-II did not saw 13. 0-0 is losing?
The real loss is certainly *NOT* 30-35 plies deep!!
Let's continue the game...
13... Rh8-g8
14. Kg1-h1 Rg8-g6
15. Qh5-d1 Ra8-c8
16. c2-c4 Qd8-h4
17. g2-g3 Qh4-h3
18. Qd1-d2 f4-f3
19. Rf1-g1 Rg6-h6
20. Qd2xh6
And white loses the queen....
Losing main line of the game just *15* plies!!
Bottem line...
- No 18 ply brute force by DT-II;
- No 30-35 ply selective search by DT-II;
- Not even 15 plies.
The loss maybe can be postponed a few plies but I do hope I made my point.
Maybe just maybe DT / DB isn't such a deep thinker as we all *assume*
I also remember DT lost from Lang (on slow hardware) 3-4 years ago.
Robert Hyatt
Ed Schroder (rebc...@xs4all.nl) wrote:
: Deep Blue and ply depth...
: Computer Corner :) wrote..
: -------------------------------------------------------------------
: Ed Schroder wrote...
: - Ed Schroder -
: ---------------------------------------------------------------------
: Bob Hyatt wrote...
: It *is* intimidating...
: -----------------------------------------------------------------------
: 1. e2-e4 c7-c5
: 2. Pg1-f3 Pb8-c6
: 3. d2-d4 c5xd4
: 4. Pf3xd4 Pg8-f6
: 5. Pb1-c3 e7-e5
: 6. Pd4-b5 d7-d6
: 7. Lc1-g5 a7-a6
: 8. Pb5-a3 b7-b5
: 9. Lg5xf6 g7xf6
: 10. Pc3-d5 f6-f5
: 11. Lf1-d3 Lc8-e6
: 12. Dd1-h5 f5-f4
: 13. O-O ???
: After this move white is lost, hope we all agree on that??
: My point is why DT-II did not saw 13. 0-0 is losing?
: The real loss is certainly *NOT* 30-35 plies deep!!
First, you aren't talking about deep blue, but deep thought. Second, the
loss is *very* deep. If you follow the line, the queen loss is not exactly
forced at that point, but the search finds out that it's in deep trouble
and tosses the queen as one way out. I believe that they did see that they
were in trouble one move later.
: Let's continue the game...
: 13... Th8-g8
: 14. Kg1-h1 Tg8-g6
: 15. Dh5-d1 Ta8-c8
: 16. c2-c4 Dd8-h4
: 17. g2-g3 Dh4-h3
: 18. Dd1-d2 f4-f3
: 19. Tf1-g1 Tg6-h6
: 20. Dd2xh6
: And white loses the queen....
: Losing main line of the game just *15* plies!!
: Bottem line...
: - No 18 ply brute force by DT-II;
: - No 30-35 ply selective search by DT-II;
: - Not even 15 plies.
: The loss maybe can be postponed a few plies but I do hope I made my point.
: Maybe just maybe DT / DB isn't such a deep thinker as we all *assume*
I'm not assuming. I've *seen*. *personally*. Hsu doesn't hide his
terminal when he's playing.
: I also remember DT lost from Lang (on slow hardware) 3-4 years ago.
: I am not against Deep Blue, I am a big fan of them. I will enjoy the
: next match against Kasparov. Hsu and IBM made it all possible. My thanks
: go to both of them. But I like to keep the balance... :)
: - Ed Schroder -
Me too. Here's one fun diversion however. You list all the games that Deep Thought
has lost, and I'll list all the games that deep thought has
won, and let's compare lists. I'll bet you mine is 4x longer than yours...
Moritz Berger
On 6 Feb 1997 23:31:45 GMT, Ed Schroder <rebc...@xs4all.nl> wrote:
>Deep Blue and ply depth...
<snip>
>Let's take the famous Deep Thought II - Fritz game, Hong Komg 1995...
>
> 1. e2-e4 c7-c5
< snip >
> 13. O-O ???
>
>After this move white is lost, hope we all agree on that??
>My point is why DT-II did not saw 13. 0-0 is losing?
>The real loss is certainly *NOT* 30-35 plies deep!!
As far as I got it, DT played O-O out of book, without any
calculations ...
Moritz
-------------
Moritz...@msn.com
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
Subject: Re: Deep Blue and ply depth...
: Computer Corner :) wrote..
: -------------------------------------------------------------------
: Ed Schroder wrote...
: - Ed Schroder -
: ---------------------------------------------------------------------
: Bob Hyatt wrote...
: It *is* intimidating...
: -----------------------------------------------------------------------
: Let's take the famous Deep Thought II - Fritz game, Hong Komg 1995...
: 1. e2-e4 c7-c5
: 2. Pg1-f3 Pb8-c6
: 3. d2-d4 c5xd4
: 4. Pf3xd4 Pg8-f6
: 5. Pb1-c3 e7-e5
: 6. Pd4-b5 d7-d6
: 7. Lc1-g5 a7-a6
: 8. Pb5-a3 b7-b5
: 9. Lg5xf6 g7xf6
: 10. Pc3-d5 f6-f5
: 11. Lf1-d3 Lc8-e6
: 12. Dd1-h5 f5-f4
: 13. O-O ???
: After this move white is lost, hope we all agree on that??
: My point is why DT-II did not saw 13. 0-0 is losing?
: The real loss is certainly *NOT* 30-35 plies deep!!
> First, you aren't talking about deep blue, but deep thought. Second,
> the loss is *very* deep. If you follow the line, the queen loss is not
> exactly forced at that point, but the search finds out that it's in
> deep trouble and tosses the queen as one way out. I believe that
> they did see that they were in trouble one move later.
Ok I will make it easier...
On ply=9 Rebel8 already find the loss on 16.c4, score -1.xx enough to
cancel the move 13.0-0?? if Rebel8 was able to search on 9+6=15 ply.
To me 13. 0-0 needs (only) 15 ply to find the loss on 13.0-0
Deep Thought-II did not see it.
Will Deep Blue see it?
For the moment I like to stick to my opinion that DT/DB is not the deep
searcher I am supposed to believe if I read the majority of the postings
here in RGCC... :)
: Let's continue the game...
: 13... Th8-g8
: 14. Kg1-h1 Tg8-g6
: 15. Dh5-d1 Ta8-c8
: 16. c2-c4 Dd8-h4
: 17. g2-g3 Dh4-h3
: 18. Dd1-d2 f4-f3
: 19. Tf1-g1 Tg6-h6
: 20. Dd2xh6
: And white loses the queen....
: Losing main line of the game just *15* plies!!
: Bottem line...
: - No 18 ply brute force by DT-II;
: - No 30-35 ply selective search by DT-II;
: - Not even 15 plies.
: The loss maybe can be postponed a few plies but I do hope I made my
: point.
: Maybe just maybe DT / DB isn't such a deep thinker as we all *assume*
> I'm not assuming. I've *seen*. *personally*. Hsu doesn't hide his
> terminal when he's playing.
But what have you seen then?
I am interested in:
- Brute force depth
- Selective depth
A PV of 30 plies will not convince me. Look at the old ChessMachine
program of Johan de Koning which is CM4000/5000 now. I have seen solving
the ChessMachine a mate in 12 position, without ANY check, without any
recapture and all usual stuff. Just 24 silent (half) moves! All in an
amazing time of 20-30 seconds or so.
: I also remember DT lost from Lang (on slow hardware) 3-4 years ago.
: I am not against Deep Blue, I am a big fan of them. I will enjoy the
: next match against Kasparov. Hsu and IBM made it all possible. My thanks
: go to both of them. But I like to keep the balance... :)
: - Ed Schroder -
> Me too. Here's one fun diversion however. You list all the games
> that Deep Thought has lost, and I'll list all the games that deep
> thought has won, and let's compare lists. I'll bet you mine is 4x
> longer than yours...
Yes, but your list is pretty out of date and in best case filled with
486 and P90 machines... :)
Now we have:
- P200 MMX and PP-200 machimes
- And better chess programs.
Time flies...
- Ed Schroder -
Ed Schroder
From: brucemo <bru...@nwlink.com>
Subject: Re: Deep Blue and ply depth...
> Losing main line of the game just *15* plies!!
> Bottem line...
> - No 18 ply brute force by DT-II;
> - No 30-35 ply selective search by DT-II;
> - Not even 15 plies.
> The loss maybe can be postponed a few plies but I do hope I made my > point.
> Maybe just maybe DT / DB isn't such a deep thinker as we all *assume*
: I was looking over their shoulders and I saw them reaching depths that
: looked to be similar to what everyone else was doing.
: Maybe there's four plies of brute-force in every tip node though, I
: don't know.
: But that wasn't Deep Blue. Who knows what that was.
: In any case, speculating gets us nowhere. I'd rather read a few
: paragraphs of clarification from the DB team than reams of speculation
: from everyone else.
You are right, a clarification could stop this whole discussion we are
having again and again, over and over.
I just jumped in because when I read the majority of the postings here I
get the feeling that DT/DB will crush any other chess program with 10-0
or so and that DT/DB is unbeatable because of that giant ply depth.
With the game DT-Fritz Hong Kong 1995 I state that DT-II could have seen
13. 0-0??? in (just) 15 plies.
- So no 18 ply brute force;
- No selective search till 30-35 plies;
as I read here, that's bullshit or wishful thinking.
About speculations...
I did not start the speculations, others did, I just jumped in because
I have another opinion.
Hsu, please help us out...
- Ed Schroder -
: bruce
Kevin James Begley
: >Let's take the famous Deep Thought II - Fritz game, Hong Komg 1995...
: >
: > 1. e2-e4 c7-c5
: > 13. O-O ???
: >
: >After this move white is lost, hope we all agree on that??
: >My point is why DT-II did not saw 13. 0-0 is losing?
: >The real loss is certainly *NOT* 30-35 plies deep!!
: As far as I got it, DT played O-O out of book, without any
: calculations ...
No, it played O-O only because it had been thrown "out-of-book"
by a move order trick. Had it been in book, it would not
have played this move (apparently, it's ok to have a killer-
book which defends against something, but taboo to have one
that threatnes something).
Still, the original poster makes an excellent point --
apparently, DT does not posses the ability to calculate
(and avoid) this blunder.
Sorry, Bob, but you'll have to do better than threatening
to trade list of DT games if you are to make the case that
DT is as powerful as you claim. If it can really see a way
to win a piece through 45 non-forced moves, why in hell
can't it see that it's queen will fall in about 15 forcing
moves?
Perhaps in certain positions, it can see extremely far
ahead (there's no reason not to trust Bob's account of
what this machine is capable of), but for some reason,
was not able to do so in the game mentioned.
Thus, it has POTENTIAL for being extremely powerful,
as Bob suggests, but it certainly is not above a cheap
swindle from a PC program (hey, that's nothing to be
ashamed of -- even the World Champion can be snookered
on occasion).
So, the question on my mind is, why is everything so
hush-hush w/ the DB team? Are they worried that GK
would prepare a line against them using the data
they provide? Is this why they refuse to exhibit
their machine? If so, what do the rgcc folks think
about this policy -- paranoia, or smart strategy?
Kevin.
Robert Hyatt
Kevin James Begley (kjbe...@chimi.engr.ucdavis.edu) wrote:
I didn't mean to imply it can see every 40 move variation, that would take
a 40 ply search, period. However, I did, and still do, claim that it can
see forcing variations that no one else can see. As far as this position,
it would be easy to set it up and see if Crafty can spot it in 15 plies, that
is well within a reasonable search depth... might take a few hours or a day
or two, but it should reach that depth. And I'll bet it won't see it. Because
there are likely lots of subtle defenses that delay things further than you'd
suspect...
: Perhaps in certain positions, it can see extremely far
: ahead (there's no reason not to trust Bob's account of
: what this machine is capable of), but for some reason,
: was not able to do so in the game mentioned.
There will be *plenty* of such positions where it is not all-seeing. But
the fact remains, the "best of the rest" see so much less...
: Thus, it has POTENTIAL for being extremely powerful,
: as Bob suggests, but it certainly is not above a cheap
: swindle from a PC program (hey, that's nothing to be
: ashamed of -- even the World Champion can be snookered
: on occasion).
Absolutely not. That's why one game is just one game. I'd bet *nothing* on
a single game between any two players, but for a long match, I'd take DB for
any amount of money... the longer the match the better...
: So, the question on my mind is, why is everything so
: hush-hush w/ the DB team? Are they worried that GK
: would prepare a line against them using the data
: they provide? Is this why they refuse to exhibit
: their machine? If so, what do the rgcc folks think
: about this policy -- paranoia, or smart strategy?
: Kevin.
Dunno the answer to that...
Peter W. Gillgasch
Kevin James Begley <kjbe...@chimi.engr.ucdavis.edu> wrote:
> Moritz Berger (Moritz...@msn.com) wrote:
> : >Deep Blue and ply depth...
[ snip ]
> : >After this move white is lost, hope we all agree on that??
> : >My point is why DT-II did not saw 13. 0-0 is losing?
> : >The real loss is certainly *NOT* 30-35 plies deep!!
>
> : As far as I got it, DT played O-O out of book, without any
> : calculations ...
Nope.
> No, it played O-O only because it had been thrown "out-of-book"
> by a move order trick. Had it been in book, it would not
> have played this move (apparently, it's ok to have a killer-
> book which defends against something, but taboo to have one
> that threatnes something).
>
> Still, the original poster makes an excellent point --
> apparently, DT does not posses the ability to calculate
> (and avoid) this blunder.
AFAIK... FHH said that DBP did avoid the blunder basically on the last
second but the connection to Yorktown Heights broke down. They
re-connected but needed to restarted DBP and added a "time penalty" for
the time-out. Now DBP didn't have the time to see it again.
Of course this issue is a mood point and one could argue about whether
FHH is telling the truth [he claimed that this was traced down by
looking at the log file] but DBP *was* seeing the trouble coming along
*at least* after Black's next move, if not earlies. The guys and me were
present at the event with DarkThought and I recall that people were
running around stating that DBP is dead meat. We checked the situation
and the machine was very unhappy, so I believe that FHH is not putting
up any crap.
Bad luck for them. Shit happens. Anyway, they didn't put the gun at
their head. In a five round event anything can happen, especially if the
competition is so fierce.
> Sorry, Bob, but you'll have to do better than threatening
> to trade list of DT games if you are to make the case that
> DT is as powerful as you claim. If it can really see a way
> to win a piece through 45 non-forced moves,
Flaw in your argumentation here. Bob's definition of "forced move" is
significantly different of FHH's definition. FHH's definition doesn't
imply that the move is a check or a capture, they extend on search
results recursively inside the tree, which is a very costly thing to do
if you can't throw a couple of millions nodes per second around without
worrying too much...
> why in hell can't it see that it's queen will fall in about 15 forcing
> moves?
See above. First it probably can [mood issue], second "forcing moves"
and "forcing moves" are not neccessarily the same. It depends on the
"amount of forcedness", if the 2 best moves anywhere on the line are
pretty "close" then this node will not be extended, so it is very much
depending on the current position you are talking about. Obviously Bob
and CB were left hanging in a position where the beast did extend and
extend and extend...
If you ever play with singular extensions you can see this phenomenon
quite often. Sometimes the results are *spectacular*, like a quick
fraction of a second 6 plies search resulting in a 32 plies principle
variation, sometimes you see only a slowdown and an 8 ply pv... If you
end up in such a position and the position is bad for your machine then
you can only hope that you sit on the right side of the board or that
#define SINGULARITY_MARGIN (PAWN_VALUE / 2)
is off by 1/128 of a pawn :)
> Perhaps in certain positions, it can see extremely far
> ahead (there's no reason not to trust Bob's account of
> what this machine is capable of), but for some reason,
> was not able to do so in the game mentioned.
Well, it didn't pull it off in OTB play, true.
> Thus, it has POTENTIAL for being extremely powerful,
> as Bob suggests, but it certainly is not above a cheap
> swindle from a PC program (hey, that's nothing to be
> ashamed of -- even the World Champion can be snookered
> on occasion).
Right. Even 1,000 elo points advantage don't keep you safe from falling
into a swindle occasionally. The real swindle is to take one game, or
one incarnation of FHH's toys that was primarily a "transition version"
between a major hardware change and to draw any conclisions from that.
If you see their record and you look at the stuff they made public and
you use a pocket calculator then it can make you worried :)
> So, the question on my mind is, why is everything so
> hush-hush w/ the DB team? Are they worried that GK
> would prepare a line against them using the data
> they provide?
My guess is as good as yours, probably yes.
> Is this why they refuse to exhibit
> their machine? If so, what do the rgcc folks think
> about this policy -- paranoia, or smart strategy?
On an unrelated subject I recall Murray saying "we do what is
neccessary" and Hans said "you do what wins!"
Not a nice strategy, but probably not a dumb one.
-- Peter
Ed Schroder
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
Subject: Re: Deep Blue and ply depth...
: Computer Corner :) wrote..
: -------------------------------------------------------------------
: Ed Schroder wrote...
: - Ed Schroder -
: ---------------------------------------------------------------------
: Bob Hyatt wrote...
: It *is* intimidating...
: -----------------------------------------------------------------------
: Let's take the famous Deep Thought II - Fritz game, Hong Komg 1995...
: 1. e2-e4 c7-c5
: 2. Pg1-f3 Pb8-c6
: 3. d2-d4 c5xd4
: 4. Pf3xd4 Pg8-f6
: 5. Pb1-c3 e7-e5
: 6. Pd4-b5 d7-d6
: 7. Lc1-g5 a7-a6
: 8. Pb5-a3 b7-b5
: 9. Lg5xf6 g7xf6
: 10. Pc3-d5 f6-f5
: 11. Lf1-d3 Lc8-e6
: 12. Dd1-h5 f5-f4
: 13. O-O ???
: After this move white is lost, hope we all agree on that??
: My point is why DT-II did not saw 13. 0-0 is losing?
: The real loss is certainly *NOT* 30-35 plies deep!!
> First, you aren't talking about deep blue, but deep thought. Second,
> the loss is *very* deep. If you follow the line, the queen loss is not
> exactly forced at that point, but the search finds out that it's in
> deep trouble and tosses the queen as one way out. I believe that
> they did see that they were in trouble one move later.
Ok I will make it easier...
On ply=9 Rebel8 already find the loss on 16.c4, score -1.xx enough to
cancel the move 13.0-0?? if Rebel8 was able to search on 9+6=15 ply.
To me 13. 0-0 needs (only) 15 ply to find the loss on 13.0-0
Deep Thought-II did not see it.
Will Deep Blue see it?
For the moment I like to stick to my opinion that DT/DB is not the deep
thinker I am supposed to believe if I read the majority of the postings
here in RGCC... :)
: Let's continue the game...
: 13... Th8-g8
: 14. Kg1-h1 Tg8-g6
: 15. Dh5-d1 Ta8-c8
: 16. c2-c4 Dd8-h4
: 17. g2-g3 Dh4-h3
: 18. Dd1-d2 f4-f3
: 19. Tf1-g1 Tg6-h6
: 20. Dd2xh6
: And white loses the queen....
: Losing main line of the game just *15* plies!!
: Bottem line...
: - No 18 ply brute force by DT-II;
: - No 30-35 ply selective search by DT-II;
: - Not even 15 plies.
: The loss maybe can be postponed a few plies but I do hope I made my
: point.
: Maybe just maybe DT / DB isn't such a deep thinker as we all *assume*
> I'm not assuming. I've *seen*. *personally*. Hsu doesn't hide his
> terminal when he's playing.
But what have you seen then?
I am interested in:
- Brute force depth
- Selective depth
A PV of 30 plies will not convince me. Look at the old ChessMachine
program of Johan de Koning which is CM4000/5000 now. I have seen solving
the ChessMachine a mate in 12 position, without ANY check, without any
recapture and all usual stuff. Just 24 silent (half) moves! All in an
amazing time of 20-30 seconds or so.
: I also remember DT lost from Lang (on slow hardware) 3-4 years ago.
: I am not against Deep Blue, I am a big fan of them. I will enjoy the
: next match against Kasparov. Hsu and IBM made it all possible. My thanks
: go to both of them. But I like to keep the balance... :)
: - Ed Schroder -
> Me too. Here's one fun diversion however. You list all the games
> that Deep Thought has lost, and I'll list all the games that deep
> thought has won, and let's compare lists. I'll bet you mine is 4x
> longer than yours...
Yes but your list is pretty out of date and in best case filled with
brucemo
Ed Schroder wrote:
> A PV of 30 plies will not convince me. Look at the old ChessMachine
> program of Johan de Koning which is CM4000/5000 now. I have seen solving
> the ChessMachine a mate in 12 position, without ANY check, without any
> recapture and all usual stuff. Just 24 silent (half) moves! All in an
> amazing time of 20-30 seconds or so.
Please post this :-)
bruce
Robert Hyatt
Ed Schroder (rebc...@xs4all.nl) wrote:
: From: brucemo <bru...@nwlink.com>
: Subject: Re: Deep Blue and ply depth...
: > Losing main line of the game just *15* plies!!
:
: > Bottem line...
: > - No 18 ply brute force by DT-II;
: > - No 30-35 ply selective search by DT-II;
: > - Not even 15 plies.
:
: > The loss maybe can be postponed a few plies but I do hope I made my > point.
: > Maybe just maybe DT / DB isn't such a deep thinker as we all *assume*
: : I was looking over their shoulders and I saw them reaching depths that
: : looked to be similar to what everyone else was doing.
: : Maybe there's four plies of brute-force in every tip node though, I
: : don't know.
: : But that wasn't Deep Blue. Who knows what that was.
: : In any case, speculating gets us nowhere. I'd rather read a few
: : paragraphs of clarification from the DB team than reams of speculation
: : from everyone else.
: You are right, a clarification could stop this whole discussion we are
: having again and again, over and over.
: I just jumped in because when I read the majority of the postings here I
: get the feeling that DT/DB will crush any other chess program with 10-0
: or so and that DT/DB is unbeatable because of that giant ply depth.
: With the game DT-Fritz Hong Kong 1995 I state that DT-II could have seen
: 13. 0-0??? in (just) 15 plies.
: - So no 18 ply brute force;
: - No selective search till 30-35 plies;
: as I read here, that's bullshit or wishful thinking.
: About speculations...
: I did not start the speculations, others did, I just jumped in because
: I have another opinion.
: Hsu, please help us out...
: - Ed Schroder -
: : bruce
someone pointed out, and I remembered after they did, that Hsu pointed out
that DT had found that o-o lost. However they lost communication and when
they restarted, it didn't have enough time. Would be interesting to see how long it
takes any other program to find out that o-o is *bad*.. not how long to not play
it, but how long to find out that it is really a bad move. I'd bet not in any
sort of tournament time. And remember, that was DT, *not* DB that found it..
DB's only 100x+ faster
Ed Schroder
From: brucemo <bru...@nwlink.com>
Subject: Re: Deep Blue and ply depth...
: Ed Schroder wrote:
: Please post this :-)
: bruce
Your wish is my command... :)
8/2Qp4/8/8/8/1r6/kp1p4/qrbK4 w - -
8/8/8/8/8/8/ppQKPPP1/k7 w - -
or in PGN...
------------------------------------------------------------------
[Event "M12"]
[Site ""]
[Date "1994.05.09"]
[Round ""]
[White ""]
[Black ""]
[Result "1-0"]
[BlackElo ""]
[WhiteElo ""]
[FEN "8/8/8/8/8/8/ppQKPPP1/k7 w - - 0 1"]
1. Qc3 Kb1 2. Qd3+ Ka1 3. Qd4 Kb1 4. Qe4+ Ka1 5. Qe5 Kb1 6. Qf5+ Ka1 7.
Qf6 Kb1 8. Qg6+ Ka1 9. Qg7 Kb1 10. Qh7+ Ka1 11. Qh8 Kb1 12. Qh1# 1-0
---------------------------------------------------------------------
[Event "M12"]
[Site ""]
[Date "1994.02.06"]
[Round ""]
[White ""]
[Black ""]
[Result "1-0"]
[BlackElo ""]
[WhiteElo ""]
[FEN "8/2Qp4/8/8/8/1r6/kp1p4/qrbK4 w - - 0 1"]
1. Qa5+ Ra3 2. Qd5+ Rb3 3. Qa8+ Ra3 4. Qg8+ Rb3 5. Qd5 d6 6. Qa8+ Ra3 7.
Qg8+ d5 8. Qxd5+ Rb3 9. Qa8+ Ra3 10. Qg8+ Rb3 11. Qd5 Ka3 12. Qa5# 1-0
------------------------------------------------------------------------
I was wrong about "no checks", there are many, but what impresses me is
the number of "quiet moves" in position-1 and especially the "quiet
move" 11.Qd5 in position-2.
I haven't checked the 2 positions with CM4000/5000, all I rememeber is
that the ChessMachine (King) solved both positions very fast.
Very impressive and I am curious if any other program will find them and
of course I am curious if CM4000/5000 still solves them and how fast.
- Ed Schroder -
Moritz Berger
On 8 Feb 1997 07:24:12 GMT, Ed Schroder <rebc...@xs4all.nl> wrote:
>From: hy...@crafty.cis.uab.edu (Robert Hyatt)
>
>Subject: Re: Deep Blue and ply depth...
>
>
>: 18 ply full width still won't be enough. Kasparov thinks over 18 ply
>: on every move, not full width but he knows which lines to throw out.
>: The selective lines that Deep Blue prunes for special consideration
>: to 30-35 ply are not as useful as Gary's. Kasparov can be beaten,
>: but more knowledge is needed by Deep Blue. Games 5 and 6 of the last
>: match are only a small sample of the problems. Prediction:
>: 4 draws for the computer and 2 wins for Kasparov.
>: Bob Hyatt wrote...
>: Let's take the famous Deep Thought II - Fritz game, Hong Komg 1995...
>: 1. e2-e4 c7-c5
>: 13. O-O ???
>: After this move white is lost, hope we all agree on that??
>: My point is why DT-II did not saw 13. 0-0 is losing?
>: The real loss is certainly *NOT* 30-35 plies deep!!
>> First, you aren't talking about deep blue, but deep thought. Second,
>> the loss is *very* deep. If you follow the line, the queen loss is not
>> exactly forced at that point, but the search finds out that it's in
>> deep trouble and tosses the queen as one way out. I believe that
>> they did see that they were in trouble one move later.
>
>Ok I will make it easier...
>On ply=9 Rebel8 already find the loss on 16.c4, score -1.xx enough to
>cancel the move 13.0-0?? if Rebel8 was able to search on 9+6=15 ply.
>: I also remember DT lost from Lang (on slow hardware) 3-4 years ago.
>: I am not against Deep Blue, I am a big fan of them. I will enjoy the
>: next match against Kasparov. Hsu and IBM made it all possible. My thanks
>: go to both of them. But I like to keep the balance... :)
>
>: - Ed Schroder -
Maybe what I tried is not very scientific, but it certainly seemed
interesting to try: Since Fritz 3.1 (Honkong version) is available as
an engine for Fritz 4, I simply let Genius 5 take the "Deep Genius"
part and play on after 13. 0-0 to see if it was indeed losing
_against_Fritz_ (this is the big caveat, of course!).
Surprise, surprise, Genius seems to be able to defend very well in
this position. Maybe Deep Thought saw some attacks by Fritz that led
it to play what it believed the best defense under this circumstances,
but certainly Fritz has much more trouble fighting against Genius 5
than against Deep Though 2, even after the castling!
Here's a rough attempt at analyzing the position by letting the
participants fight it out at tournament conditions (40 moves in 2h).
What do you think? I'm waiting for your comments, but the verdict that
white is lost after castling certainly needs some more arguments in
its favour ... Although I'm neither a GM nor a Komputer ;-)
BTW: Analysis was done using ChessBase 6, so this PGN file with nested
comments might look best in CB 6, CB Demo or Fritz 4 (aka Extreme) ...
I also tried Fritz vs. CM 5000, CM 5000 didn't like its position after
some moves (evaluation < -3.5), so I didn't continue this game ...
I hope that this hasn't been analyzed to death before, but independent
from the possible best defense, I still can claim that the attack
found by Fritz was not good enough to beat Genius 5 while it crunched
the big monster DT II ...
Moritz
[Event "?"]
[Site "Moritz Berger,Bonn"]
[Date "1995.??.??"]
[Round "?"]
[White "Genius 5 Windows, P5/166 32MB hash"]
[Black "Fritz3.1 Hongkong Version (F4), P5/133 32MB hash"]
[Result "*"]
[ECO "B33"]
[PlyCount "58"]
1. e4 c5 2. Nf3 Nc6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 e5 6. Ndb5 d6 7. Bg5
a6 8.
Na3 b5 9. Bxf6 gxf6 10. Nd5 f5 11. Bd3 Be6 12. Qh5 f4 13. O-O Rg8 14.
c3 ({
Chessmaster 5000 P5/166 32MB hash:} 14. Qxh7 Rg6 15. Qh5 Rh6 $19) (14.
Kh1 Rg6
15. Qd1 Rc8 16. c4 Qh4 17. g3 Qh3 18. Qd2 (18. Rg1 Qxh2+ 19. Kxh2 Rh6+
20. Qh5
Rxh5+ 21. Kg2 Bh3+ 22. Kf3 Nd4#) 18... f3 19. Rg1 Rh6 20. Qxh6 Qxh6
21. cxb5
Bxd5 22. exd5 Nb4 23. Bf5 Rc5 24. bxa6 Nxa6 25. Nc2 Qd2 26. Ne1 Rxd5
27. Nxf3
Qxf2 28. Be4 Ra5 29. Rg2 Qe3 30. Re1 Qh6 31. Bc6+ Kd8 32. a3 f5 33.
Rc2 Rc5 34.
Rxc5 Nxc5 35. Rf1 Be7 36. a4 f4 37. gxf4 Qxf4 38. Rg1 Nxa4 39. b4 Qxb4
{
0-1 Deep Blue-Fritz3/ 1995}) 14... h6 (14... Rg6 {(move entered by
operator
since Rg6 is outside the search window on P5/133 without taking
permanent
brain prolongated pondering time into account)} 15. Be2 Rc8 16. Rfd1
Rh6 17.
Qf3 Qg5 18. g3 Bg7 19. Nc2 f5 20. Nb6 fxe4 21. Qxe4 Rc7 22. Ne1 Bf5
23. Qd5 Qd8
24. a4 Ne7 25. Qf3 Qb8 26. a5 Rg6 27. Qh5 Kf8 28. Ng2 Rh6 29. Qf3 fxg3
30. hxg3
Rf6 31. Rd2 $13) 15. Be2 Rg5 16. Qf3 Rc8 17. Nc2 f5 18. exf5 Bxf5 19.
Bd3 Bg4
20. Qe4 Kf7 21. Nce3 Bh5 22. h4 Bg6 23. Qf3 Bxd3 24. hxg5 Qxg5 25.
Rfe1 Qg6 26.
Nb6 Be4 27. Qg4 Rb8 28. Qxg6+ Bxg6 29. Ned5 Be7 $14 *
-------------
Moritz...@msn.com
Kylm{l{ Jari
Ed Schroder (rebc...@xs4all.nl) wrote:
: > A PV of 30 plies will not convince me. Look at the old ChessMachine
: > program of Johan de Koning which is CM4000/5000 now. I have seen solving
: > the ChessMachine a mate in 12 position, without ANY check, without any
: > recapture and all usual stuff. Just 24 silent (half) moves! All in an
: > amazing time of 20-30 seconds or so.
: : Please post this :-)
: : bruce
: [FEN "8/8/8/8/8/8/ppQKPPP1/k7 w - - 0 1"]
: 1. Qc3 Kb1 2. Qd3+ Ka1 3. Qd4 Kb1 4. Qe4+ Ka1 5. Qe5 Kb1 6. Qf5+ Ka1 7.
: Qf6 Kb1 8. Qg6+ Ka1 9. Qg7 Kb1 10. Qh7+ Ka1 11. Qh8 Kb1 12. Qh1# 1-0
: [FEN "8/2Qp4/8/8/8/1r6/kp1p4/qrbK4 w - - 0 1"]
: 1. Qa5+ Ra3 2. Qd5+ Rb3 3. Qa8+ Ra3 4. Qg8+ Rb3 5. Qd5 d6 6. Qa8+ Ra3 7.
: Qg8+ d5 8. Qxd5+ Rb3 9. Qa8+ Ra3 10. Qg8+ Rb3 11. Qd5 Ka3 12. Qa5# 1-0
: ------------------------------------------------------------------------
: I was wrong about "no checks", there are many, but what impresses me is
: the number of "quiet moves" in position-1 and especially the "quiet
: move" 11.Qd5 in position-2.
: I haven't checked the 2 positions with CM4000/5000, all I rememeber is
: that the ChessMachine (King) solved both positions very fast.
: Very impressive and I am curious if any other program will find them and
: of course I am curious if CM4000/5000 still solves them and how fast.
: - Ed Schroder -
I tried these positions with CM5000 using P5 133MHz. I used different
selective search depths to see CM's ability to select "quiet moves".
Here are the results:
Pos 1: Pos 2:
selective selective
search time used search time used
0 47 secs 0 8 secs
1 37 secs 1 7 secs
2 29 secs 2 9 secs
3 22 secs 3 15 secs
. .
6 21 secs 6 26 secs
. .
10 5 secs 10 2 mins 28 secs
Seems that CM5000 can successfully select the quiet moves in Pos1
(1.Qc3) but losts the track in Pos2(1.Qa5+).
--
Jari Kylmälä, gur...@cc.tut.fi, Tampere University of Technology
mclane
Moritz...@msn.com (Moritz Berger) wrote:
>On 6 Feb 1997 23:31:45 GMT, Ed Schroder <rebc...@xs4all.nl> wrote:
>>Deep Blue and ply depth...
><snip>
>>Let's take the famous Deep Thought II - Fritz game, Hong Komg 1995...
>>
>> 1. e2-e4 c7-c5
>< snip >
>> 13. O-O ???
>>
>>After this move white is lost, hope we all agree on that??
>>My point is why DT-II did not saw 13. 0-0 is losing?
>>The real loss is certainly *NOT* 30-35 plies deep!!
>As far as I got it, DT played O-O out of book, without any
>calculations ...
>Moritz
>-------------
>Moritz...@msn.com
Of course this is another example of artificial intelligence.
Not to compute but to play a loser-move.
Databases = quantity.
mclane
Moritz...@msn.com (Moritz Berger) wrote:
>Maybe what I tried is not very scientific, but it certainly seemed
>interesting to try: Since Fritz 3.1 (Honkong version) is available as
>an engine for Fritz 4, I simply let Genius 5 take the "Deep Genius"
>part and play on after 13. 0-0 to see if it was indeed losing
>_against_Fritz_ (this is the big caveat, of course!).
>Surprise, surprise, Genius seems to be able to defend very well in
>this position. Maybe Deep Thought saw some attacks by Fritz that led
>it to play what it believed the best defense under this circumstances,
>but certainly Fritz has much more trouble fighting against Genius 5
>than against Deep Though 2, even after the castling!
Right. Sometimes it is better NOT to see anything. If you always
expect the opponent to make the best moves, you would resign if the
opponent can mate you in 34 plies.
But: will the opponent see this ?!
Francesco Di Tolla
Ed Schroder wrote:
[... big snip...]
> Now we have:
> - P200 MMX and PP-200 machimes
> - And better chess programs.
I guess this is the real fact. The gap between
mainframe/parallel/anything-
your-wallet-cant-buy and a PC is getting narrower.
Well for a while we will have only minor improovements (MMX gives only
10% more due to extra cache and this doesen't help much a chess program)
for a big step we really need next tecnology (P7 which should
double or triple present clocks), but we have to wait until next
century.
In the meanwhile what about Crafty on a SGI Origin 2000 R10000 at 200
MHz
(faster than an Alpha 500 MHz on a bench)?
regards
Franz
ShaktiFire
:::Well for a while we will have only minor improvements (MMX gives only
:::10% more due to extra cache and this doesen't help much a chess
program)
:::double or triple present clocks), but we have to wait until next
:::century.
Well there will be some improvements in speed before P7 comes out.
April 1 is due the PP233, July 1 PP266, and late 97 the PP300.
Further improvements likely in 1998.
Ronald de Man
Ed Schroder <rebc...@xs4all.nl> writes:
>From: brucemo <bru...@nwlink.com>
>Subject: Re: Deep Blue and ply depth...
>: Ed Schroder wrote:
>> A PV of 30 plies will not convince me. Look at the old ChessMachine
>> program of Johan de Koning which is CM4000/5000 now. I have seen solving
>> the ChessMachine a mate in 12 position, without ANY check, without any
>> recapture and all usual stuff. Just 24 silent (half) moves! All in an
>> amazing time of 20-30 seconds or so.
>: Please post this :-)
>: bruce
>Your wish is my command... :)
>8/2Qp4/8/8/8/1r6/kp1p4/qrbK4 w - -
>8/8/8/8/8/8/ppQKPPP1/k7 w - -
>or in PGN...
>------------------------------------------------------------------
>[Event "M12"]
>[Site ""]
>[Date "1994.05.09"]
>[Round ""]
>[White ""]
>[Black ""]
>[Result "1-0"]
>[BlackElo ""]
>[WhiteElo ""]
>[FEN "8/8/8/8/8/8/ppQKPPP1/k7 w - - 0 1"]
>1. Qc3 Kb1 2. Qd3+ Ka1 3. Qd4 Kb1 4. Qe4+ Ka1 5. Qe5 Kb1 6. Qf5+ Ka1 7.
>Qf6 Kb1 8. Qg6+ Ka1 9. Qg7 Kb1 10. Qh7+ Ka1 11. Qh8 Kb1 12. Qh1# 1-0
My program takes less than 5 seconds for this position (on a Sparc 20):
0:00.015 1 3 +2.066 Qxb2+ Kxb2
0:00.028 1 30 +2.273 Qb3 b1=Q
0:00.036 1 31 +11.860 Qc3
0:00.048 1. 72 +11.860 Qc3
0:00.057 2 76 +11.926 Qc3 Kb1
0:00.079 2. 261 +11.926 Qc3 Kb1
0:00.099 3 511 +11.946 Qc3 Kb1 Qd3+ Ka1 Qd4 Kb1
0:00.158 3. 1485 +11.946 Qc3 Kb1 Qd3+ Ka1 Qd4 Kb1
0:00.242 4 3256 +12.000 Qc3 Kb1 Qd3+ Ka1 Qd4 Kb1 Qe4+ Ka1
0:00.367 4. 5850 +12.000 Qc3 Kb1 Qd3+ Ka1 Qd4 Kb1 Qe4+ Ka1
0:00.733 5 14768 +11.980 Qc3 Kb1 Qd3+ Ka1 Qd4 Kb1 Qe4+ Ka1 Qe5 Kb1 Qf5+ Ka1
0:00.855 5. 18011 +11.980 Qc3 Kb1 Qd3+ Ka1 Qd4 Kb1 Qe4+ Ka1 Qe5 Kb1 Qf5+ Ka1
0:01.360 6 32372 +11.940 Qc3 Kb1 Qd3+ Ka1 Qd4 Kb1 Qe4+ Ka1 Qe5 Kb1 Qf5+ Ka1 Qf6 Kb1 Qg6+ Ka1
0:01.480 6. 36208 +11.940 Qc3 Kb1 Qd3+ Ka1 Qd4 Kb1 Qe4+ Ka1 Qe5 Kb1 Qf5+ Ka1 Qf6 Kb1 Qg6+ Ka1
0:01.998 7 54416 +11.920 Qc3 Kb1 Qd3+ Ka1 Qd4 Kb1 Qe4+ Ka1 Qe5 Kb1 Qf5+ Ka1 Qf6 Kb1 Qg6+ Ka1 Qg7 Kb1 Qh7+ Ka1
0:02.344 7. 64873 +11.920 Qc3 Kb1 Qd3+ Ka1 Qd4 Kb1 Qe4+ Ka1 Qe5 Kb1 Qf5+ Ka1 Qf6 Kb1 Qg6+ Ka1 Qg7 Kb1 Qh7+ Ka1
0:03.829 8 110738 +Mate12 Qc3 Kb1 Qd3+ Ka1 Qd4 Kb1 Qe4+ Ka1 Qe5 Kb1 Qf5+ Ka1 Qf6 Kb1 Qg6+ Ka1 Qg7 Kb1 Qh7+ Ka1 Qh8 Kb1 Qh1#
0:04.398 8. 127839 +Mate12 Qc3 Kb1 Qd3+ Ka1 Qd4 Kb1 Qe4+ Ka1 Qe5 Kb1 Qf5+ Ka1 Qf6 Kb1 Qg6+ Ka1 Qg7 Kb1 Qh7+ Ka1 Qh8 Kb1 Qh1#
>---------------------------------------------------------------------
>[Event "M12"]
>[Site ""]
>[Date "1994.02.06"]
>[Round ""]
>[White ""]
>[Black ""]
>[Result "1-0"]
>[BlackElo ""]
>[WhiteElo ""]
>[FEN "8/2Qp4/8/8/8/1r6/kp1p4/qrbK4 w - - 0 1"]
>1. Qa5+ Ra3 2. Qd5+ Rb3 3. Qa8+ Ra3 4. Qg8+ Rb3 5. Qd5 d6 6. Qa8+ Ra3 7.
>Qg8+ d5 8. Qxd5+ Rb3 9. Qa8+ Ra3 10. Qg8+ Rb3 11. Qd5 Ka3 12. Qa5# 1-0
After 8 minutes and 16 plies, my program hadn't found this. Turning off
nullmove here helps a lot:
0:00.002 1 2 -19.430 Qxd7
0:00.014 1. 36 -19.430 Qxd7
0:00.023 2 61 -19.383 Qxd7 Ka3
0:00.057 2 156 > +0.000 Qa5+ Ra3 Qd5+ Rb3 Qa5+
0:00.102 2 326 +0.000 Qa5+ Ra3 Qd5+ Rb3 Qa5+ etc.
0:00.115 2. 380 +0.000 Qa5+ Ra3 Qd5+ Rb3 Qa5+ etc.
0:00.210 3 1460 +0.000 Qa5+ Ra3 Qd5+ Rb3 Qa5+ etc.
0:00.237 3. 1768 +0.000 Qa5+ Ra3 Qd5+ Rb3 Qa5+ etc.
0:00.391 4 4446 +0.000 Qa5+ Ra3 Qd5+ Rb3 Qa5+ etc.
0:00.454 4. 5515 +0.000 Qa5+ Ra3 Qd5+ Rb3 Qa5+ etc.
0:00.761 5 13226 +0.000 Qa5+ Ra3 Qd5+ Rb3 Qa5+ etc.
0:00.872 5. 15967 +0.000 Qa5+ Ra3 Qd5+ Rb3 Qa5+ etc.
0:01.611 6 35223 +0.000 Qa5+ Ra3 Qd5+ Rb3 Qa5+ etc.
0:01.856 6. 41572 +0.000 Qa5+ Ra3 Qd5+ Rb3 Qa5+ etc.
0:04.228 7 103676 +Mate12 Qa5+ Ra3 Qd5+ Rb3 Qa8+ Ra3 Qg8+ Rb3 Qd5 d6 Qa8+ Ra3 Qg8+ d5 Qxd5+ Rb3 Qa8+ Ra3 Qg8+ Rb3 Qd5 Ka3 Qa5#
0:05.171 7. 128510 +Mate12 Qa5+ Ra3 Qd5+ Rb3 Qa8+ Ra3 Qg8+ Rb3 Qd5 d6 Qa8+ Ra3 Qg8+ d5 Qxd5+ Rb3 Qa8+ Ra3 Qg8+ Rb3 Qd5 Ka3 Qa5#
Ronald de Man
Robert Hyatt
ShaktiFire (shakt...@aol.com) wrote:
: :::Well for a while we will have only minor improvements (MMX gives only
The above calculation is also slightly broken. Example:
1983 Belle: 160,000 NPS.. typical commercial chess program: 3,000 NPS
difference: factor of roughly 50.
1997 DB: 1,000,000,000 NPS.. typical commercial program 100,000 NPS
difference: factor of 10,000. The Special-purpose machines are pulling
away, not "getting caught"...
A sobering thought when you look at it...
brucemo
Lonnie wrote:
> Vincent,
> Wow!! If I recall right and this is not to slam ya' but your program DIEP came in at
> the bottom end of the barrel in the Dutch championships and crafty was tied for third
> place I believe at Jakrata which in my opinion was a stronger field. All programs
> need improvement.
Fourth, tied with another instance of Crafty :-)
bruce
Harald Faber
quoting a mail from rebchess # xs4all.nl
Hello Ed,
ES> From: Ed Schroder <rebc...@xs4all.nl>
ES> Subject: Re: Deep Blue and ply depth...
ES> Organization: Schroder BV
ES> [FEN "8/8/8/8/8/8/ppQKPPP1/k7 w - - 0 1"]
ES>
ES> 1. Qc3 Kb1 2. Qd3+ Ka1 3. Qd4 Kb1 4. Qe4+ Ka1 5. Qe5 Kb1 6. Qf5+ Ka1 7.
ES> Qf6 Kb1 8. Qg6+ Ka1 9. Qg7 Kb1 10. Qh7+ Ka1 11. Qh8 Kb1 12. Qh1# 1-0
ES> ---------------------------------------------------------------------
Immediately (0-1sec) found with settings look below.
As I heard Genius5 solves this also immediately but only sees draw in
second position.
ES> [FEN "8/2Qp4/8/8/8/1r6/kp1p4/qrbK4 w - - 0 1"]
ES>
ES> 1. Qa5+ Ra3 2. Qd5+ Rb3 3. Qa8+ Ra3 4. Qg8+ Rb3 5. Qd5 d6 6. Qa8+ Ra3 7.
ES> Qg8+ d5 8. Qxd5+ Rb3 9. Qa8+ Ra3 10. Qg8+ Rb3 11. Qd5 Ka3 12. Qa5# 1-0
ES> ------------------------------------------------------------------------
ES>
ES> I was wrong about "no checks", there are many, but what impresses me is
ES> the number of "quiet moves" in position-1 and especially the "quiet
ES> move" 11.Qd5 in position-2.
ES> I haven't checked the 2 positions with CM4000/5000, all I rememeber is
ES> that the ChessMachine (King) solved both positions very fast.
CM5000 on 486/133 16MB hash but with changed style (mobility 138, pawn
weakness 81, attacker/defender -20, selectivity 6) finds move in 54
seconds, meaning finds move ad hoc but first eval is draw and after 54 sec
announcing mate in 11
Harald
--
Ziggy
MMX is not designed to help chess programs. Only games written for it,
graphic intensive games. Graphic designers benefit a lot from it. It speeds
certain filters up to %500. In any event, a $200 Matrox Millenium card does
better anyways.
Robert Hyatt
Ziggy (je...@mursuky.campus.mci.net) wrote:
: MMX is not designed to help chess programs. Only games written for it,
: graphic intensive games. Graphic designers benefit a lot from it. It speeds
: certain filters up to %500. In any event, a $200 Matrox Millenium card does
: better anyways.
MMX really doesn't help chess programs, but the larger on-chip cache that
comes with it does. I still don't have any results that put a P5/200/mmx
faster than a p6/200, as has been claimed here.
Valavan Manohararajah
In article <5dsduh$t...@juniper.cis.uab.edu>,
Are you sure about this? From my look at the MMX instruction set there are
some 64 bit operations that a BitBoard program should benefit from.... I
even thought about scrapping my 0x88 program and doing a bitboard version of
Rajah...
--
-------------------------------------------------------------------------------
man...@ecf.utoronto.ca | 3rd year Comp Eng., University of Toronto
Valavan Manohararajah |
-------------------------------------------------------------------------------
Michael Nyman
I'm not really a guru at computer chess programming, so you can
correct me if I'm wrong. You say that in 1983, Belle searched
about 160K NPS and now in 1997 typical commercial programs (by
typical I'm assuming you mean programs such as Chess Genius) search
100K NPS, which is 37.5 percent less. Correct so far? Then your
conclusion is basically that it's sobering how special-purpose
machines are pulling away.
Well, I don't make the same conclusions
that you do. I'm just guessing here, but I believe that Chess
Genius would have little or no problem beating the 1983 Belle
machine, even though genius(according to your calculations)
is about 37.5 percent slower.
My conclusion is that special-purpose machines are not pulling
away in chess skill, only in NPS. But the stats you show
seem to demonstrate that there is a lot more to computer chess
than NPS. So, am I wrong? Would '83 Belle crush Chess Genius
because it searchs so many more NPS? Or what did you mean by
typical commercial programs?
I think that until a program has
enough "NPS" so see a forced win, there will always be a horizon
effect that might screw a program up. I'm not even convinced that
searching farther and farther past a certain point won't make a
program worse. With PC programs doing a heck of a lot better than
I feel you give them credit for, I believe skill comes mostly
from a great evaluation function.
Maybe if instead of looking at NPS, you looked at some ratio of
NPS to ELO rating. Then I imagine you'd have a different opinion
of whether the gap were closing or not. I know you have all the
information available to you to make these calculations. Could
you please tell me how NPS to ELO rating is going for PC programs
versus special purpose programs?
What do you think?
Cheers!
-Michael
Ziggy
Most people recognize that a well built chess engine will compete with
mainframe chess engines. Deep Blue can run brute force circles around PC's,
but that doesn't stop the PC's from winning and drawing games against Deep
Blue. Also, the Deep Blue engineers could easily hook up 1000 or more
processors to the mainframe, to boost the computing power. That is not the
reason for Deep Blue. It's an attempt to harness the existing power of 256
processors. Rumored is that the current voice recognition software from IBM
was born out of some of the programming techniques discovered on the Deep
Blue project. Of course, other IBM projects also contributed to the voice
recognition engine.
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
<5dppl3$q...@juniper.cis.uab.edu>...
Harald Faber
quoting a mail from hyatt # crafty.cis.uab.edu
Hello Robert,
RH> From: hy...@crafty.cis.uab.edu (Robert Hyatt)
RH> Subject: Re: Deep Blue and ply depth...
RH> Organization: CIS, University of Alabama at Birmingham
RH> 1983 Belle: 160,000 NPS.. typical commercial chess program: 3,000 NPS
RH> difference: factor of roughly 50.
RH>
RH> 1997 DB: 1,000,000,000 NPS.. typical commercial program 100,000 NPS
RH> difference: factor of 10,000. The Special-purpose machines are pulling
RH> away, not "getting caught"...
RH> A sobering thought when you look at it...
If you are right. But did the db team improve that much?? I remember last
year they had 1.000.000nps so they are supposed to have accelerated so
much? I cannot believe.
Harald
--
Robert Hyatt
Michael Nyman (michae...@techne-sys.com) wrote:
: Robert Hyatt wrote:
: >
: > : :::Well for a while we will have only minor improvements (MMX gives only
: > : :::10% more due to extra cache and this doesen't help much a chess
: > : program)
: > : :::or a big step we really need next tecnology (P7 which should
: > : :::double or triple present clocks), but we have to wait until next
: > : :::century.
: >
: > : Well there will be some improvements in speed before P7 comes out.
: >
: > : April 1 is due the PP233, July 1 PP266, and late 97 the PP300.
: >
: > : Further improvements likely in 1998.
: >
: > The above calculation is also slightly broken. Example:
: > 1983 Belle: 160,000 NPS.. typical commercial chess program: 3,000 NPS
: > difference: factor of roughly 50.
: >
: > 1997 DB: 1,000,000,000 NPS.. typical commercial program 100,000 NPS
: > difference: factor of 10,000. The Special-purpose machines are pulling
: > away, not "getting caught"...
: >
: > A sobering thought when you look at it...
: I'm not really a guru at computer chess programming, so you can
: correct me if I'm wrong. You say that in 1983, Belle searched
: about 160K NPS and now in 1997 typical commercial programs (by
: typical I'm assuming you mean programs such as Chess Genius) search
: 100K NPS, which is 37.5 percent less. Correct so far? Then your
: conclusion is basically that it's sobering how special-purpose
: machines are pulling away.
: Well, I don't make the same conclusions
: that you do. I'm just guessing here, but I believe that Chess
: Genius would have little or no problem beating the 1983 Belle
: machine, even though genius(according to your calculations)
: is about 37.5 percent slower.
Possibly, possibly not. I remember the 1986 ACM event which was so close
to the world championship in 1986 that CB and HiTech didn't go. Monty talked
Ken into entering Belle, and it won easily. Remember, this program/machine
earned a 2200+ uscf rating and had a couple of 2350+ performance ratings,
one at the world open I believe. Main thing is that 1983 belle was before
null-move, but it could be modified or redone using the same PLA architecture
if Ken had the inclination...
: My conclusion is that special-purpose machines are not pulling
: away in chess skill, only in NPS. But the stats you show
: seem to demonstrate that there is a lot more to computer chess
: than NPS. So, am I wrong? Would '83 Belle crush Chess Genius
: because it searchs so many more NPS? Or what did you mean by
: typical commercial programs?
The point I was after was that special purpose hardware is not being
caught performance wise by the micros. the special purpose stuff is so
much faster overall today than 10 years ago. The gap has widened. There
is *nothing* that prevents DB from copying the genius eval and genius
search if they wanted to. The hardware doesn't have any particular limits,
and for those that are aggravating, the hardware could be modified. And
then you'd have genius at 100K vs genius at 1B nodes per second. Care to
make a wager on the outcome?
: I think that until a program has
: enough "NPS" so see a forced win, there will always be a horizon
: effect that might screw a program up. I'm not even convinced that
: searching farther and farther past a certain point won't make a
: program worse. With PC programs doing a heck of a lot better than
: I feel you give them credit for, I believe skill comes mostly
: from a great evaluation function.
I'm not down on the PC programs you should notice, as I'm the author of one
that's not playing badly at all. However, I know it's not going to take on
DB in a match and win, but I'd be more than willing to try any other program
and not expect to get crushed...
: Maybe if instead of looking at NPS, you looked at some ratio of
: NPS to ELO rating. Then I imagine you'd have a different opinion
: of whether the gap were closing or not. I know you have all the
: information available to you to make these calculations. Could
: you please tell me how NPS to ELO rating is going for PC programs
: versus special purpose programs?
I think you misread the original. The point was that PC hardware is closer
to special purpose hardware today than 10 years ago. I pointed out that that
was not just incorrect, but grossly incorrect. The pure numbers bear this out.
Chess skill is related to NPS, because any program is better if it is faster, and
there's no evidence to counter this. This isn't a deep blue vs genius
comparison, just DB hardware vs PC hardware. There the winner is easy to see...
: What do you think?
: Cheers!
: -Michael
Robert Hyatt
Valavan Manohararajah (man...@ecf.toronto.edu) wrote:
: In article <5dsduh$t...@juniper.cis.uab.edu>,
: Robert Hyatt <hy...@crafty.cis.uab.edu> wrote:
: >Ziggy (je...@mursuky.campus.mci.net) wrote:
: >: MMX is not designed to help chess programs. Only games written for it,
: >: graphic intensive games. Graphic designers benefit a lot from it. It speeds
: >: certain filters up to %500. In any event, a $200 Matrox Millenium card does
: >: better anyways.
: >
: >
: >MMX really doesn't help chess programs, but the larger on-chip cache that
: >comes with it does. I still don't have any results that put a P5/200/mmx
: >faster than a p6/200, as has been claimed here.
: Are you sure about this? From my look at the MMX instruction set there are
: some 64 bit operations that a BitBoard program should benefit from.... I
: even thought about scrapping my 0x88 program and doing a bitboard version of
: Rajah...
problem is that no compilers emit this stuff yet. the MMX stuff might
help the 64bit stuff then, but it will probably be a while before we get
a compiler to produce such code. We have not yet seen the P6's conditional
move instruction show up, so MMX instructions are probably a ways away, although
it will all help sooner or later...
j...@sysaid.nl
In article <330217...@techne-sys.com>,
Michael Nyman <michae...@techne-sys.com> wrote:
> > 1983 Belle: 160,000 NPS.. typical commercial chess program: 3,000 NPS
> > difference: factor of roughly 50.
> >
> > 1997 DB: 1,000,000,000 NPS.. typical commercial program 100,000 NPS
> > difference: factor of 10,000. The Special-purpose machines are pulling
> > away, not "getting caught"...
> >
> > A sobering thought when you look at it...
>
> I'm not really a guru at computer chess programming, so you can
> correct me if I'm wrong. You say that in 1983, Belle searched
> about 160K NPS and now in 1997 typical commercial programs (by
> typical I'm assuming you mean programs such as Chess Genius) search
> 100K NPS, which is 37.5 percent less. Correct so far? Then your
> conclusion is basically that it's sobering how special-purpose
> machines are pulling away.
> Well, I don't make the same conclusions
> that you do. I'm just guessing here, but I believe that Chess
> Genius would have little or no problem beating the 1983 Belle
> machine, even though genius(according to your calculations)
> is about 37.5 percent slower.
> My conclusion is that special-purpose machines are not pulling
> away in chess skill, only in NPS. But the stats you show
> seem to demonstrate that there is a lot more to computer chess
> than NPS. So, am I wrong? Would '83 Belle crush Chess Genius
> because it searchs so many more NPS? Or what did you mean by
> typical commercial programs?
Special purpose machines aren't falling behind in chess skill either.
I think it's safe to say that PC programs aren't the only ones who
have increased in "knowledge" (skill, evaluation, whatever -- I'll
call it knowledge here).
No, DB isn't stronger than PC programs *just* because of its speed.
As Hsu et.al. spend quite some time working on DB, you can expect that
the knowledge built into DB has evolved just as well. Perhaps a little
more than in PC programs, perhaps a little less, but for all we know,
about as much. If DB were "merely" an 80's program running on 2000's
hardware, I'd agree that there is no clear case for it being much
stronger than modern PC programs (other than its results OTB against
Kasparov of course). But it isn't. It's a 2000's program running
on 2000's hardware.
Let's hope that DB will start playing public games after the match
against Kasparov. Particularly, in the interest of science, against
PC programs. I for one can't wait to see that. (:
Jack van Rijswijck
j...@sysaid.nl
-------------------==== Posted via Deja News ====-----------------------
http://www.dejanews.com/ Search, Read, Post to Usenet
Robert Hyatt
Harald Faber (Harald...@p21.f2.n1.z1001.fidonet.org) wrote:
: quoting a mail from hyatt # crafty.cis.uab.edu
: Hello Robert,
: RH> From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: RH> Subject: Re: Deep Blue and ply depth...
: RH> Organization: CIS, University of Alabama at Birmingham
: RH> 1983 Belle: 160,000 NPS.. typical commercial chess program: 3,000 NPS
: RH> difference: factor of roughly 50.
: RH>
: RH> 1997 DB: 1,000,000,000 NPS.. typical commercial program 100,000 NPS
: RH> difference: factor of 10,000. The Special-purpose machines are pulling
: RH> away, not "getting caught"...
: RH> A sobering thought when you look at it...
: If you are right. But did the db team improve that much?? I remember last
: year they had 1.000.000nps so they are supposed to have accelerated so
: much? I cannot believe.
your data's bad. they were doing 2-3 million nodes per second *per chip*.
last year they had 256 chips. This year 1024 supposedly. *big* numbers...
: Harald
: --
Tom C. Kerrigan
Ziggy (je...@mursuky.campus.mci.net) wrote:
> Most people recognize that a well built chess engine will compete with
> mainframe chess engines. Deep Blue can run brute force circles around PC's,
> but that doesn't stop the PC's from winning and drawing games against Deep
> Blue. Also, the Deep Blue engineers could easily hook up 1000 or more
> processors to the mainframe, to boost the computing power. That is not the
The implication here, as I see it, is that DB is a mainframe. Not exactly
the case.
Running brute force circles around something WILL keep the something from
winning.
Cheers,
Tom
Robert Hyatt
Tom C. Kerrigan (kerr...@merlin.pn.org) wrote:
: Cheers,
: Tom
It's a hybrid. the SP is a classic mainframe type architecture, but it has
many processors, and each SP processor has DB chess processors attached as
peripheral devices. It's cute, fast as hell, and strong as hell.
One funny thing I haven't heard in a while is a common statement here a few
years back when I was running Cray Blitz in ACM events. It was so common to
hear "what if Lang had a cray, how good would genius be" and "you big iron guys
only win because you have huge computers." It seems that some of us "big iron"
guys are doing quite well with a PC now. Does anyone *really* believe that the
DB guys couldn't put together a PC program that's competitive? I did it. They
could. And if you believe that, is it such a stretch to think that DB might
just be a monster after all, and not just a slightly fast dumb program?
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
> The implication here, as I see it, is that DB is a mainframe. Not
> exactly the case.
> Running brute force circles around something WILL keep the something
> from winning.
> Cheers,
> Tom
: It's a hybrid. the SP is a classic mainframe type architecture, but
: it has many processors, and each SP processor has DB chess processors
: attached as peripheral devices. It's cute, fast as hell, and strong as
: hell.
: One funny thing I haven't heard in a while is a common statement here a
: few years back when I was running Cray Blitz in ACM events. It was so
: common to hear "what if Lang had a cray, how good would genius be" and
: "you big iron guys only win because you have huge computers." It seems
: that some of us "big iron" guys are doing quite well with a PC now.
: Does anyone *really* believe that the DB guys couldn't put together a
: PC program that's competitive? I did it. They could. And if you
: believe that, is it such a stretch to think that DB might just be a
: monster after all, and not just a slightly fast dumb program?
It happened before...
I remember the 1970-1992 period. Nobody thought it would be possible that
micro computers could beat the main frame computers.
Still it happened in Madrid 1992 and now it's common.
Bob, what about 16. c4???
I counted the plies, the losing line is just 12 silly plies!
Deep Blue Prototype did not see it.
Genius5 sees the loss in just 8 seconds!
Do you have an explanation for that?
I mean with 2,000,000 - 3,000,000 NPS DB should see this in one second.
Puzzled...
- Ed Schroder -
BTW, I did an overnight analysis on 16. c4??
Rebel8 on 9 plies fails low and the score drops till -1.xx
On ply=14 the score for 16.c4 drops again now to -3.xx
Robert Hyatt
Ed Schroder (rebc...@xs4all.nl) wrote:
: From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: > The implication here, as I see it, is that DB is a mainframe. Not
: > exactly the case.
: > Running brute force circles around something WILL keep the something
: > from winning.
: > Cheers,
: > Tom
: : It's a hybrid. the SP is a classic mainframe type architecture, but
: : it has many processors, and each SP processor has DB chess processors
: : attached as peripheral devices. It's cute, fast as hell, and strong as
: : hell.
: : One funny thing I haven't heard in a while is a common statement here a
: : few years back when I was running Cray Blitz in ACM events. It was so
: : common to hear "what if Lang had a cray, how good would genius be" and
: : "you big iron guys only win because you have huge computers." It seems
: : that some of us "big iron" guys are doing quite well with a PC now.
: : Does anyone *really* believe that the DB guys couldn't put together a
: : PC program that's competitive? I did it. They could. And if you
: : believe that, is it such a stretch to think that DB might just be a
: : monster after all, and not just a slightly fast dumb program?
: It happened before...
: I remember the 1970-1992 period. Nobody thought it would be possible that
: micro computers could beat the main frame computers.
: Still it happened in Madrid 1992 and now it's common.
What happened in 1992? Two of the strongest mainframe programs didn't
participate, Cray Blitz and Deep Thought. Tis funny that Jakarta was
tainted because a couple of commercial programs didn't attend, but
the '92 WCCC was proof that the micros are better than the mainframes
when the mainframes weren't there, excepting HiTech as I recall, and a
couple of parallel programs that have not been consistently high finishers
in open competition..
: Bob, what about 16. c4???
: I counted the plies, the losing line is just 12 silly plies!
: Deep Blue Prototype did not see it.
: Genius5 sees the loss in just 8 seconds!
could be any of a dozen things. How about Rc6 in the Kasparov game where
all the micros wanted to play (either Rc6 or Rc7 whichever DB *didn't* play,
and all would have lost about 20 plies later, because the rook was a critical
piece there, a little later. Remember Kaspy had mate in 1, so when DB ran out
of checks, the game was over if Kaspy wasn't mated. But DB saw it *all*...
for c4, DB might have looked deep enough to find that it was not a bad move.
IE, it might look bad at 10, good at 14, bad at 16 plies. If they were at
14, that would be bad for them. They might have seen lots of things. We
won't know until Hsu one day tells us.
: Do you have an explanation for that?
: I mean with 2,000,000 - 3,000,000 NPS DB should see this in one second.
: Puzzled...
I see this all the time in Crafty vs other programs. I've seen Rebel 8,
Hiarcs 5, genius, and so forth make a move after a long think that Crafty
instantly fails high on and continues to fail high on iteration after
iteration. I also see the opposite, where Crafty is happy, and makes a move,
and Lonnie will tell me that he's at +1.5 almost instantly, yet Crafty doesn't
see it. If you are using any selective search pruning, you are going to get
zapped, and also you will do some zapping at times. DB's more consistent since
they don't rely on *any* tricks to get the depth they are seeing, no null move,
no forward pruning, etc...
It might be that if you let the micros reach 13-14-15-16 or whatever, they
might suddenly think that c4 looks better again. There's always a danger when
you outsearch your opponents so badly, that you see things they will never see,
and play accordingly. On occasion you will lose because of how *you* play,
rather than how your opponent plays, because you see all moves lose a pawn at
30 plies, and often will simply offer the pawn *now* as the best alternative,
even if your opponent could *never* see how to win it...
: - Ed Schroder -
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: It happened before...
: I remember the 1970-1992 period. Nobody thought it would be possible
: that
: micro computers could beat the main frame computers.
: Still it happened in Madrid 1992 and now it's common.
> What happened in 1992? Two of the strongest mainframe programs didn't
> participate, Cray Blitz and Deep Thought. Tis funny that Jakarta was
> tainted because a couple of commercial programs didn't attend, but
> the '92 WCCC was proof that the micros are better than the mainframes
> when the mainframes weren't there, excepting HiTech as I recall, and a
> couple of parallel programs that have not been consistently high
> finishers in open competition..
Hitech was one of the favorites for the WC title (150,000 NPS)
Zugzwang was there (over 100,000 NPS)
There was a Cray computer (nodes 40,000 - 50,000)
But instead a micro searching just 2,000 NPS won in Madrid!
NPS are not holy...
: Bob, what about 16. c4???
: I counted the plies, the losing line is just 12 silly plies!
: Deep Blue Prototype did not see it.
: Genius5 sees the loss in just 8 seconds!
> could be any of a dozen things. How about Rc6 in the Kasparov game
> where all the micros wanted to play (either Rc6 or Rc7 whichever DB
> *didn't* play, and all would have lost about 20 plies later, because
> the rook was a critical piece there, a little later. Remember Kaspy
> had mate in 1, so when DB ran out of checks, the game was over if Kaspy
> wasn't mated. But DB saw it *all*...
Sure, but maybe you can help me out here and explain which move number,
the wrong move, the right move and perhaps the why.
I think this is the game you are talking about:
[Event "Man Vs. Machine 1996"]
[Site "ACM round1"]
[Date "1996.02.11"]
[Round "1"]
[White "DEEP BLUE"]
[Black "Kasparov, G."]
[Result "1-0"]
1. e4 c5 2. c3 d5 3. exd5 Qxd5 4. d4 Nf6 5. Nf3 Bg4 6. Be2 e6 7. h3 Bh5
8. O-O Nc6 9. Be3 cxd4 10. cxd4 Bb4 11. a3 Ba5 12. Nc3 Qd6 13. Nb5 Qe7
14. Ne5 Bxe2 15. Qxe2 O-O 16. Rac1 Rac8 17. Bg5 Bb6 18. Bxf6 gxf6 19. Nc4
Rfd8 20. Nxb6 axb6 21. Rfd1 f5 22. Qe3 Qf6 23. d5 Rxd5 24. Rxd5 exd5 25.
b3 Kh8 26. Qxb6 Rg8 27. Qc5 d4 28. Nd6 f4 29. Nxb7 Ne5 30. Qd5 f3 31. g3
Nd3 32. Rc7 Re8 33. Nd6 Re1+ 34. Kh2 Nxf2 35. Nxf7+ Kg7 36. Ng5+ Kh6 37.
Rxh7# 1-0
I assume you mean 32. Rc7 is the critical move?
And 32. Rc6 is the bad move concerning DB losing after 20 plies?
Could be another very interesting position to discover more about the
power of Deep Blue.
> for c4, DB might have looked deep enough to find that it was not a bad
> move. IE, it might look bad at 10, good at 14, bad at 16 plies. If
> they were at 14, that would be bad for them. They might have seen lots
> of things. We won't know until Hsu one day tells us.
Not in this case, see comment below.
: Do you have an explanation for that?
: I mean with 2,000,000 - 3,000,000 NPS DB should see this in one second.
: Puzzled...
> It might be that if you let the micros reach 13-14-15-16 or whatever,
> they might suddenly think that c4 looks better again.
You say "could be any of a dozen things", "might be", etc but did you
studied the position? I counted the plies, it's just 12 plies. This is
not a "maybe" but a fact!
Here are some analysis:
16.c4 Qh4 17.cxb5 Rh6 18.Nf6+ Kd8 19.h3 Bxh3 20.gxh3 Qxh3 21.Kg1 Qh2# (12)
16.c4 Qh4 17.cxb5 Rh6 18.Nf6+ Kd8 19.h3 Bxh3 20.Be6+ 21.Kg1 Qh2# (12)
16.c4 Qh4 17.g3 Qh3 18.Qe2 fxg3 19.fxg3 Rgx3 (8)
16.c4 Qh4 17.g3 Qh3 18.Qf3 Rh6 19.Qg2 Qxg2 20.Kxg2 Bh3+ (10)
If you find a deeper (better) defense for white I am all ears.
> There's always a danger when you outsearch your opponents so badly,
> that you see things they will never see, and play accordingly.
> On occasion you will lose because of how *you* play, rather than how
> your opponent plays, because you see all moves lose a pawn at
> 30 plies, and often will simply offer the pawn *now* as the best
> alternative, even if your opponent could *never* see how to win it...
Not valid in this case:
- Rebel8 after 14 ply says -3.xx
- DB on move 20 has to give the queen for a rook (game over)
For clearness sake the whole game once again...
[Event ""]
[Site ""]
[Date "1995.??.??"]
[White "Deep Blue"]
[Black "Fritz"]
[Result "0-1"]
1. e4 c5 2. Nf3 Nc6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 e5 6. Ndb5 d6 7. Bg5
a6 8. Na3 b5 9. Bxf6 gxf6 10. Nd5 f5 11. Bd3 Be6 12. Qh5 f4 13. O-O Rg8
14. Kh1 Rg6 15. Qd1 Rc8 16. c4 Qh4 17. g3 Qh3 18. Qd2 f3 19. Rg1 Rh6 20.
Qxh6 Qxh6 21. cxb5 Bxd5 22. exd5 Nb4 23. Bf5 Rc5 24. bxa6 Nxa6 25. Nc2
Qd2 26. Ne1 Rxd5 27. Nxf3 Qxf2 28. Be4 Ra5 29. Rg2 Qe3 30. Re1 Qh6 31.
Bc6+ Kd8 32. a3 f5 33. Rc2 Rc5 34. Rxc5 Nxc5 35. Rf1 Be7 36. a4 f4 37.
gxf4 Qxf4 38. Rg1 Nxa4 39. b4 Qxb4 0-1
I simply can't understand 16. c4??
It's losing and very quick!
I see 3 possibilities:
1.. DB had a bug which was critical for 16.c4??
2.. My "12 plies" analyse is in error (white may have a longer defense?)
3.. DB (HK edition) is not the deep searcher I am always told.
- Ed Schroder -
Robert Hyatt
Ed Schroder (rebc...@xs4all.nl) wrote:
: From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: : It happened before...
: : I remember the 1970-1992 period. Nobody thought it would be possible
: : that
: : micro computers could beat the main frame computers.
: : Still it happened in Madrid 1992 and now it's common.
: > What happened in 1992? Two of the strongest mainframe programs didn't
: > participate, Cray Blitz and Deep Thought. Tis funny that Jakarta was
: > tainted because a couple of commercial programs didn't attend, but
: > the '92 WCCC was proof that the micros are better than the mainframes
: > when the mainframes weren't there, excepting HiTech as I recall, and a
: > couple of parallel programs that have not been consistently high
: > finishers in open competition..
: Hitech was one of the favorites for the WC title (150,000 NPS)
: Zugzwang was there (over 100,000 NPS)
: There was a Cray computer (nodes 40,000 - 50,000)
: But instead a micro searching just 2,000 NPS won in Madrid!
: NPS are not holy...
they never have been. *BUT*... if you take two programs with reasonable
adequate evaluations, and reasonable adequate searches, *THEN* NPS
rules... There have always been fast programs that didn't play that great.
and slow programs that did. But there have also been fast programs that
did play decent chess. Cray Blitz certainly played some outstanding games
over the years, including beating GM's in tournament games, among other
things...
: : Bob, what about 16. c4???
: : I counted the plies, the losing line is just 12 silly plies!
: : Deep Blue Prototype did not see it.
: : Genius5 sees the loss in just 8 seconds!
: > could be any of a dozen things. How about Rc6 in the Kasparov game
: > where all the micros wanted to play (either Rc6 or Rc7 whichever DB
: > *didn't* play, and all would have lost about 20 plies later, because
: > the rook was a critical piece there, a little later. Remember Kaspy
: > had mate in 1, so when DB ran out of checks, the game was over if Kaspy
: > wasn't mated. But DB saw it *all*...
: Sure, but maybe you can help me out here and explain which move number,
: the wrong move, the right move and perhaps the why.
I really have no way to find this out. However I have seen positions that
go like this: 10 ply, fail low... 11 ply fail low again, 12 plies fail
high and discover that the losing move is really a winner. I've also seen
the exact opposite, but it's all depth dependent. I lost a game to chiptest
because of this happening to me. A particular move looked good at depth=10
on the Cray (1987 I believe), but failed low at depth=11 and we switched to
a move that would have really failed low at 12. And at 12, the original move
would have failed high big-time. ChipTest (deep thought predecessor) saw that
Ne6 was a killer, and so did we, then we didn't, and we didn't play it and
lost. Either one ply less, or one ply more and we'd have at least won that
game. It happens... Luck has always been a factor. and always will be a
factor in a few games...
: I think this is the game you are talking about:
: [Event "Man Vs. Machine 1996"]
: [Site "ACM round1"]
: [Date "1996.02.11"]
: [Round "1"]
: [White "DEEP BLUE"]
: [Black "Kasparov, G."]
: [Result "1-0"]
: 1. e4 c5 2. c3 d5 3. exd5 Qxd5 4. d4 Nf6 5. Nf3 Bg4 6. Be2 e6 7. h3 Bh5
: 8. O-O Nc6 9. Be3 cxd4 10. cxd4 Bb4 11. a3 Ba5 12. Nc3 Qd6 13. Nb5 Qe7
: 14. Ne5 Bxe2 15. Qxe2 O-O 16. Rac1 Rac8 17. Bg5 Bb6 18. Bxf6 gxf6 19. Nc4
: Rfd8 20. Nxb6 axb6 21. Rfd1 f5 22. Qe3 Qf6 23. d5 Rxd5 24. Rxd5 exd5 25.
: b3 Kh8 26. Qxb6 Rg8 27. Qc5 d4 28. Nd6 f4 29. Nxb7 Ne5 30. Qd5 f3 31. g3
: Nd3 32. Rc7 Re8 33. Nd6 Re1+ 34. Kh2 Nxf2 35. Nxf7+ Kg7 36. Ng5+ Kh6 37.
: Rxh7# 1-0
: I assume you mean 32. Rc7 is the critical move?
: And 32. Rc6 is the bad move concerning DB losing after 20 plies?
Could be, yes. The thing is, after Rc7, later a knight move (a critical
knight move) is a discovered check. With the rook on c6 it isn't... I
think that Rc6 appears to win a pawn, or something, because most programs
would stick the rook on the 7th no matter what. I'll run this thru crafty
and post the results however...
: Could be another very interesting position to discover more about the
: power of Deep Blue.
: > for c4, DB might have looked deep enough to find that it was not a bad
: > move. IE, it might look bad at 10, good at 14, bad at 16 plies. If
: > they were at 14, that would be bad for them. They might have seen lots
: > of things. We won't know until Hsu one day tells us.
: Not in this case, see comment below.
: : Do you have an explanation for that?
: : I mean with 2,000,000 - 3,000,000 NPS DB should see this in one second.
: : Puzzled...
: > It might be that if you let the micros reach 13-14-15-16 or whatever,
: > they might suddenly think that c4 looks better again.
: You say "could be any of a dozen things", "might be", etc but did you
: studied the position? I counted the plies, it's just 12 plies. This is
: not a "maybe" but a fact!
: Here are some analysis:
: 16.c4 Qh4 17.cxb5 Rh6 18.Nf6+ Kd8 19.h3 Bxh3 20.gxh3 Qxh3 21.Kg1 Qh2# (12)
: 16.c4 Qh4 17.cxb5 Rh6 18.Nf6+ Kd8 19.h3 Bxh3 20.Be6+ 21.Kg1 Qh2# (12)
: 16.c4 Qh4 17.g3 Qh3 18.Qe2 fxg3 19.fxg3 Rgx3 (8)
: 16.c4 Qh4 17.g3 Qh3 18.Qf3 Rh6 19.Qg2 Qxg2 20.Kxg2 Bh3+ (10)
: If you find a deeper (better) defense for white I am all ears.
I'd discount all mate lines. I can't imagine any program overlooking those
except with severe bugs.... It's the other lines that would be iteresting
to see what they saw...
: > There's always a danger when you outsearch your opponents so badly,
: > that you see things they will never see, and play accordingly.
: > On occasion you will lose because of how *you* play, rather than how
: > your opponent plays, because you see all moves lose a pawn at
: > 30 plies, and often will simply offer the pawn *now* as the best
: > alternative, even if your opponent could *never* see how to win it...
: Not valid in this case:
: - Rebel8 after 14 ply says -3.xx
: - DB on move 20 has to give the queen for a rook (game over)
Don't forget, this wasn't DB... this was deep thought. Actually it was
deep thought hardware with the deep blue software running. Could easily
have bugs of course, and it was still fast as hell at only 10M nodes per
second (or less actually)...
: For clearness sake the whole game once again...
mclane
Ed Schroder <rebc...@xs4all.nl> wrote:
>> they never have been. *BUT*... if you take two programs with
>> reasonable adequate evaluations, and reasonable adequate searches,
>> *THEN* NPS rules...
>NPS rules...???
>Was Hitech a bad program then?
>You must be kidding!
>Hitech from Hans Berliner was *by far* the main candidate for the WCM
>title in Madrid 1992!! (because of the absent of Deep Thought)
>Hitech (150,000 NPS) lost from a 2,000 NPS program with just 512 Kb
>hash table. Also this 2,000 NPS program won the tournament with 4.5
>points out of 5.
>NPS rules...???
>NPS is important but it does not rule.
>Same tournament 3 years later, Deep Blue Prototype participated in
>Hong Kong. Same result, a micro won.
>Same NPS fight ended in favor for the micro.
>Deep Blue Prototype NPS 3,000,000
>Fritz NPS 100,000
>Madrid 1992 : 150,000 / 2000 = factor 75
>Hong Kong 1995 : 3,000,000 / 100,000 = factor 30
>NPS rules???
You see Ed, they suggest always the same wrong picture of
computerchess.
Deep Blue was killed by a program that is NOT the best chess program
in the world. It was killed by a program that knows not much about
chess. Any chessplayer who is playing against Fritz4 can easily see
/recognize the lack of knowledge.
But Deep Blue , the big iron, falls into the trap. Loses.
Over and over they try to convince people about IMPORTANCE of
Nodes/second.
Why do they do that ?! What is the interest they have ? Why is there a
MUST for them to show us this wrong but materialistic image ?
mclane
hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>Any one micro had a very small chance of winning the event, probably the
>best was at the 10% probability level. But there were a *bunch* of
>micros there, and the liklihood that one would win, with only one strong
>"mainframe" there suddenly became quite high. The winner in 1992 was
>basically chosen by a coin toss, not by playing skill.
Ich denke Ed, es ist sinnlos weiter zu diskutieren. Sie leben in IHRER
Welt und werden es nicht verstehen !
Sie wollen nichts sehen was *entgegen* ihrer, scheinbar mit einem
hohen Preis bezahlten, Ideologie an Ihre mit Scheuklappen versehenen
Augen dringt.
Ein Kamel geht nicht durch ein Nadelöhr. Und kein Reicher kommt in den
Himmel. Und kein Materialist wird Qualität verstehen.
Harald Faber
## Nachricht vom : 14.02.97 weitergeleitet
## Ursprung : /FIDO/REC.GAMES.CHESS.COMPUTER
## Ersteller war : hyatt # crafty.cis.uab.edu@1001:1/0
RH> From: hy...@crafty.cis.uab.edu (Robert Hyatt)
RH> Subject: Re: Deep Blue and ply depth...
RH> Organization: CIS, University of Alabama at Birmingham
: RH>> 1983 Belle: 160,000 NPS.. typical commercial chess program: 3,000
: RH>> NPS difference: factor of roughly 50.
: RH>>
: RH>> 1997 DB: 1,000,000,000 NPS.. typical commercial program 100,000 NPS
: RH>> difference: factor of 10,000. The Special-purpose machines are pulling
: RH>> away, not "getting caught"...
: RH>> A sobering thought when you look at it...
RH> : If you are right. But did the db team improve that much?? I remember
RH> last : year they had 1.000.000nps so they are supposed to have accelerated
RH> so : much? I cannot believe.
RH>
RH> your data's bad. they were doing 2-3 million nodes per second *per chip*.
RH> last year they had 256 chips. This year 1024 supposedly. *big*
RH> numbers...
Hae? 2-3mio nps x 256 makes at least half a billion, now they have 1
billion, just a factor of 2. Am I wrong again?
Tschuessikowski
Harald
--
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: Hitech was one of the favorites for the WC title (150,000 NPS)
: Zugzwang was there (over 100,000 NPS)
: There was a Cray computer (nodes 40,000 - 50,000)
: But instead a micro searching just 2,000 NPS won in Madrid!
: NPS are not holy...
> they never have been. *BUT*... if you take two programs with
> reasonable adequate evaluations, and reasonable adequate searches,
> *THEN* NPS rules...
NPS rules...???
Was Hitech a bad program then?
You must be kidding!
Hitech from Hans Berliner was *by far* the main candidate for the WCM
title in Madrid 1992!! (because of the absent of Deep Thought)
Hitech (150,000 NPS) lost from a 2,000 NPS program with just 512 Kb
hash table. Also this 2,000 NPS program won the tournament with 4.5
points out of 5.
NPS rules...???
NPS is important but it does not rule.
Same tournament 3 years later, Deep Blue Prototype participated in
Hong Kong. Same result, a micro won.
Same NPS fight ended in favor for the micro.
Deep Blue Prototype NPS 3,000,000
Fritz NPS 100,000
Madrid 1992 : 150,000 / 2000 = factor 75
Hong Kong 1995 : 3,000,000 / 100,000 = factor 30
NPS rules???
: > could be any of a dozen things. How about Rc6 in the Kasparov game
: > where all the micros wanted to play (either Rc6 or Rc7 whichever DB
: > *didn't* play, and all would have lost about 20 plies later, because
: > the rook was a critical piece there, a little later. Remember Kaspy
: > had mate in 1, so when DB ran out of checks, the game was over if
: > Kaspy
: > wasn't mated. But DB saw it *all*...
: Sure, but maybe you can help me out here and explain which move number,
: the wrong move, the right move and perhaps the why.
> I really have no way to find this out.
I do not understand this.
In the above you claim DB was by far superior than the micro's because
of a very deep combination of 20 plies. I show you the game and ask for
proof so I can check it (and learn something from DB!) and you say
I don't know!
Does anybody else remember?
: I think this is the game you are talking about:
: [Event "Man Vs. Machine 1996"]
: [Site "ACM round1"]
: [Date "1996.02.11"]
: [Round "1"]
: [White "DEEP BLUE"]
: [Black "Kasparov, G."]
: [Result "1-0"]
: 1. e4 c5 2. c3 d5 3. exd5 Qxd5 4. d4 Nf6 5. Nf3 Bg4 6. Be2 e6 7. h3 Bh5
: 8. O-O Nc6 9. Be3 cxd4 10. cxd4 Bb4 11. a3 Ba5 12. Nc3 Qd6 13. Nb5 Qe7
: 14. Ne5 Bxe2 15. Qxe2 O-O 16. Rac1 Rac8 17. Bg5 Bb6 18. Bxf6 gxf6 19.
: Nc4
: Rfd8 20. Nxb6 axb6 21. Rfd1 f5 22. Qe3 Qf6 23. d5 Rxd5 24. Rxd5 exd5 25.
: b3 Kh8 26. Qxb6 Rg8 27. Qc5 d4 28. Nd6 f4 29. Nxb7 Ne5 30. Qd5 f3 31. g3
: Nd3 32. Rc7 Re8 33. Nd6 Re1+ 34. Kh2 Nxf2 35. Nxf7+ Kg7 36. Ng5+ Kh6 37.
: Rxh7# 1-0
: I assume you mean 32. Rc7 is the critical move?
: And 32. Rc6 is the bad move concerning DB losing after 20 plies?
> Could be, yes. The thing is, after Rc7, later a knight move (a
> critical knight move) is a discovered check. With the rook on c6 it
> isn't... I think that Rc6 appears to win a pawn, or something, because
> most programs would stick the rook on the 7th no matter what. I'll run
> this thru crafty and post the results however...
Please do, I will also...
: 16.c4 Qh4 17.cxb5 Rh6 18.Nf6+ Kd8 19.h3 Bxh3 20.gxh3 Qxh3 21.Kg1 Qh2#
(12)
: 16.c4 Qh4 17.cxb5 Rh6 18.Nf6+ Kd8 19.h3 Bxh3 20.Be6+ 21.Kg1 Qh2# (12)
: 16.c4 Qh4 17.g3 Qh3 18.Qe2 fxg3 19.fxg3 Rgx3 (8)
: 16.c4 Qh4 17.g3 Qh3 18.Qf3 Rh6 19.Qg2 Qxg2 20.Kxg2 Bh3+ (10)
: If you find a deeper (better) defense for white I am all ears.
> I'd discount all mate lines. I can't imagine any program overlooking
> those except with severe bugs.... It's the other lines that would be
> iteresting to see what they saw...
Please show me "the other lines", as I said I am all ears.
Till now for me 16.c4?? is just 12 plies (see above variations)
If I am wrong I like to know.
- Ed Schroder -
Robert Hyatt
Ed Schroder (rebc...@xs4all.nl) wrote:
: From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: : Hitech was one of the favorites for the WC title (150,000 NPS)
: : Zugzwang was there (over 100,000 NPS)
: : There was a Cray computer (nodes 40,000 - 50,000)
: : But instead a micro searching just 2,000 NPS won in Madrid!
: : NPS are not holy...
: > they never have been. *BUT*... if you take two programs with
: > reasonable adequate evaluations, and reasonable adequate searches,
: > *THEN* NPS rules...
: NPS rules...???
: Was Hitech a bad program then?
: You must be kidding!
: Hitech from Hans Berliner was *by far* the main candidate for the WCM
: title in Madrid 1992!! (because of the absent of Deep Thought)
Not particularly. HiTech is not doing 150K nps any more, and has not in
a long time It was, at last account, below 100K. However, your math is
*way* off.
Let's say HiTech was rated 200 points better than any other program there.
Then the P(win) was .75... per round. .75^5 is not a big number since
.75^2 is only .5, so HiTech didn't have a good chance, even if it was
200 points higher rated than anyone else, which I don't believe it was.
Any one micro had a very small chance of winning the event, probably the
best was at the 10% probability level. But there were a *bunch* of
micros there, and the liklihood that one would win, with only one strong
"mainframe" there suddenly became quite high. The winner in 1992 was
basically chosen by a coin toss, not by playing skill.
: Hitech (150,000 NPS) lost from a 2,000 NPS program with just 512 Kb
: hash table. Also this 2,000 NPS program won the tournament with 4.5
: points out of 5.
: NPS rules...???
: NPS is important but it does not rule.
See the above explanation about two fairly even programs. Would you care
to challenge HiTech to a 10 game match with your 2K nps program? I suspect
HiTech's eval is better than Crafty's at present. Would you want to play
Crafty at 50:1 time odds? A Swiss tournament is one thing, and winning it
is not easy. But it certainly doesn't produce the best program with a small
number of rounds...
: Same tournament 3 years later, Deep Blue Prototype participated in
: Hong Kong. Same result, a micro won.
: Same NPS fight ended in favor for the micro.
Again, do you believe Fritz would win a match? The math for a Swiss is
difficult, because there are few rounds, and a luck win can move you to the
top without enough rounds to let an unluck loss be overcome.
: Deep Blue Prototype NPS 3,000,000
: Fritz NPS 100,000
: Madrid 1992 : 150,000 / 2000 = factor 75
: Hong Kong 1995 : 3,000,000 / 100,000 = factor 30
: NPS rules???
yep... just try crafty at 75:1 or 30:1. In a match, not a single
game...
: : > could be any of a dozen things. How about Rc6 in the Kasparov game
: : > where all the micros wanted to play (either Rc6 or Rc7 whichever DB
: : > *didn't* play, and all would have lost about 20 plies later, because
: : > the rook was a critical piece there, a little later. Remember Kaspy
: : > had mate in 1, so when DB ran out of checks, the game was over if
: : > Kaspy
: : > wasn't mated. But DB saw it *all*...
: : Sure, but maybe you can help me out here and explain which move number,
: : the wrong move, the right move and perhaps the why.
: > I really have no way to find this out.
: I do not understand this.
: In the above you claim DB was by far superior than the micro's because
: of a very deep combination of 20 plies. I show you the game and ask for
: proof so I can check it (and learn something from DB!) and you say
: I don't know!
: Does anybody else remember?
It's not that I don't remember, it's that I never knew what they saw. That
loss could have been any of a dozen things, from an actual bug, to a time
allocation bug after a restart, to a shallow search depth caused by search
extensions going out of control. I don't have any idea what happened, because
I wasn't there, but I've seen plenty go wrong in the past. You are making
much too much noise about a single game. I'd take deep thought in a match
any time, and so would most any logical-thinking person...
: : I think this is the game you are talking about:
: : [Event "Man Vs. Machine 1996"]
: : [Site "ACM round1"]
: : [Date "1996.02.11"]
: : [Round "1"]
: : [White "DEEP BLUE"]
: : [Black "Kasparov, G."]
: : [Result "1-0"]
: : 1. e4 c5 2. c3 d5 3. exd5 Qxd5 4. d4 Nf6 5. Nf3 Bg4 6. Be2 e6 7. h3 Bh5
: : 8. O-O Nc6 9. Be3 cxd4 10. cxd4 Bb4 11. a3 Ba5 12. Nc3 Qd6 13. Nb5 Qe7
: : 14. Ne5 Bxe2 15. Qxe2 O-O 16. Rac1 Rac8 17. Bg5 Bb6 18. Bxf6 gxf6 19.
: : Nc4
: : Rfd8 20. Nxb6 axb6 21. Rfd1 f5 22. Qe3 Qf6 23. d5 Rxd5 24. Rxd5 exd5 25.
: : b3 Kh8 26. Qxb6 Rg8 27. Qc5 d4 28. Nd6 f4 29. Nxb7 Ne5 30. Qd5 f3 31. g3
: : Nd3 32. Rc7 Re8 33. Nd6 Re1+ 34. Kh2 Nxf2 35. Nxf7+ Kg7 36. Ng5+ Kh6 37.
: : Rxh7# 1-0
: : I assume you mean 32. Rc7 is the critical move?
: : And 32. Rc6 is the bad move concerning DB losing after 20 plies?
: > Could be, yes. The thing is, after Rc7, later a knight move (a
: > critical knight move) is a discovered check. With the rook on c6 it
: > isn't... I think that Rc6 appears to win a pawn, or something, because
: > most programs would stick the rook on the 7th no matter what. I'll run
: > this thru crafty and post the results however...
: Please do, I will also...
Crafty liked Rc6 forever. I didn't give it forever to search, however, but
Rc6 (it thinks) wins a pawn. Obviously, if DB played Rc7, it saw something
better, but I don't have a clue as to what.
: : 16.c4 Qh4 17.cxb5 Rh6 18.Nf6+ Kd8 19.h3 Bxh3 20.gxh3 Qxh3 21.Kg1 Qh2#
: (12)
: : 16.c4 Qh4 17.cxb5 Rh6 18.Nf6+ Kd8 19.h3 Bxh3 20.Be6+ 21.Kg1 Qh2# (12)
: : 16.c4 Qh4 17.g3 Qh3 18.Qe2 fxg3 19.fxg3 Rgx3 (8)
: : 16.c4 Qh4 17.g3 Qh3 18.Qf3 Rh6 19.Qg2 Qxg2 20.Kxg2 Bh3+ (10)
: : If you find a deeper (better) defense for white I am all ears.
I didn't... Crafty things that c4 loses a pawn to Qh4, at depth 10/11
oor thereabouts. However, the position is tactical enough that I wouldn't
be surprised if there's some defense that appears to work for a while, only
to fail at deeper depth. Without DT analysis, there's no way to know what
it saw or thought it saw, and why...
: > I'd discount all mate lines. I can't imagine any program overlooking
brucemo
Robert Hyatt wrote:
> Any one micro had a very small chance of winning the event, probably the
> best was at the 10% probability level. But there were a *bunch* of
> micros there, and the liklihood that one would win, with only one strong
> "mainframe" there suddenly became quite high. The winner in 1992 was
> basically chosen by a coin toss, not by playing skill.
Your last sentence is true, but it is true of any of these tournaments,
and doesn't take anything away from the winners.
The strongest programs will tend to do better, but there are no
guarantees. All you can hope for is that your program is good enough to
give the coin a nudge in the right direction. If it is, you are to be
congratulated.
Fritz clearly was good enough in 1995, and so was CM Schroder was in 1992.
bruce
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: The winner in 1992 was basically chosen by a coin toss, not by
: playing skill.
Thanks, the best compliment I ever had...
Bob, I am a little bit confused because of our discussion below.
Bob Hyatt wrote:
"How about Rc6 in the Kasparov game where all the micros wanted to
play (either Rc6 or Rc7 whichever DB *didn't* play, and all would
have lost about 20 plies later, because the rook was a critical
piece there, a little later."
Ed Schroder anwsered:
"Sure, but maybe you can help me out here and explain which move
number, the wrong move, the right move and perhaps the why."
Bob Hyatt answered:
"I really have no way to find this out."
Ed Schroder asked again:
"I do not understand this.
In the above you claim DB was by far superior than the micro's because
of a very deep combination of 20 plies. I show you the game and ask for
proof so I can check it (and learn something from DB!) and you say
I don't know!
Bob Hyatt answered:
"It's not that I don't remember, it's that I never knew what they saw.
and
"Crafty liked Rc6 forever. I didn't give it forever to search,
however, but Rc6 (it thinks) wins a pawn. Obviously, if DB played
Rc7, it saw something better, but I don't have a clue as to what.
You made 2 claims:
1.. Deep Blue played 32. Rc7 because 32. Rc6 would lose after 20 plies.
2.. Deep Blue saw it, the micro's did not.
These are your words and not mine.
I asked you to come up with proof.
You say 2 things:
1.. I don't know;
2.. DB played 32. Rc7 but I don't have a clue.
This is quite different than your original claim that 32. Rc6 (the move
of the micro's) is a losing move after 20 plies, implying that DB saw it
all, the micro's did not, meaning DB is by far superior...
To me the facts are different, 32. Rc6 is *NOT* losing. I did not
see any 20 ply combination, but you claim it and then not proof it. I
expect at least to see the supposed losing main line for 32. Rc6
Where is the analysis that proves your 20 ply claim of DB?
In the light of this whole discussion which is about ply depth / NPS
versus playing strength I am a little bit pissed off about statements
of 20 plies without proof which I find misleading information is this
case since it's so essential to the whole discussion.
: You are making much too much noise about a single game.
So now a serious and importamt discussion is suddenly noise?
Compliment (2) taken, can I expect more?
I gave you the following info (proof) that 16. c4?? just needs
12 plies to see it's a bad move. Here is the analysis again:
16.c4 Qh4 17.cxb5 Rh6 18.Nf6+ Kd8 19.h3 Bxh3 20.gxh3 Qxh3 21.Kg1 Qh2# (12)
16.c4 Qh4 17.cxb5 Rh6 18.Nf6+ Kd8 19.h3 Bxh3 20.Be6+ 21.Kg1 Qh2# (12)
16.c4 Qh4 17.g3 Qh3 18.Qe2 fxg3 19.fxg3 Rgx3 (8)
16.c4 Qh4 17.g3 Qh3 18.Qf3 Rh6 19.Qg2 Qxg2 20.Kxg2 Bh3+ (10)
If you find a deeper (better) defense for white I am all ears.
You answered with: I didn't...
Come on Bob, DB prototype failed miserabely on a 12 ply deep variation.
Rebel, Hiarcs, Fritz, Genius all see it, DB prototype didn't and had
to give the queen for a rook 4 moves later (game over).
I find it useful if you or somebody else would post a set of positions
where DB has seen deep combinations so we can learn more by simply
check these positions with the available chess programs.
- Ed Schroder -
Robert Hyatt
brucemo (bru...@nwlink.com) wrote:
: Robert Hyatt wrote:
: > Any one micro had a very small chance of winning the event, probably the
: > best was at the 10% probability level. But there were a *bunch* of
: > micros there, and the liklihood that one would win, with only one strong
: > "mainframe" there suddenly became quite high. The winner in 1992 was
: > basically chosen by a coin toss, not by playing skill.
: Your last sentence is true, but it is true of any of these tournaments,
: and doesn't take anything away from the winners.
Of course not. because to take advantage of that 10% you still have to play
quite well...
: The strongest programs will tend to do better, but there are no
: guarantees. All you can hope for is that your program is good enough to
: give the coin a nudge in the right direction. If it is, you are to be
: congratulated.
: Fritz clearly was good enough in 1995, and so was CM Schroder was in 1992.
I've never dissented here. Cray Blitz won in 1983 when it was not clear it
was better than Belle. It also won in 1986 when I thought it was the best
program there. I saw Chess 4.6 lose in 1974, when it was clearly better,
but fell into the famous knight sacrifice against Chaos and was eliminated
from contention. No one was really close, but the coin came up tails...
Crafty might have been lucky to finish 4th at Jakarta. It also might have
been very unlucky to finish 4th. Fortunately, there are enough rounds that
the eventual winner is more likely to be the best, than at the WCCC which is
only 5 rounds, with as many as 28-30 programs. There it's a crap shoot...
In any case, in *every* ACM / WCCC event Cray Blitz won, I can show you at least
one game where the coin could have come up heads or tails, because there was an
element of chance in at least one game in every tourney. I had my fair share
of luck to be sure...
: bruce
Robert Hyatt
: : The winner in 1992 was basically chosen by a coin toss, not by
: : playing skill.
: Thanks, the best compliment I ever had...
Don't be insulted. I posted another discussion where I said exactly
the same thing about myself in 1983, 1986 WCCC's and in the ACM events
I won or finished 2nd in...
You were lucky to win in 1992, plain and simple, just as I was lucky
in 1983 and 1986, because the odds are so stacked against *anyone*
winning a 5 rounder than it is really amazing that anyone does. But
even a lottery has a winner of course...
Skill does play an important part of winning of course, but even if you
have a lot of skill, you also need a lot of luck in 5 rounds. I had
my share. Your number came up in 1992.
: Bob, I am a little bit confused because of our discussion below.
: and
Ed, I don't have the analysis. There was mucho discussion on ICC the day
that happened. We had GM's analyzing, computers analyzing, and so forth.
We were also getting feedback from the skittles room at the actual site.
I simply remember when all the micros including crafty wanted to play
Rc6, and there was a lot of discussion. Later, one of the GM's noticed
that in the game, Rc7 was critical, because a knight move was a discovered
check. He then did some analysis and pointed out that the Rc6 move really
was a lemon after the fact.
I don't know how to "prove" this because I certainly can't do an exhaustive
20+ ply search to "prove" that this move won and Rc6 lost. However, I'll
stand by my premise that DB's over 200 rating points better than the next closest
machine. When you, or I, or anyone else can show a win and draws against the
current world champion, we can then claim to join that elite group of "1" at
present. So far, most micros get blasted even at game/30 or game/60 against
the best. This is *slowly* changing... But DB did it in a real game, at a
real time control, with real money on the line. A couple of months ago, Crafty
scored 3.5/5.0 against 5 GM's in game/30. Do I believe it is that good. Hardly.
There were no prizes, there were no incentives, there were no reasons for the
GM's to study crafty (or the other programs) and find cooks to their favorite
openings... there were no reasons for the GM's to play their pet prepared line
since nothing was at stake. THe DB/Kasparov match is a different thing entirely,
because Kasparov's reputation *and* a bunch of money are on the line. DB's
playing against a different GM opponent than the rest of us, until we drum up
that kind of support for a reasonable prize fund. And *then* you, or I, or anyone
else will find out just how far behind the top GM's our programs are. Believe me,
it's a lot further than most are thinking... A *lot* further...
: I asked you to come up with proof.
: You say 2 things:
: 1.. I don't know;
: 2.. DB played 32. Rc7 but I don't have a clue.
: This is quite different than your original claim that 32. Rc6 (the move
: of the micro's) is a losing move after 20 plies, implying that DB saw it
: all, the micro's did not, meaning DB is by far superior...
No... but you asked for "proof". I didn't record the analysis, although I
did remember saying "aha"... knight moves is a disco check... wonder if that
was the reason...
: To me the facts are different, 32. Rc6 is *NOT* losing. I did not
: see any 20 ply combination, but you claim it and then not proof it. I
: expect at least to see the supposed losing main line for 32. Rc6
: Where is the analysis that proves your 20 ply claim of DB?
: In the light of this whole discussion which is about ply depth / NPS
: versus playing strength I am a little bit pissed off about statements
: of 20 plies without proof which I find misleading information is this
: case since it's so essential to the whole discussion.
If you really think you are in the same class with DB, so be it. We will
just have to agree to disagree. I hope, after the Kasparov match, they have
enough machine time to drop in on ICC or whatever and play a series of
matches against the top programs, just to put this to rest. As I mentioned
before, are you interested in playing Crafty at 50:1 time odds? I'd suspect
not. Yet that's a far cry from playing DB at 10000:1 time odds.
As I've also said, you might want to think that us "big iron" types only rely
on speed and nothing else. Crafty's a classic counter to that point of view.
It's not the best micro program around, but it's close to any of them. Close
enough that it can't be taken for granted. Close enough that no one can afford
to give it significant time odds. And yet you want to believe that Hsu and
company can't do the same, and I'll be danged if I can figure out the logic in
that. I'm an exception to the "big iron only wins because of speed." That's
enough of a warning that I might not be the *only* exception. I know Murray
quite well, Hsu pretty well, and the rest of the group less so, but there's not
a dummy in the group, which seems to be contrary to what you or others seem to
think.
: : You are making much too much noise about a single game.
: So now a serious and importamt discussion is suddenly noise?
: Compliment (2) taken, can I expect more?
No, it's not noise. Here's the *point*. If you look back over every
DT game played, and now at the DB games (although there are precious few of
them that I've seen, although there have been a few exhibitions like Byrne and
others last year), you draw one conclusion... the machine was strong 5 years
ago, it is stronger today. If you find that offensive, I'm sorry. I, too, wish
it weren't true, because I'd love to have a chance with a Micro against them,
but it's wishful thinking. Do you actually think rebel on a 386/16 has a chance
against rebel on a P6/200? Of course not. But then, in the same breath, you
discount the 10000x improvement over the P6/200 as "nothing"... The only con-
clusion I can draw is you think the DB guys are incompetent... there's no other
real option I can see. I think this is based on an uninformed opinion that has
no facts (nor even fiction) to support it.
If you want a real test, try one of two games... either the last *Socrates vs
DT game at an ACM event, or the last (I believe) Cray Blitz vs DT game at an
ACM. And try to see if Rebel (or any other program) can find the long path
that DT found against *Socrates and produce the same move... or see if you can
find the pawn push DT made, where they saw winning a bishop against CB. We
finally saw it about 5-8 full moves (maybe even 10, it was a while ago) *after*
they had failed high, and seen the big score. I'm not speculating, I was there.
I was on the receiving end of one of those. I know CB couldn't do it, and I know
that CB can search deeper than Rebel will ever hope to do (5 M nodes per sec
at present), and I know that Crafty is not ever going to see either, and the
funny thing is, it doesn't bother me. I'd love to have a Ferrari F5, or a 300+
horsepower Merc outboard for my boat, but I'm not willing (or able perhaps. :) )
to spend that kind of money. I'll take my $3000 P6/200 and have a ball with it,
and in 6 months I'll be using a multi-processor machine that is not a lot more
expensive and having fun there.
: I gave you the following info (proof) that 16. c4?? just needs
: 12 plies to see it's a bad move. Here is the analysis again:
: 16.c4 Qh4 17.cxb5 Rh6 18.Nf6+ Kd8 19.h3 Bxh3 20.gxh3 Qxh3 21.Kg1 Qh2# (12)
: 16.c4 Qh4 17.cxb5 Rh6 18.Nf6+ Kd8 19.h3 Bxh3 20.Be6+ 21.Kg1 Qh2# (12)
: 16.c4 Qh4 17.g3 Qh3 18.Qe2 fxg3 19.fxg3 Rgx3 (8)
: 16.c4 Qh4 17.g3 Qh3 18.Qf3 Rh6 19.Qg2 Qxg2 20.Kxg2 Bh3+ (10)
: If you find a deeper (better) defense for white I am all ears.
: You answered with: I didn't...
: Come on Bob, DB prototype failed miserabely on a 12 ply deep variation.
: Rebel, Hiarcs, Fritz, Genius all see it, DB prototype didn't and had
: to give the queen for a rook 4 moves later (game over).
: I find it useful if you or somebody else would post a set of positions
: where DB has seen deep combinations so we can learn more by simply
: check these positions with the available chess programs.
: - Ed Schroder -
If you can find the two games above, we can discuss them. I'll check my files
at the office, and I'll prompt Hsu to see if he remembers when the CB game was
played, but I do believe it was the last time we met, whenever that was. I
can honestly say I've *never* seen CB out-calculated that badly. If you'd like
to compare notes, 5 years ago it solved 298/300 win at chess in < 1 second each.
The other two were solved in (I believe 3 and 12 seconds each.) On the T90, that
would put *every* one at < 1 second. And CB is *nowhere* near capable of the
calculation prowess of DB. In contrast, Crafty solves 297/300 in one minute or
less, with the *right* score. it solves 298/300 if you count one position where
it plays the right move, but does not see the winning variation. It's quick, but
it's not a CB, and it is absolutely not a DB.
I wish Hsu would jump in and report some of his Nolot solutions, because it is
eye-popping... but that is his data to report, not mine. One day, you'll see
just how silly all this really is. I just hope they can get access to the SP
after the Kasparov match, and play a series of 10 game matches against the best
micros, on a server for all to see. That would quickly and clearly solve the
issue once and for all...
Richard Bean
Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
: : To me the facts are different, 32. Rc6 is *NOT* losing. I did not
: : see any 20 ply combination, but you claim it and then not proof it. I
: : expect at least to see the supposed losing main line for 32. Rc6
I don't think Rc6 was losing... also take a look at fhh's posting
http://xp6.dejanews.com/getdoc.xp?recnum=532256&server=dnserver.db96q2&
CONTEXT=856148095.10295&hitnum=0
All fhh says is: "there is a flaw with rc6. any guess to what black's
reply should be? Hint: counter play". His tone was similar to his posts
after the Polgar-Kasparov "he took his hand off the knight" game,
where he was saying something like "Well, no micro could ever see that Qh4
draws in several months, but DT2 does" and then not ultimately posting any
exhaustive analysis.
: No, it's not noise. Here's the *point*. If you look back over every
: DT game played, and now at the DB games (although there are precious few of
: them that I've seen, although there have been a few exhibitions like Byrne
: and others last year),
I thought fhh made it clear that DB only existed in its match configuration
3 weeks before the Kasparov match, that they only fixed all the bugs
the night before the first game (!) and that it hasn't played any
games since. Do an altavista search on "Deep Blue Junior"; that's played
a few exhibition games in '96 (search on "kuan-kuan tian"?) but it's only a
toy compared to Deep Blue. (Robert only made it clear a few posts ago that
the HK game was Deep Blue software running on Deep Thought II hardware,
so it can't be used to get an indication of DB's strength).
: If you want a real test, try one of two games... either the last *Socrates
: vs DT game at an ACM event, or the last (I believe) Cray Blitz vs DT game
: at an ACM. And try to see if Rebel (or any other program) can find the
: long path that DT found against *Socrates and produce the same move...
: or see if you can
: find the pawn push DT made, where they saw winning a bishop against CB. We
: finally saw it about 5-8 full moves (maybe even 10, it was a while ago)
: *after*
: they had failed high, and seen the big score. I'm not speculating, I
: was there. I was on the receiving end of one of those.
Cray Blitz vs DT ftp://ftp.pitt.edu/group/chess/TECL/91acomcom.zip
or ftp://ftp.pitt.edu/group/chess/PGN/Events/1992/92acm-pg.zip
(which?)
Star Socrates, as fhh pointed out, got more nps than DT2 but DT2 wasted it
in a q-side attack in a sicilian. So singular extensions beat nps :-)
http://www.ie.cuhk.hk/~chess95 (round 1) for one example (DT2 was white)
but the one Robert is thinking of is
http://theory.lcs.mit.edu/~cilk/sample_games.ascii ... this is awesome.
: I wish Hsu would jump in and report some of his Nolot solutions, because
: it is eye-popping... but that is his data to report, not mine. One day,
: you'll see just how silly all this really is. I just hope they can get
: access to the SP after the Kasparov match, and play a series of 10 game
: matches against the best
: micros, on a server for all to see. That would quickly and clearly solve
: the issue once and for all...
http://xp6.dejanews.com/getdoc.xp?recnum=%3c32AEF...@nwlink.com%3e
&server=dnserver.db96q5&CONTEXT=856147858.8860&hitnum=0
is the Nolot suite on DT, reposted by brucemo, reposted by Marc-Francois
Baudot (Virtual Chess), posted by fhh.
BTW I agree with Robert about the 11-12 ply estimate for DT1 after seeing the
"Chip vs the Chess Master" film. (~3M nps). It'd be exciting to see
some of the search information for the DB-Kasparov match, but I suppose
that's all kept under lock and key :-/.
Richard.
Robert Hyatt
Ed:
Here's the game I mentioned between Cray Blitz and Deep Thought II. I
believe (if my file naming convention is correct) that this was played in
1991. I think in Orlando, but don't remember exactly.
In any case, when blacked played 27. ... c5, they did so with a +2 eval
after a fail high. More details are below, but it was about 26 *plies*
later that CB finally reached -2. Notice that... 26 *plies*. It is
certainly ugly beyond move 27, but it does illustrate just how powerful that
machine is. You can draw your own conclusions of course...
[Event "ACM 1991"]
[Site ""]
[Date ""]
[Round ""]
[White "Cray Blitz"]
[WhiteElo ""]
[Black "Deep Thought II"]
[BlackElo ""]
[Result "0-1"]
1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 Nc6 6. f4 e5
7. Nxc6 bxc6 8. fxe5 Ng4 9. Be2 Nxe5 10. Be3 Be7 11. o-o Be6
12. Qd4 o-o 13. Rad1 f6 14. b3 Qe8 15. Na4 Qg6 16. Bf4 Rf7 17. Qe3
Raf8 18. Qxa7 Qxe4 19. Bd3 Qb4 20. Qe3 Ra8 21. c3 Qb7 22. Rf2
Qa7 23. Qxa7 Rxa7 24. Be3 Ra5 25. Bb6 Ra8 26. Bc2 Bf8 27. Re1
c5 28. Be4 Ra6 29. Rb1 f5 30. Bc2 Rb7 31. Bd8 g6 32. Re1 c4
33. Rb1 Bd7 34. Nb2 Ra8 35. Bg5 Rxa2 36. b4 Bb5 37. Re2 Bg7
38. Nd1 Ra6 39. Bd2 Nd3 40. Ne3 Ra2 41. Bxd3 cxd3 42. Rf2 Rxd2
43. Rxd2 Bxc3 44. Nf1 Bxd2 45. Nxd2 Re7 46. Nf3 h6 47. Rb2 Re4
48. Kf2 g5 49. g3 f4 50. gxf4 Rxf4 51. Kg3 h5 52. Nd2 h4+ 53. Kg2
Bc6+ 54. Kg1 Rg4+ 55. Kf2 Rg2+ 56. Ke3 Bb5 57. Ra2 Rxh2 58. Ra5
Re2+ 59. Kd4 h3 60. Rxb5 Rg2 61. Rb8+ Kg7 62. Rb7+ Kg6 63. Rd7
Here's the results I've reconstructed from going thru the log from Cray
Blitz, with a couple of comments I entered as the game was in progress.
When Black (Deep Thought II) played 27. c5, they had failed high, with a score
> 2.000 according to my notes. They couldn't tell what was happening, because
part of their search is in software, but part is in hardware, and the hardware
part does not "back up" a PV, just returns the score and best move for a given
position it searches. As a result, they saw the > +2, but didn't see anything
captured in their PV as Hsu, Murray, Harry, Bert and I looked at their output.
If you look, the bishop is obviously in trouble to a human's eye, but there are
*lots* of delaying moves before thing get serious... such as the next white
move which attacks the black rook on a8 and burns two plies of search...
After making move 33. Rb1, Cray Blitz's score was still -.092, nearly zero.
While pondering, it failed low, and saw itself losing a pawn, which was likely
a horizon effect move to postpone the bishop problem a ways further... By move
43, we were down to -1.9, which agreed approximately with what they had seen
when they played c5... However, this is *one hell* of a long way from where
they saw this. Hsu and Murray thought it was just an example of where the
singular extension algorithm happened to work well. I'd certainly agree...
I didn't make any further notes, so this is really all I have to go on. I
personally saw the +2 from them, and 10+ moves later I saw the -2 for us.
They did *not* have huge positional scores back then, neither did CB, so this
was material, plain and simple.
I let Crafty run on the position after c5 and it sees *nothing* bad for white at
all. This may give you some insight into what that machine is capable of doing.
This version of Cray Blitz was doing about 250K nodes per second, on a Cray YMP
with 8 processors, for comparison. It used null-move, R=1, non-recursive, but
did have a lot of search extensions. It was typically searching 10-11 plies
deep each move, although it varies as most programs do.
When I saw the c5 move, I immediately got interested, because the bishop suddenly
has very few places it can stay safely, depending on a knight to hold it for a
while. It appears to be a very deep forced line, however, that just happens to
play right into their strong suit. This is somewhat like the *Socrates game that
I mentioned, because they were saying they were winning, *socrates thought it
was even, but at the end of the variation, there was a "stinger" that was being
overlooked by everyone but Deep Thought... I'll try to find that position as
well for discussion, but after looking at this one, you might be sick enough
that it's not needed. I know I was. :)
Tord Kallqvist Romstad
Ed Schroder (rebc...@xs4all.nl) wrote:
: I find it useful if you or somebody else would post a set of positions
: where DB has seen deep combinations so we can learn more by simply
: check these positions with the available chess programs.
: - Ed Schroder -
Hsu posted the results on the Nolot positions for DT2 some time ago:
#1: 6 hours.
#2: 2 minutes
#3: Not found within one hour
#4: 2.5 hours
#5: 2 hours
#6: Not found
#7: 6 hours
#8: Bxh7 is not found, but Hsu believes that Bxh7 is not the best move.
DT2 plays c4, which threaten Bxh7
#9: 9 minutes
#10: Less than 2 minutes
#11: 5 minutes
Note that this is DT2 --- DB should solve most of these problems at tournament
time control!
Tord
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
[ snip ]
: Ed, I don't have the analysis. There was mucho discussion on ICC
: the day that happened. We had GM's analyzing, computers analyzing,
: and so forth. We were also getting feedback from the skittles room at
: the actual site.
Ok, no 20 plies combination for DB then on 32. Rc6
A pity since it would have given me more information about DB.
I am not against DB, I want info to build my own opinion and to judge
how strong the thing really is.
: However, I'll stand by my premise that DB's over 200 rating points
: better than the next closest machine.
200 ???
50 not more!
: If you really think you are in the same class with DB, so be it.
I never said the current Pc programs are *better* than DB. See above
statement, you say 200 ELO, I say 50 ELO which "touches the heart" of
our discussion!
: As I mentioned before, are you interested in playing Crafty at 50:1
: time odds? I'd suspect not.
Yes I am interested as long as Rebel may play on tournament level!
Yes I take your challenge, Rebel will win a ten game match... :)
KK, were are you?
We need you!
I will make it even more interesting, the latest version of Crafty
against the latest version of Rebel on a PP-200. Rebel at tournament
level and Crafty 100 x more time, so 100:1 instead of 50.
The KK NPS Kup...
NPS 70,000 against NPS 7,000,000, I love your idea!
NPS rules... your chance to prove it!
[ snip ]
: The only conclusion I can draw is you think the DB guys are
: incompetent...
Hey Bob don't put words in my mouth I never said!!
We have a discussion about ply depth / NPS versus playing strength, you
state "NPS rules..." which is speculative, I only counter attack all
your "maybe", "most likely", "probably" with the naked truth, the FACTS.
The current facts are:
- Madrid 1992, a 2,000 NPS micro won against 150,000 NPS;
- Hong Kong 1995, a 100,000 NPS micro won against 3,000,000 NPS;
- 16.c4??, just a 12 ply tactics.
I don't say things I can't prove, fact is you did with 32.Rc6
I don't fight Hsu, I fight against YOUR CLAIMS!
Quite a difference.
[ snip ]
- Ed Schroder -
Robert Hyatt
Harald Faber (Harald...@p21.f2.n1.z1001.fidonet.org) wrote:
: ## Nachricht vom : 14.02.97 weitergeleitet
: Tschuessikowski
: Harald
: --
Yes and no. Remember they are doing a parallel search. they could potentially
search faster, but they don't. They claimed a max of 200M nodes per second
for 256 processors, with an average of around 100M. they should now be at a
max of 800M, plus they have had a year to tune the parallel search, which I'd
suspect will get them close to 1B nodes per sec, maybe even a sustained speed
that's that fast...
and remember, speed is not *all* they are doing.
Robert Hyatt
Ed Schroder (rebc...@xs4all.nl) wrote:
: From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: [ snip ]
: : Ed, I don't have the analysis. There was mucho discussion on ICC
: : the day that happened. We had GM's analyzing, computers analyzing,
: : and so forth. We were also getting feedback from the skittles room at
: : the actual site.
: Ok, no 20 plies combination for DB then on 32. Rc6
: A pity since it would have given me more information about DB.
: I am not against DB, I want info to build my own opinion and to judge
: how strong the thing really is.
: : However, I'll stand by my premise that DB's over 200 rating points
: : better than the next closest machine.
: 200 ???
: 50 not more!
Utter and complete hogwash. I only hope we end up with a chance for
you to see just how wrong you are... If you believe 50 is correct, that
you would win 40% of the games against them, hmmm....
: : If you really think you are in the same class with DB, so be it.
: I never said the current Pc programs are *better* than DB. See above
: statement, you say 200 ELO, I say 50 ELO which "touches the heart" of
: our discussion!
: : As I mentioned before, are you interested in playing Crafty at 50:1
: : time odds? I'd suspect not.
: Yes I am interested as long as Rebel may play on tournament level!
: Yes I take your challenge, Rebel will win a ten game match... :)
You hve the machines. However, back the trolly up. YOu were claiming
that 2K nodes per second was better than them in 1992. Why do you suddenly
want to go 50x faster? IE, give Crafty 250 seconds per move, and Rebel
5, and that should be a close approximation to what was happening in 1992.
However, I'll take the odds at 3 mins vs 150 mins. Of course, you do know
you have to not think on opponent's time... If you can win at that handicap,
I'd likely go back to writing compilers... :)
: KK, were are you?
: We need you!
: I will make it even more interesting, the latest version of Crafty
: against the latest version of Rebel on a PP-200. Rebel at tournament
: level and Crafty 100 x more time, so 100:1 instead of 50.
: The KK NPS Kup...
: NPS 70,000 against NPS 7,000,000, I love your idea!
: NPS rules... your chance to prove it!
Somewhat broken test, of course... unless someone has a P6/200 with something
like 64 gigs if memory, otherwise hashing is going to effectively turned off
in Crafty. However, I'd still take the odds... You should talk to Lonnie
first, however. They are playing reasonably closely at equal time controls in
the games I've seen on ICC and Chess.net...
: [ snip ]
: : The only conclusion I can draw is you think the DB guys are
: : incompetent...
: Hey Bob don't put words in my mouth I never said!!
but it's clearly implied. You want to play Crafty, but you want to use
the full factor of 50 we've gotten in the past 5 years. And you claim it
is important. And then you claim that the much larger performance gains
made by the DB guys is worth only 50 rating points. The only conclusion
I can see is that they must be incompetent, if you can get so much from a
factor of 50, but they get nothing from an additional factor of 2,000...
Right?
: We have a discussion about ply depth / NPS versus playing strength, you
: state "NPS rules..." which is speculative, I only counter attack all
: your "maybe", "most likely", "probably" with the naked truth, the FACTS.
No, reread what I wrote. I said that for two programs with an adequate
search, and an adequate evaluation, that NPS *does* rule. Otherwise, why
are you running on a P6/200 instead of a 386/16? Our major point of
disagreement is the following:
I claim their search is *better* than yours. That it makes fewer tactical
errors than yours (note that you can take "yours" and replace with "mine" and
it doesn't change my opinion here.) I claim their evaluation is better than
yours, because I know what you can do, and run at the speed you are running
at, while their hardware eval can do as much as they care to without their
worrying about slowing things down at all...
You've done well with Rebel, and it plays well, but if you *really* believe
you would win 4 out of every 10 games against them, your judgement is really
clouded. Perhaps by not understanding what they are doing, but I can read
your mind.
However, see my post from last night. Let's see Rebel make any progress on
that position, because that was solved by DT II in 1991, when you are lucky
to do 1K nodes per second.
: The current facts are:
: - Madrid 1992, a 2,000 NPS micro won against 150,000 NPS;
one game remember, just one game. Cray Blitz lost to "Bobby" in Cologne
Germany due to a bug I introduced. It won the rest and the tourament, but
*one game* doesn't mean a lot. Try 'em in a match if we can ever get Hsu
to play on ICC or whatever.
: - Hong Kong 1995, a 100,000 NPS micro won against 3,000,000 NPS;
: - 16.c4??, just a 12 ply tactics.
: I don't say things I can't prove, fact is you did with 32.Rc6
: I don't fight Hsu, I fight against YOUR CLAIMS!
: Quite a difference.
But you fight with feathers. You claim 50 points. Can you *prove* that?
of course not. Relevant questions: (1) when was the last time you beat
Kasparov in a reasonable time control? (2) when was the last time you drew
Kasparov in a reasonable time control? (3) can you show an even record against
GM's in any sort of standard time control to come anywhere near justifying that
you are within 50 points of them? (4) have you ever beaten them OTB yourself?
(5) do you have any test suites that suggest you are close? Nolot? Can you
approach what DT did?
And with 5 out of 5 "no" answers, you would still claim to be "right up there
with them." That's difficult for me, a scientist, to fathom. I've been able
to observe them, I've been able to observe Rebel, Crafty, Mchess Pro, Hiarcs,
and the rest. and in my opinion, based on observed facts, none are close.
: [ snip ]
: - Ed Schroder -
Robert Hyatt
Tord Kallqvist Romstad (tor...@ifi.uio.no) wrote:
: : - Ed Schroder -
: Tord
Also notice that Hsu doesn't consider a position solved *just* because DT
would play the right move. He evaluates just as I do, so that if it doesn't
have a score to prove that it understands that a move is best, it's not right.
The is as opposed to many that claim success just because they get the right
move, period...
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: 200 ???
: 50 not more!
> Utter and complete hogwash. I only hope we end up with a chance for
> you to see just how wrong you are... If you believe 50 is correct,
> that you would win 40% of the games against them, hmmm....
In previous postings I said: DB-PC 6.5 - 3.5
You answered with: DB-PC 9 - 1
And for the moment I like to stick to my opinion.
: : As I mentioned before, are you interested in playing Crafty at 50:1
: : time odds? I'd suspect not.
: Yes I am interested as long as Rebel may play on tournament level!
: Yes I take your challenge, Rebel will win a ten game match... :)
> You hve the machines. However, back the trolly up. YOu were claiming
> that 2K nodes per second was better than them in 1992. Why do you
> suddenly want to go 50x faster? IE, give Crafty 250 seconds per move,
> and Rebel 5, and that should be a close approximation to what was
> happening in 1992. However, I'll take the odds at 3 mins vs 150 mins.
> Of course, you do know you have to not think on opponent's time...
> If you can win at that handicap, I'd likely go back to writing
> compilers... :)
Very funny, but I won't let you go away that easy... :)
The discussion is of course (and has always been!) how well will the best
PC programs perform against DB *NOW*. In this light we were always
talking about tournament level.
Right?
Now you suddenly introduce an average time of 5 seconds??
: Somewhat broken test, of course... unless someone has a P6/200 with
: something like 64 gigs if memory, otherwise hashing is going to
: effectively turned off in Crafty. However, I'd still take the odds...
: You should talk to Lonnie first, however. They are playing reasonably
: closely at equal time controls in the games I've seen on ICC and
: Chess.net...
Well, you challenged me 2 times, the first time I ignored your challenge,
the second time I accepted. I even proposed to you a handicap factor of
100 instead of your suggested factor of 50.
And now you complain about hash table size?
Is a factor 100 not enough for you?
What do you want then?
A 150, 200 factor??
Or do you simply don't want to play at all?
It is your chance to prove your "NPS rules..." statement.
: : The only conclusion I can draw is you think the DB guys are
: : incompetent...
: Hey Bob don't put words in my mouth I never said!!
> but it's clearly implied. You want to play Crafty, but you want
> to use the full factor of 50 we've gotten in the past 5 years. And you
> claim it is important. And then you claim that the much larger
> performance gains made by the DB guys is worth only 50 rating points.
> The only conclusion I can see is that they must be incompetent, if you
> can get so much from a factor of 50, but they get nothing from an
> additional factor of 2,000...
> Right?
Bob this are your words and your conclusions, not mine. The deep blue
team have my highest admiration for what they are doing now and have
done in the past. As I said before I am not fighting Hsu but your
unfounded claims, did you missed this in my previous posting? It was
there in UPPERCASE... :)
So how about the 10 game match?
Rebel at tournament level, Crafty 100 x more time, all on a PP-200.
You are right about playing without permanent brain. To compensate this
for both I suggest the following schedule:
Rebel 4:30 fixed time per move.
Crafty 4:30 x 100 = 7.5 hours per move (nice overnight solution)
How about it Bob?
Remember you challenged me, I accepted...
- Ed Schroder -
Robert Hyatt
: : Yes I take your challenge, Rebel will win a ten game match... :)
: You hve the machines. However, back the trolly up. YOu were claiming
: that 2K nodes per second was better than them in 1992. Why do you suddenly
: want to go 50x faster? IE, give Crafty 250 seconds per move, and Rebel
: 5, and that should be a close approximation to what was happening in 1992.
: However, I'll take the odds at 3 mins vs 150 mins. Of course, you do know
: you have to not think on opponent's time... If you can win at that handicap,
: I'd likely go back to writing compilers... :)
I wasn't thinking so clearly above, because there's an obvious problem
in playing a match at such a long time control. If we take your
"tournament mode" which I assume means 40 moves in 2 hours, then at
100:1, crafty would be playing 40 moves in 200 hours." Since there's
not an easy way to "autoplay" this, we'd be lucky to get in two moves
per day or so, which turns into a game per month if we are lucky.
There's two problems:
1. by the time a 10 game match is over, Rebel 9 and Crafty v12 will be
out. And we could then argue over the old program results... :)
2. 10 games is probably not enough to draw any real conclusions from,
we really need maybe 100 games. That's not going to happen unless we
can find a bunch of people with P6/200's lying around that want to do
this, which I doubt...
ERGO, what about the following: a much faster game, say 5:1 time odds
with crafty getting 10 seconds, Rebel 2? If you believe your knowledge
is much superior to mine, that is a perfectly valid test. In fact,
knowledge-based programs should do better at fast time controls than at
longer ones, because the effect of knowledge is more pronounced when you
can't search, and make moves based only on knowledge. Try playing a
rank beginner and not thinking at all, just follow normal practices such
as controlling the center, watching pawn advances, piece placement and
the like.
At least at that time control, we can see the results of a 100 game
match before we are old and retired... :)
However, I hope you have a chance to look over the Cray Blitz vs Deep
Thought II game I found and posted last night. Fortunately I kept the
log files from Cray Blitz for that event, so I can see exactly what it
was seeing, and I manually inserted the note when DT failed high, because
I thought it was something we should see. I didn't realize when I put
the note in just how deep the "plot" really was...
However, if you look at the Deep Thought Nolot test results, and then
extrapolate with Deep Blue being maybe 300x faster, and then being
pessimistic and scaling this back to 100x, you get a frightening
result. How many do you solve? Last time I tried Crafty it was not
doing well. It certainly didn't solve all of them inside of 5 minutes
each... And then if you look at the game mentioned above, and then at
the game vs *Socrates, you begin to get an inkling of just how deep that
thing searches.
ANd then if you look at some of the Kasparov games. One in particular,
the Bxh7 game... I've heard, indirectly, that if you ask Kasparov about
that position you get an alarming answer. namely that he saw that Bxh7
didn't win, but even more interesting is that a3 *does* outright win.
Rumor has it that DB saw this as well. Let's find a micro that can figure
out just how strong a3 is and why (hint: some *very* deep threats that a3
prevents.) The conclusion is still quite obvious...
Crafty, Rebel, Mchess Pro, Genius, and the rest can fight it out for the
best micro, but DB's playing a different game entirely...
Moritz Berger
On 17 Feb 1997 04:18:01 GMT, hy...@crafty.cis.uab.edu (Robert Hyatt)
wrote:
>Ed:
>
>Here's the game I mentioned between Cray Blitz and Deep Thought II. I
>believe (if my file naming convention is correct) that this was played in
>1991. I think in Orlando, but don't remember exactly.
>
>In any case, when blacked played 27. ... c5, they did so with a +2 eval
>after a fail high. More details are below, but it was about 26 *plies*
>later that CB finally reached -2. Notice that... 26 *plies*. It is
>certainly ugly beyond move 27, but it does illustrate just how powerful that
>machine is. You can draw your own conclusions of course...
Chessmaster 5000 play 27. .. c5 after about 2 or 3 minutes on my
P5/166 ... It needs about 2 or 3 million nodes to see it. Of course,
this is not enough to justify a +2 eval (too much speculation or
intelligence involved, too less exact "knowledge", therefore only a
meagre 0.14 eval ...), but my point is, that micros can play the best
moves without calculating zillions of nps. Point taken?
BTW: Yes, I knew that Chessmaster does about the best mobility
evaluation you will find on a micro, so I knew with which program to
test this position ;-) Together with its hellish extensions, this is
enough to play 27. .. c5 even on a P90 at regular time controls ...
Moritz
-------------
Moritz...@msn.com
John Michael Sauder
> In the meanwhile what about Crafty on a SGI Origin 2000 R10000 at
> 200 MHz
>(faster than an Alpha 500 MHz on a bench)?
>Franz
Has anyone ported Crafty to SGI? I've had problems getting
it to compile on a 150 MHz Indy R4400. I'd love to try it out on
our new 150 MHz R10000 O2, but we don't have a C compiler on it...
If anyone has compliled it on an SGI, please e-mail me!
--
-- Mike S. (M_Sa...@fccc.edu)
http://www.fccc.edu/research/labs/roder/mike/
Don Fong
In article <5e9l02$e...@juniper.cis.uab.edu>,
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote:
[to Ed Schroder]
>: : The only conclusion I can draw is you think the DB guys are
>: : incompetent...
>
>: Hey Bob don't put words in my mouth I never said!!
>
>but it's clearly implied.
no, Ed Schroder is dead right here. time and again i have seen
you play this game. you escalate the hostilities by drawing some extreme
hypothetical conclusion based on YOUR beliefs plus THE OTHER PERSON's
facts, then you try to impose this unviable hybrid on the other person.
the focus of the debate then shifts from being an illuminating battle
between two people's ideas, to a meaningless argument over your
made-up fallacies.
Ed Schroder did NOT say the DB team was incompetent. it is
NOT clearly implied by anything he said. it may be implied by YOUR
beliefs, but not HIS.
it's like when someone disagrees with you and you say ``ARE YOU
CALLING ME A LIAR?'' it is the tactic of a schoolyard bully.
mclane
Ed Schroder <rebc...@xs4all.nl> wrote:
>From: hy...@crafty.cis.uab.edu (Robert Hyatt)
>: The winner in 1992 was basically chosen by a coin toss, not by
>: playing skill.
>Thanks, the best compliment I ever had...
>Bob, I am a little bit confused because of our discussion below.
>Bob Hyatt wrote:
> "How about Rc6 in the Kasparov game where all the micros wanted to
> play (either Rc6 or Rc7 whichever DB *didn't* play, and all would
> have lost about 20 plies later, because the rook was a critical
> piece there, a little later."
>Ed Schroder anwsered:
> "Sure, but maybe you can help me out here and explain which move
> number, the wrong move, the right move and perhaps the why."
>Bob Hyatt answered:
> "I really have no way to find this out."
>Ed Schroder asked again:
> "I do not understand this.
> In the above you claim DB was by far superior than the micro's because
> of a very deep combination of 20 plies. I show you the game and ask for
> proof so I can check it (and learn something from DB!) and you say
> I don't know!
>Bob Hyatt answered:
> "It's not that I don't remember, it's that I never knew what they saw.
>and
> "Crafty liked Rc6 forever. I didn't give it forever to search,
> however, but Rc6 (it thinks) wins a pawn. Obviously, if DB played
> Rc7, it saw something better, but I don't have a clue as to what.
>You made 2 claims:
>1.. Deep Blue played 32. Rc7 because 32. Rc6 would lose after 20 plies.
>2.. Deep Blue saw it, the micro's did not.
>These are your words and not mine.
>I asked you to come up with proof.
>You say 2 things:
>1.. I don't know;
>2.. DB played 32. Rc7 but I don't have a clue.
>This is quite different than your original claim that 32. Rc6 (the move
>of the micro's) is a losing move after 20 plies, implying that DB saw it
>all, the micro's did not, meaning DB is by far superior...
>To me the facts are different, 32. Rc6 is *NOT* losing. I did not
>see any 20 ply combination, but you claim it and then not proof it. I
>expect at least to see the supposed losing main line for 32. Rc6
>Where is the analysis that proves your 20 ply claim of DB?
>In the light of this whole discussion which is about ply depth / NPS
>versus playing strength I am a little bit pissed off about statements
>of 20 plies without proof which I find misleading information is this
>case since it's so essential to the whole discussion.
>: You are making much too much noise about a single game.
It will be difficult to show evidence that Deep Blue or any other big
iron is superiour because this IS NOT TRUE.
Therefore people involved in this business can only claim but not show
in CHESS-LINES even because they don't play chess or they claim
without knowing much about.
>So now a serious and importamt discussion is suddenly noise?
>Compliment (2) taken, can I expect more?
>I gave you the following info (proof) that 16. c4?? just needs
>12 plies to see it's a bad move. Here is the analysis again:
>16.c4 Qh4 17.cxb5 Rh6 18.Nf6+ Kd8 19.h3 Bxh3 20.gxh3 Qxh3 21.Kg1 Qh2# (12)
>16.c4 Qh4 17.cxb5 Rh6 18.Nf6+ Kd8 19.h3 Bxh3 20.Be6+ 21.Kg1 Qh2# (12)
>16.c4 Qh4 17.g3 Qh3 18.Qe2 fxg3 19.fxg3 Rgx3 (8)
>16.c4 Qh4 17.g3 Qh3 18.Qf3 Rh6 19.Qg2 Qxg2 20.Kxg2 Bh3+ (10)
>If you find a deeper (better) defense for white I am all ears.
>You answered with: I didn't...
>Come on Bob, DB prototype failed miserabely on a 12 ply deep variation.
>Rebel, Hiarcs, Fritz, Genius all see it, DB prototype didn't and had
>to give the queen for a rook 4 moves later (game over).
>I find it useful if you or somebody else would post a set of positions
>where DB has seen deep combinations so we can learn more by simply
>check these positions with the available chess programs.
They cannot show us that DB is strong, DB is only strong because of
the hardware difference. It is not superiour. IMO it plays weaker than
other programs and only the hardware difference hides this.
But sometimes games against micros show that this is the truth.
>- Ed Schroder -
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: I wasn't thinking so clearly above, because there's an obvious problem
: in playing a match at such a long time control. If we take your
: "tournament mode" which I assume means 40 moves in 2 hours, then at
: 100:1, crafty would be playing 40 moves in 200 hours." Since there's
: not an easy way to "autoplay" this, we'd be lucky to get in two moves
: per day or so, which turns into a game per month if we are lucky.
Yes, unfortunately you are right...
About Crafty...
Is Crafty now fully compatible with AUTO232?
If so we can also keep the permament brain and set Rebel on 40/2:00
and Crafty on 40 in 200 hours. In this case we will have 4-5 moves
a day, 2.5 games a month.
: There's two problems:
: 1. by the time a 10 game match is over, Rebel 9 and Crafty v12 will be
: out. And we could then argue over the old program results... :)
I disagree here. It's not only the score that will count but also it's
my curiosity how well 70,000 NPS will perform against 7,000,000 NPS. I
am very interested in this! Aren't you? And of course I like to
win also... :)
: 2. 10 games is probably not enough to draw any real conclusions from,
: we really need maybe 100 games. That's not going to happen unless we
: can find a bunch of people with P6/200's lying around that want to do
: this, which I doubt...
That's an old discussion and I agree. But 10 is better than ZERO and
it will tell us a lot about our NPS argument.
: ERGO, what about the following: a much faster game, say 5:1 time odds
: with crafty getting 10 seconds, Rebel 2? If you believe your knowledge
: is much superior to mine, that is a perfectly valid test. In fact,
: knowledge-based programs should do better at fast time controls than at
: longer ones, because the effect of knowledge is more pronounced when you
: can't search, and make moves based only on knowledge. Try playing a
: rank beginner and not thinking at all, just follow normal practices such
: as controlling the center, watching pawn advances, piece placement and
: the like.
I am not interested in a 0:10 / 0:02 match.
I don't like your idea because I am only interested in tournament level
play. Fritz, Genius will win at blitz-level against Rebel but on
tournament level... You know what I mean.
: However, I hope you have a chance to look over the Cray Blitz vs Deep
: Thought II game I found and posted last night. Fortunately I kept the
: log files from Cray Blitz for that event, so I can see exactly what it
: was seeing, and I manually inserted the note when DT failed high,
: because I thought it was something we should see. I didn't realize
: when I put the note in just how deep the "plot" really was...
Of course I will give you my comments...
: However, if you look at the Deep Thought Nolot test results, and then
: extrapolate with Deep Blue being maybe 300x faster, and then being
: pessimistic and scaling this back to 100x, you get a frightening
: result. How many do you solve? Last time I tried Crafty it was not
: doing well. It certainly didn't solve all of them inside of 5 minutes
: each... And then if you look at the game mentioned above, and then at
: the game vs *Socrates, you begin to get an inkling of just how deep
: that thing searches.
: ANd then if you look at some of the Kasparov games. One in
: particular, the Bxh7 game... I've heard, indirectly, that if you
: ask Kasparov about that position you get an alarming answer. namely
: that he saw that Bxh7 didn't win, but even more interesting is that
: a3 *does* outright win. Rumor has it that DB saw this as well. Let's
: find a micro that can figure out just how strong a3 is and why (hint:
: some *very* deep threats that a3 prevents.) The conclusion is
: still quite obvious...
I like to avoid new circles in this stage.
I like to stay to the facts if you don't mind.
I also hate to repeat myself... :)
: Crafty, Rebel, Mchess Pro, Genius, and the rest can fight it out for
: the best micro, but DB's playing a different game entirely...
New circles...
- Ed Schroder -
Tom C. Kerrigan
> In any case, when blacked played 27. ... c5, they did so with a +2 eval
> after a fail high. More details are below, but it was about 26 *plies*
> later that CB finally reached -2. Notice that... 26 *plies*. It is
> certainly ugly beyond move 27, but it does illustrate just how powerful that
> machine is. You can draw your own conclusions of course...
I see this a lot in computer vs. computer games. One program all of a
sudden says +2 for no obvious reason, the other program disagrees until
about a gazillion moves later. Now I just assume that the program with +2
is searching a bit deeper and makes a lot of moves that don't
"immediately" result in a win of material, but it can still see the win
because of an extension or whatever. This annoys me a lot because it makes
me think I have a bug in my search, but I see it happen to other programs,
too, so I'm not quite as worried.
I'm not trying to rip on DT/DTII/DBP/DB/whatever here, but the 26 plies
may be a bit exaggerated...
Cheers,
Tom
Moritz Berger
On 17 Feb 1997 00:01:46 GMT, hy...@crafty.cis.uab.edu (Robert Hyatt)
wrote:< snip >
>: Bob Hyatt wrote:
>: "How about Rc6 in the Kasparov game where all the micros wanted to
>: play (either Rc6 or Rc7 whichever DB *didn't* play, and all would
>: have lost about 20 plies later, because the rook was a critical
>: piece there, a little later."
>: Bob Hyatt answered:
>: "I really have no way to find this out."
>
>: Ed Schroder asked again:
>: "I do not understand this.
>: In the above you claim DB was by far superior than the micro's because
>: of a very deep combination of 20 plies. I show you the game and ask for
>: proof so I can check it (and learn something from DB!) and you say
>: I don't know!
>: "Crafty liked Rc6 forever. I didn't give it forever to search,
>: however, but Rc6 (it thinks) wins a pawn. Obviously, if DB played
>: Rc7, it saw something better, but I don't have a clue as to what.
>: You made 2 claims:
>: 1.. Deep Blue played 32. Rc7 because 32. Rc6 would lose after 20 plies.
>: 2.. Deep Blue saw it, the micro's did not.
>: These are your words and not mine.
>Ed, I don't have the analysis. There was mucho discussion on ICC the day
>that happened. We had GM's analyzing, computers analyzing, and so forth.
>We were also getting feedback from the skittles room at the actual site.
>I simply remember when all the micros including crafty wanted to play
>Rc6, and there was a lot of discussion. Later, one of the GM's noticed
>that in the game, Rc7 was critical, because a knight move was a discovered
>check. He then did some analysis and pointed out that the Rc6 move really
>was a lemon after the fact.
>I don't know how to "prove" this because I certainly can't do an exhaustive
>20+ ply search to "prove" that this move won and Rc6 lost. However, I'll
>stand by my premise that DB's over 200 rating points better than the next closest
>machine.
"Micro Moritz" strikes again ;-)
M-Chess 6 prefers the move 32. Rc7 (over the previous Rc6) after 20:37
(minutes, not hours) at I think 10 ply depth (plus extensions ...). I
didn't try out other programs and this is a type of position where
M-Chess certainly plays very well. The score keeps getting better and
better with each additional ply ...
So far, I have found at least one micro program that would have played
as well or better in positions that Bob identified as crucial examples
of Deep Blue's (resp. DT's in the Hong Kong case) superiority.
The above time for M-Chess is certainly sufficient OTB on some faster
PC hardware, I only used my old Toshiba P5/133 notebook.
-------------
Moritz...@msn.com
mclane
Wolfgang Kuechle <kue...@theochem.uni-stuttgart.de> wrote:
>mclane wrote:
>> Ich denke Ed, es ist sinnlos weiter zu diskutieren. Sie leben in IHRER
>> Welt und werden es nicht verstehen !
>>
>> Sie wollen nichts sehen was *entgegen* ihrer, scheinbar mit einem
>> hohen Preis bezahlten, Ideologie an Ihre mit Scheuklappen versehenen
>> Augen dringt.
>Ist das bei dir anders ? :)
>Gruesse,
>Wolfgang Kuechle
Ich denke schon Wolfgang. Schau einmal in die schwedische Liste und
sage mir, wieviele von den schnell/dumm Programmen OBEN mitmischen.
Weder Mchess, noch Rebel, noch Hiarcs oder Genius oder The
King=ChessMaster sind SCHWACHE Programme, lediglich Fritz und Nimzo
machen es wie Deep Blue.
Aber mit welchem Erfolg ? Meinst du die Schweden hätten die Ergebnisse
extra für mich frisiert ?
Hast du schon mal Fernschach gespielt und Nimzo oder Fritz dafür als
Analyseengines benutzt ?!
Ist dir dabei nichts aufgefallen ?!
Robert Hyatt
: : 200 ???
: : 50 not more!
: > Utter and complete hogwash. I only hope we end up with a chance for
: > you to see just how wrong you are... If you believe 50 is correct,
: > that you would win 40% of the games against them, hmmm....
: In previous postings I said: DB-PC 6.5 - 3.5
: You answered with: DB-PC 9 - 1
: And for the moment I like to stick to my opinion.
correct. I still stick with 9-1 if we are talking about Deep Blue.
: : : As I mentioned before, are you interested in playing Crafty at 50:1
: : : time odds? I'd suspect not.
: : Yes I am interested as long as Rebel may play on tournament level!
: : Yes I take your challenge, Rebel will win a ten game match... :)
: > You hve the machines. However, back the trolly up. YOu were claiming
: > that 2K nodes per second was better than them in 1992. Why do you
: > suddenly want to go 50x faster? IE, give Crafty 250 seconds per move,
: > and Rebel 5, and that should be a close approximation to what was
: > happening in 1992. However, I'll take the odds at 3 mins vs 150 mins.
: > Of course, you do know you have to not think on opponent's time...
: > If you can win at that handicap, I'd likely go back to writing
: > compilers... :)
: Very funny, but I won't let you go away that easy... :)
: The discussion is of course (and has always been!) how well will the best
: PC programs perform against DB *NOW*. In this light we were always
: talking about tournament level.
: Right?
Yes, but Crafty at 50:1 is *not* DB now. Crafty at 10,000:1 time odds is
more like DB. So we aren't quite so close...
: Now you suddenly introduce an average time of 5 seconds??
I'm only thinking of how to make this happen. One move per day at
100:1 doesn't seem very attractive unless we are willing to accept the
outcome after 2 or 3 games, which will take until the Summer to pull
off.
How about a different offer: 10 seconds to 60 seconds? that's only 6:1,
but I'd still claim it's a serious winning advantage.
: : Somewhat broken test, of course... unless someone has a P6/200 with
: : something like 64 gigs if memory, otherwise hashing is going to
: : effectively turned off in Crafty. However, I'd still take the odds...
: : You should talk to Lonnie first, however. They are playing reasonably
: : closely at equal time controls in the games I've seen on ICC and
: : Chess.net...
: Well, you challenged me 2 times, the first time I ignored your challenge,
: the second time I accepted. I even proposed to you a handicap factor of
: 100 instead of your suggested factor of 50.
: And now you complain about hash table size?
: Is a factor 100 not enough for you?
: What do you want then?
: A 150, 200 factor??
No... you are going the wrong way. I'd like to finish the match in this
century if possible... :)
Here's *real* analysis. You claim, in 1992, that you were doing 2K nodes
per second, and you won the WCCC that year, which is certainly true. You
claim that proved that the micros were superior, even though you were doing
only 2K and Hitech was doing 100K. What I proposed was to test *that*
hypothesis, which would be crafty at any usual time control you want, and
Rebel at 1/50th of that time control. Again I'd suggest 5 secs for rebel,
which is about like 250 seconds in 1992. and something reasonable for Crafty,
such as 50 seconds, which is way less than the advantage HiTech had back
then. That will a game or two a day doable. And will simply give insight
into where micros *really* stood in 1992. And, by extrapolation, where they
stand today. Remember, if you give me 100X the time, you are talking *4*
plies, at least. I'm not aware of any experiment that says that 4 plies means
*nothing*. There have been extrapolations for N and N+1 done a couple of times,
but never N and N+4. What little data was available at that spread in Berliner's
paper was overwhelming for the 4 ply advantage...
: Or do you simply don't want to play at all?
Here's my take: I am interested if we can do it in a reasonable way. one or
two moves per day is not reasonable for several reasons. It's not "productive"
as far as Crafty is concerned, because it's a search done outside of any normal
time constraint I'd ever expect to play under. I get more information from playing
a few GM games or games against computers on the server at more normal time
controls, particularly when they are automatic and don't require any attention
from me.
: It is your chance to prove your "NPS rules..." statement.
I believe my statement fully, because it's been proven over and over... you just
have to look up Berliner's paper and Thompson's paper on what an additional ply
of search gains, rating wise. Between relatively equal programs, a ply is
significant. 4 plies is huge, and 6-8 plies is overwhelming. It's easy to
prove however... Just play any time odds you want that keeps the longer
time to something reasonable, as in under 5 mins/move or so, so that the game
won't go on forever, and so we can get a large enough sample that you will be
convinced the result isn't a fluke. If you are willing, I'll set crafty up
on ICC so it will only accept matches from you, and will not think on your time.
It will be easy to operate, and you can run Rebel to be sure you are getting the
best you can get. I ran a short such test this morning with Crafty vs Crafty,
1 second to 100 seconds, 10 games, result was 9 wins, 1 draw. The draw went
45 moves in book and the first move out of book was a King, Knight and three
pawns vs a King knight and 3 pawns. the extra depth did nothing there as all
the pawns were on the same side. In the other games, the 1 second side simply
got crushed tactically and positionally....
: : : The only conclusion I can draw is you think the DB guys are
: : : incompetent...
: : Hey Bob don't put words in my mouth I never said!!
: > but it's clearly implied. You want to play Crafty, but you want
: > to use the full factor of 50 we've gotten in the past 5 years. And you
: > claim it is important. And then you claim that the much larger
: > performance gains made by the DB guys is worth only 50 rating points.
: > The only conclusion I can see is that they must be incompetent, if you
: > can get so much from a factor of 50, but they get nothing from an
: > additional factor of 2,000...
: > Right?
: Bob this are your words and your conclusions, not mine. The deep blue
: team have my highest admiration for what they are doing now and have
: done in the past. As I said before I am not fighting Hsu but your
: unfounded claims, did you missed this in my previous posting? It was
: there in UPPERCASE... :)
you claim 50 points for a factor of 10,000 faster according to your exact
words. Am I correct? That's taken directly from what you posted....
: So how about the 10 game match?
: Rebel at tournament level, Crafty 100 x more time, all on a PP-200.
: You are right about playing without permanent brain. To compensate this
: for both I suggest the following schedule:
: Rebel 4:30 fixed time per move.
: Crafty 4:30 x 100 = 7.5 hours per move (nice overnight solution)
: How about it Bob?
: Remember you challenged me, I accepted...
: - Ed Schroder -
I'm game for the match, but the time is simply too long. one move per night,
figuring 40-60 moves per game, is 1.5 years to finish 10 games. Not worth the
trouble. Scale it back as I proposed. Scale the handicap back as well, to only
10:1, or even 5:1... then we can play enough games to at least conclude what a
5:1 or 10:1 handicap is worth...
1.5 years would mean the programs playing the last game are nothing at all like
the programs playing the first game... not to mention the loss of quality games
on the servers this would give up. I don't mind taking time to prove a point, but
one move per day is *really* slow and takes a *lot* of time away from normal
testing...
3 mins to 30 seconds, 1 min to 10 seconds... anything down in that range is
o.k... I don't mind playing them on even time controls because it would be
interesting, but it doesn't address the point of "speed kills" of course...
Robert Hyatt
Don Fong (df...@cse.ucsc.edu) wrote:
: In article <5e9l02$e...@juniper.cis.uab.edu>,: >: : The only conclusion I can draw is you think the DB guys are
: >
: >: Hey Bob don't put words in my mouth I never said!!
: >
: >but it's clearly implied.
: no, Ed Schroder is dead right here. time and again i have seen
: you play this game. you escalate the hostilities by drawing some extreme
: hypothetical conclusion based on YOUR beliefs plus THE OTHER PERSON's
: facts, then you try to impose this unviable hybrid on the other person.
: the focus of the debate then shifts from being an illuminating battle
: between two people's ideas, to a meaningless argument over your
: made-up fallacies.
The reread what he wrote. (a) DB is only 50 points better than Rebel,
with a computing advantage of 10,000... (b) yet Rebel is significantly
better now than in 1992, with only a 50X performance increase.
With those two statements, what other conclusion is there? They have a
bigger hardware advantage over rebel97 than rebel97 did over rebel92, yet
rebel97 is much better than rebel92, but DB is not. *what* other conclusion
is there? You jump in, now answer. Do you see any other explanation for the
above disparity? There's hardware, and there's programmers. The hardware
advantage is provable by counting nodes. So what can possibly offset that
factor of 10,000, except for incompetent programmers? This *has* been a common
claim, by you and others, that don't believe that DB is clearly the best. I
simply challenge that assertion. period...
: Ed Schroder did NOT say the DB team was incompetent. it is
Robert Hyatt
Moritz Berger (Moritz...@msn.com) wrote:
: On 17 Feb 1997 04:18:01 GMT, hy...@crafty.cis.uab.edu (Robert Hyatt)
: wrote:
: >Ed:
: >
: >Here's the game I mentioned between Cray Blitz and Deep Thought II. I
: >believe (if my file naming convention is correct) that this was played in
: >1991. I think in Orlando, but don't remember exactly.
: >
: >In any case, when blacked played 27. ... c5, they did so with a +2 eval
: >after a fail high. More details are below, but it was about 26 *plies*
: >later that CB finally reached -2. Notice that... 26 *plies*. It is
: >certainly ugly beyond move 27, but it does illustrate just how powerful that
: >machine is. You can draw your own conclusions of course...
: Chessmaster 5000 play 27. .. c5 after about 2 or 3 minutes on my
: P5/166 ... It needs about 2 or 3 million nodes to see it. Of course,
: this is not enough to justify a +2 eval (too much speculation or
: intelligence involved, too less exact "knowledge", therefore only a
: meagre 0.14 eval ...), but my point is, that micros can play the best
: moves without calculating zillions of nps. Point taken?
No. the discussion was about the computational prowess of DB, and the fact
that they didn't just play the best move, they played the best move *knowing*
why it was best. A better test would be to back up one move and see if it will
retreat the bishop rather than lose it, because it obviously doesn't know it
will be lost...
And of course, c5 doesn't win the bishop. There are many more moves that have
to be found first, including a couple that were somewhat "ugly" on the surface,
but which took away key squares for later threats...
The key is that on c5 they weren't just "equal" they knew they were winning.
I don't think any micro around, given a week to compute, would have a prayer
of finding that...
: BTW: Yes, I knew that Chessmaster does about the best mobility
Robert Hyatt
: : I wasn't thinking so clearly above, because there's an obvious problem
: : in playing a match at such a long time control. If we take your
: : "tournament mode" which I assume means 40 moves in 2 hours, then at
: : 100:1, crafty would be playing 40 moves in 200 hours." Since there's
: : not an easy way to "autoplay" this, we'd be lucky to get in two moves
: : per day or so, which turns into a game per month if we are lucky.
: Yes, unfortunately you are right...
: About Crafty...
: Is Crafty now fully compatible with AUTO232?
: If so we can also keep the permament brain and set Rebel on 40/2:00
: and Crafty on 40 in 200 hours. In this case we will have 4-5 moves
: a day, 2.5 games a month.
If you can do this, I believe crafty will work. It *will not work* with
thinking on the opponent's time due to some quirk I have not found, but
with thinking on the opponent's time disabled, I've had reports that it
works well. Doing it like that is fine by me... I can send detailed
instructions (and the latest version which appears to work with auto232
without pondering) if you'd like...
: : There's two problems:
: : 1. by the time a 10 game match is over, Rebel 9 and Crafty v12 will be
: : out. And we could then argue over the old program results... :)
: I disagree here. It's not only the score that will count but also it's
: my curiosity how well 70,000 NPS will perform against 7,000,000 NPS. I
: am very interested in this! Aren't you? And of course I like to
: win also... :)
yes I am... and the way you propose doing it is perfect in my book... it doesn't
take my machine offline at all, and also will be 2x-3x faster than shipping a move
back and forth every day.
: : 2. 10 games is probably not enough to draw any real conclusions from,
: : we really need maybe 100 games. That's not going to happen unless we
: : can find a bunch of people with P6/200's lying around that want to do
: : this, which I doubt...
: That's an old discussion and I agree. But 10 is better than ZERO and
: it will tell us a lot about our NPS argument.
: : ERGO, what about the following: a much faster game, say 5:1 time odds
: : with crafty getting 10 seconds, Rebel 2? If you believe your knowledge
: : is much superior to mine, that is a perfectly valid test. In fact,
: : knowledge-based programs should do better at fast time controls than at
: : longer ones, because the effect of knowledge is more pronounced when you
: : can't search, and make moves based only on knowledge. Try playing a
: : rank beginner and not thinking at all, just follow normal practices such
: : as controlling the center, watching pawn advances, piece placement and
: : the like.
: I am not interested in a 0:10 / 0:02 match.
: I don't like your idea because I am only interested in tournament level
: play. Fritz, Genius will win at blitz-level against Rebel but on
: tournament level... You know what I mean.
no problem as you've proposed it. Assuming you don't mind the lost machine
time there...
: : However, I hope you have a chance to look over the Cray Blitz vs Deep
Robert Hyatt
Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
: : : I wasn't thinking so clearly above, because there's an obvious problem
: : : in playing a match at such a long time control. If we take your
: : : "tournament mode" which I assume means 40 moves in 2 hours, then at
: : : 100:1, crafty would be playing 40 moves in 200 hours." Since there's
: : : not an easy way to "autoplay" this, we'd be lucky to get in two moves
: : : per day or so, which turns into a game per month if we are lucky.
: : Yes, unfortunately you are right...
: : About Crafty...
: : Is Crafty now fully compatible with AUTO232?
: : If so we can also keep the permament brain and set Rebel on 40/2:00
: : and Crafty on 40 in 200 hours. In this case we will have 4-5 moves
: : a day, 2.5 games a month.
: If you can do this, I believe crafty will work. It *will not work* with
: thinking on the opponent's time due to some quirk I have not found, but
: with thinking on the opponent's time disabled, I've had reports that it
: works well. Doing it like that is fine by me... I can send detailed
: instructions (and the latest version which appears to work with auto232
: without pondering) if you'd like...
to get this rolling, here's a couple of ideas:
1. do you want to go to the trouble of downloading the book I use, and all
the learned stuff? Or would you prefer to use a small book, and simply abort
a game that looks like a book bust? If you are using any sort of book from
my ftp site, for 10 games, I'd suggest you simply watch, and any opening that
pops out of book at > +1.5 or < -1.5 gets aborted to avoid wasting time. For
a long match, Crafty would learn to play the right moves from any book, but
since Rebel doesn't, it would be easier to simply make sure that each game
starts off on a relatively even position...
2. I can tell you how to ftp the current version which has a tweak or two
for auto232 that seems to make it work. One of the SSDF guys reported 10
games between Crafty and genius worked, so long as no thinking on the opponent's
time was allowed for Crafty (he was simply trying everything he could to help me
figure out where this thing is broken since I can't test it here.)
If you want the current book, you need about 1/2 gig to build it, but < 100mb
to actually store it. Tablebases are probably not worth the trouble...
Let me know how you'd like to proceed. I do believe that there's a mirror
site in Europe for my ftp site that has the book stuff, so all you'd really
need from me is the current version (11.17) with the latest auto232 mods,
which is not yet on my ftp machine for public access...
Perhaps someone from Europe might give us a URL for the mirror site there
if Ed wants to grab the book...
Robert Hyatt
I'd like to appologize for some of my responses being "out of sequence" today.
I'm not sure what's been going on, but we got almost no news articles during the
day, I could not connect to ICC, chess.net or anywhere else except EICS and FICS.
Now, as I am responding to one article, 18 new ones show up, and it's obvious
that the order is screwed up. I'm not nuts, and not suffering from jet-lag.
I just figured out that things are out of sequence here and things are still
pouring in as I write this. If you asked something of me, it will eventually
show up and I'll respond.
Tom C. Kerrigan
Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
> Also notice that Hsu doesn't consider a position solved *just* because DT
> would play the right move. He evaluates just as I do, so that if it doesn't
> have a score to prove that it understands that a move is best, it's not right.
> The is as opposed to many that claim success just because they get the right
> move, period...
Just a sec, here...
If a program gets the right move, then it has the score to prove it,
unless it's a particularly buggy program.
Granted, it may be a step down from a material win, but a material win is
a step down from mate, too...
Cheers,
Tom
Tom C. Kerrigan
Ed Schroder (rebc...@xs4all.nl) wrote:
> The current facts are:
> - Madrid 1992, a 2,000 NPS micro won against 150,000 NPS;
Rebel searches 540K nodes per move, Stobor searches 40.5M...
> - Hong Kong 1995, a 100,000 NPS micro won against 3,000,000 NPS;
Rebel searches 27M, Stobor searches 810M...
This might be fun... Ed?
Cheers,
Tom
Tom C. Kerrigan
Wolfgang Kuechle (kue...@theochem.uni-stuttgart.de) wrote:
> mclane wrote:
> > Ich denke Ed, es ist sinnlos weiter zu diskutieren. Sie leben in IHRER
> > Welt und werden es nicht verstehen !
> > Sie wollen nichts sehen was *entgegen* ihrer, scheinbar mit einem
> > hohen Preis bezahlten, Ideologie an Ihre mit Scheuklappen versehenen
> > Augen dringt.
> Ist das bei dir anders ? :)
LOL... good point!
Thorsten: "Private" conversations with Ed should be conducted via e-mail,
not German.
Cheers,
Tom
Dave Gomboc
In article <5e7tjf$pnb$1...@news1.xs4all.nl>,
Ed Schroder <rebc...@xs4all.nl> wrote:
[snippage of conversation between the heavyweights ;-]
#12 plies to see it's a bad move. Here is the analysis again:
#
#16.c4 Qh4 17.cxb5 Rh6 18.Nf6+ Kd8 19.h3 Bxh3 20.gxh3 Qxh3 21.Kg1 Qh2# (12)
#16.c4 Qh4 17.cxb5 Rh6 18.Nf6+ Kd8 19.h3 Bxh3 20.Be6+ 21.Kg1 Qh2# (12)
#16.c4 Qh4 17.g3 Qh3 18.Qe2 fxg3 19.fxg3 Rgx3 (8)
#16.c4 Qh4 17.g3 Qh3 18.Qf3 Rh6 19.Qg2 Qxg2 20.Kxg2 Bh3+ (10)
#
#If you find a deeper (better) defense for white I am all ears.
#
#You answered with: I didn't...
#
#Come on Bob, DB prototype failed miserabely on a 12 ply deep variation.
#Rebel, Hiarcs, Fritz, Genius all see it, DB prototype didn't and had
#to give the queen for a rook 4 moves later (game over).
#
#I find it useful if you or somebody else would post a set of positions
#where DB has seen deep combinations so we can learn more by simply
#check these positions with the available chess programs.
#
#- Ed Schroder -
A few years ago FHH posted what DT-2 thought about the Nolot positions.
Perhaps someone has this post archived. It would be interesting to see
who can match DT-2's results today.
It would also be interesting to see how DB does on the Nolot positions...
Dave Gomboc
drgo...@a.stu.athabascau.ca
Robert Hyatt
Tom C. Kerrigan (kerr...@merlin.pn.org) wrote:
: Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
: Just a sec, here...
: Cheers,
: Tom
No. not what I meant at all. If you play the "right" move but don't see anything
good as a result, you just flipped a coin N times, and on the right move it happened
to turn up heads. Which do you want to do, play the right move for a random
positional reason, or play the right move because you see it solves a tactical
(and very deep) problem? There's one WAC position that Crafty solves positionally
and which is about .06 better than the previous best move. A couple of more plies
and it sees a +2.7 or so, which is the right solution.
In the Cray Blitz vs DT game, crafty would play c5, but it wouldn't know that it
leads to a material win, and it might not follow up with the correct next move.
You can count what you want right, of course. but if you don't see *why* it's
right, you aren't solving the position, just playing the right move. There's a
subtle difference that's important... the former "solution" can be lost by a minor
eval change, which is what happens to me in the WAC position that causes problems.
The latter "solution" will never be lost by eval tweaks, which is the way I'd
want it... Positional problems are more difficult to analyze, but tactical
problems should have the complete tactical solution... or at least a tactical
solution that is convincing. IE, +15 rather than mate in 19 might be acceptable.
I have seen one person post a solution where the right move was -2, while the
previous best moves was -2.3... that's not very convincing...
Robert Hyatt
Moritz Berger (Moritz...@msn.com) wrote:
: On 17 Feb 1997 04:18:01 GMT, hy...@crafty.cis.uab.edu (Robert Hyatt)
: wrote:
: >Ed:
: >
: >Here's the game I mentioned between Cray Blitz and Deep Thought II. I
: >believe (if my file naming convention is correct) that this was played in
: >1991. I think in Orlando, but don't remember exactly.
: >
: >In any case, when blacked played 27. ... c5, they did so with a +2 eval
: >after a fail high. More details are below, but it was about 26 *plies*
: >later that CB finally reached -2. Notice that... 26 *plies*. It is
: >certainly ugly beyond move 27, but it does illustrate just how powerful that
: >machine is. You can draw your own conclusions of course...
: Chessmaster 5000 play 27. .. c5 after about 2 or 3 minutes on my
: P5/166 ... It needs about 2 or 3 million nodes to see it. Of course,
: this is not enough to justify a +2 eval (too much speculation or
: intelligence involved, too less exact "knowledge", therefore only a
: meagre 0.14 eval ...), but my point is, that micros can play the best
: moves without calculating zillions of nps. Point taken?
It's not hard to play c5, as the following shows:
depth time score variation (1)
1 0.02 0.095 d5
1-> 0.02 0.095 d5
2 0.02 -0.117 d5 Rfe2
2-> 0.02 -0.117 d5 Rfe2
3 0.03 0.019 d5 Bf5 Re8
3-> 0.03 0.019 d5 Bf5 Re8
4 0.06 -0.084 d5 b4 Bg4 a3
4-> 0.08 -0.084 d5 b4 Bg4 a3
5 0.12 -0.058 d5 b4 Bg4 h3 Be6
5-> 0.29 -0.058 d5 b4 Bg4 h3 Be6
6 0.44 -0.166 d5 b4 Bg4 h3 Bh5 a3
6 0.86 -0.160 g6 Bd4 c5 Be4 Re8 Bxe5 dxe5
6 1.32 -0.150 Re7 Bd4 Rae8 Bxe5 fxe5 Be4 d5
6 1.70 -0.144 Re8 Rfe2 g6 c4 d5 cxd5 Bxd5
6-> 1.94 -0.144 Re8 Rfe2 g6 c4 d5 cxd5 Bxd5
7 2.95 -0.129 Re8 Bd4 c5 Bxe5 dxe5 Be4 f5 Bc6
7 7.42 -0.127 c5 Be4 d5 Bc2 c4 bxc4 dxc4 Rfe2
7-> 8.81 -0.127 c5 Be4 d5 Bc2 c4 bxc4 dxc4 Rfe2
8 20.46 -0.097 c5 c4 Ra6 Rf4 Bd7 Bd8 Ra8 Bb6 Bxa4
bxa4
8-> 23.44 -0.097 c5 c4 Ra6 Rf4 Bd7 Bd8 Ra8 Bb6 Bxa4
bxa4
9 34.58 -0.141 c5 Be4 d5 Bc2 c4 bxc4 dxc4 Nc5 Bxc5
Bxc5 Rd7
9-> 57.25 -0.141 c5 Be4 d5 Bc2 c4 bxc4 dxc4 Nc5 Bxc5
Bxc5 Rd7
time: 1:03 cpu:99% mat:0 n:4703583 nps:74235
ext-> checks:60123 recaps:18389 pawns:0 1rep:26513
nodes full:691229 quiescence:4012354 evals:3206991
endgame tablebase-> probes done: 0 successful: 0
Of course, a big point here is that Crafty also does bishop mobility which
makes c5 inteesting. However, note that after a minute (no big hash or
anything however) it has no clue. And it won't have a clue for 20+
more plies that this is actually winning.
that was the point of my DB post, that Deep Thought actually saw the whole
forcing line that none of us could figure out at the time because it was too
deep. I remember Valvo looking at it too, and scratching his head.
This is one of the few positions I have where I have CB output and knew what
DT was seeing at the same point in time. I was impressed, but I can honestly
say I didn't enjoy it at all. :)
Ed Schroder
From: kerr...@merlin.pn.org (Tom C. Kerrigan)
> The current facts are:
> - Madrid 1992, a 2,000 NPS micro won against 150,000 NPS;
: Rebel searches 540K nodes per move, Stobor searches 40.5M...
> - Hong Kong 1995, a 100,000 NPS micro won against 3,000,000 NPS;
: Rebel searches 27M, Stobor searches 810M...
: This might be fun... Ed?
: Cheers,
: Tom
I have already accepted Bob's challenge...
I can only handle one... :)
Maybe next time?
- Ed Schroder -
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
Ed Schroder wrote:
"Hey Bob don't put words in my mouth I never said!!"
: no, Ed Schroder is dead right here. time and again i have seen
: you play this game. you escalate the hostilities by drawing some
: extreme
: hypothetical conclusion based on YOUR beliefs plus THE OTHER PERSON's
: facts, then you try to impose this unviable hybrid on the other person.
: the focus of the debate then shifts from being an illuminating battle
: between two people's ideas, to a meaningless argument over your
: made-up fallacies.
> The reread what he wrote. (a) DB is only 50 points better than Rebel,
> with a computing advantage of 10,000... (b) yet Rebel is significantly
> better now than in 1992, with only a 50X performance increase.
> With those two statements, what other conclusion is there? They
> have a bigger hardware advantage over rebel97 than rebel97 did over
> rebel92, yet rebel97 is much better than rebel92, but DB is not.
> *what* other conclusion is there? You jump in, now answer. Do you
> see any other explanation for the above disparity? There's hardware,
> and there's programmers. The hardware advantage is provable by
> counting nodes. So what can possibly offset that factor of 10,000,
> except for incompetent programmers? This *has* been a common
> claim, by you and others, that don't believe that DB is clearly
> the best. I simply challenge that assertion. period...
Wish you never had written this Bob!
I simply have to agree on Don Fong's posting NOW I have read your answer.
Yes you trough with mud if you are on the losing side, attack the
person's creditibilty. In this case you actually are implying "Ed has
only one goal attack DB and Hsu". This is very dirty of you Bob!
I am a chess programmer for 15 years now. When I started I had only one
goal to get the best out of myself and make a chess program that plays
as strong as the good Lord will allow me.
After 15 years I still have this passion for playing strength. I respect
any of my collegues because I know the prize everybody has to pay writing
a chess program. It's blood sweat and tears. No?
I love Deep Blue, I have said that many times here in RGCC, I have
covered the Kasparov - Deep Blue match on my home page (it's still there)
and that cost me weeks of work. Why? Because I love Deep Blue and all
the work they are doing.
I told you many times I am not fighting Hsu or DB, I am fighting YOUR
CLAIMS! I am not even fighting Bob Hyatt which I respect, NO I am
figthing your opinions because I disagree! Can't you see that
difference Bob?
Sofar I have been enjoying our discussion, one strong opinion against
another strong opinion. It was a good and honest discussion sofar!
I love strong discussions it forces both participants to get the most
out of the subject with just one goal; a better understanding!
I am not interested about who is wrong and who is right, I am interested
in how strong Deep Blue is and I always will say "I was wrong" if it
turns out to be I was wrong. You must know that by now don't you?
Despite of that you keep on saying that my intensions are different! You
imply that I have said "Hsu is incompetent". I never said that neither
it has been my goal. It's your conclusion and your words!
I have written this now 3 times, still you keep on saying this!
Don't you believe me Bob?
Look what Don Fong wrote:
"Ed Schroder did NOT say the DB team was incompetent. it is
NOT clearly implied by anything he said. it may be implied by YOUR
beliefs, but not HIS.
it's like when someone disagrees with you and you say ``ARE YOU
CALLING ME A LIAR?'' it is the tactic of a schoolyard bully."
He is right, since you keep on saying your LIE about my intentions you
are calling me indeed a liar. How many times must I tell you that I
love Hsu and his wonderful hardware?
No you simply continue your mud trying to convince the reader I have
said Hsu is incompetent. I did not! and I never implied it.
Wish you will agree.
- Ed Schroder -
brucemo
mclane wrote:
>
> hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>
> >Any one micro had a very small chance of winning the event, probably the
> >best was at the 10% probability level. But there were a *bunch* of
> >micros there, and the liklihood that one would win, with only one strong
> >"mainframe" there suddenly became quite high. The winner in 1992 was
> >basically chosen by a coin toss, not by playing skill.
>
> Welt und werden es nicht verstehen !
>
> Sie wollen nichts sehen was *entgegen* ihrer, scheinbar mit einem
> hohen Preis bezahlten, Ideologie an Ihre mit Scheuklappen versehenen
> Augen dringt.
>
> Himmel. Und kein Materialist wird Qualität verstehen.
Are you trying to make some point by only posting in German these days?
bruce
Moritz Berger
On Tue, 18 Feb 1997 00:55:18 GMT, mcl...@prima.ruhr.de (mclane) wrote:
<snip>
>Ich denke schon Wolfgang. Schau einmal in die schwedische Liste und
>sage mir, wieviele von den schnell/dumm Programmen OBEN mitmischen.
>Weder Mchess, noch Rebel, noch Hiarcs oder Genius oder The
>King=ChessMaster sind SCHWACHE Programme, lediglich Fritz und Nimzo
>machen es wie Deep Blue.
>Aber mit welchem Erfolg ? Meinst du die Schweden hätten die Ergebnisse
>extra für mich frisiert ?
Come on, Thorsten - may I remind you of your own words that Fritz 4.01
plays 80-100 SSDF ELO points better with 32 MB hash? You wrote it here
in this newsgroup, I can look it up for you if necessary ... Of
course, this is yesterday's talk, maybe you have reached different
conclusions in the meantime ...
Moritz
>Hast du schon mal Fernschach gespielt und Nimzo oder Fritz dafür als
>Analyseengines benutzt ?!
>Ist dir dabei nichts aufgefallen ?!
>
>
>
-------------
Moritz...@msn.com
Vincent Diepeveen
>Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
>: Ed Schroder (rebc...@xs4all.nl) wrote:
>: : Yes I am interested as long as Rebel may play on tournament level!
>: : Yes I take your challenge, Rebel will win a ten game match... :)
>
>: You hve the machines. However, back the trolly up. YOu were claiming
>: that 2K nodes per second was better than them in 1992. Why do you suddenly
>: want to go 50x faster? IE, give Crafty 250 seconds per move, and Rebel
>: 5, and that should be a close approximation to what was happening in 1992.
>: However, I'll take the odds at 3 mins vs 150 mins. Of course, you do know
>: you have to not think on opponent's time... If you can win at that handicap,
>: I'd likely go back to writing compilers... :)
>
>I wasn't thinking so clearly above, because there's an obvious problem
>in playing a match at such a long time control. If we take your
>"tournament mode" which I assume means 40 moves in 2 hours, then at
>100:1, crafty would be playing 40 moves in 200 hours." Since there's
>not an easy way to "autoplay" this, we'd be lucky to get in two moves
>per day or so, which turns into a game per month if we are lucky.
>
>There's two problems:
>
>1. by the time a 10 game match is over, Rebel 9 and Crafty v12 will be
>out. And we could then argue over the old program results... :)
>
>2. 10 games is probably not enough to draw any real conclusions from,
>we really need maybe 100 games. That's not going to happen unless we
>can find a bunch of people with P6/200's lying around that want to do
>this, which I doubt...
>
>ERGO, what about the following: a much faster game, say 5:1 time odds
>with crafty getting 10 seconds, Rebel 2? If you believe your knowledge
>is much superior to mine, that is a perfectly valid test. In fact,
>knowledge-based programs should do better at fast time controls than at
>longer ones, because the effect of knowledge is more pronounced when you
>can't search, and make moves based only on knowledge. Try playing a
>rank beginner and not thinking at all, just follow normal practices such
>as controlling the center, watching pawn advances, piece placement and
>the like.
>
>At least at that time control, we can see the results of a 100 game
>match before we are old and retired... :)
>
>However, I hope you have a chance to look over the Cray Blitz vs Deep
>Thought II game I found and posted last night. Fortunately I kept the
>log files from Cray Blitz for that event, so I can see exactly what it
>was seeing, and I manually inserted the note when DT failed high, because
>I thought it was something we should see. I didn't realize when I put
>the note in just how deep the "plot" really was...
>
>However, if you look at the Deep Thought Nolot test results, and then
>extrapolate with Deep Blue being maybe 300x faster, and then being
>pessimistic and scaling this back to 100x, you get a frightening
>result. How many do you solve? Last time I tried Crafty it was not
>doing well. It certainly didn't solve all of them inside of 5 minutes
>each... And then if you look at the game mentioned above, and then at
>the game vs *Socrates, you begin to get an inkling of just how deep that
>thing searches.
>
>ANd then if you look at some of the Kasparov games. One in particular,
>the Bxh7 game... I've heard, indirectly, that if you ask Kasparov about
>that position you get an alarming answer. namely that he saw that Bxh7
>didn't win, but even more interesting is that a3 *does* outright win.
>Rumor has it that DB saw this as well. Let's find a micro that can figure
>out just how strong a3 is and why (hint: some *very* deep threats that a3
>prevents.) The conclusion is still quite obvious...
a3 is a sure and slow win, majority on the queenside, closed centre,
and better kingside. Bxh7 is a spectacular quick and very dangerous win, as
for several moves black is a piece up, which in the foreseable future cannot
move. If you play a bad move however in the bxh7 attack...
So if a sure win is awarded points for the best move prize,
then 1.a3 is better, definitely. In all other cases Bxh7 is clearly better.
If Bxh7 would not have won, then Kasparov would have shown it,
either immediately after the game or in an analyzes.
I have so far not seen a serious move that prevents black from loosing.
Someone reported that after Bxh7,Kxh7...etc. ...,Qe8 is a better defence
than reported by Berliner in ICCA, but still looses after a move or 10/20
(black piece up doesn't do a thing, black king in centre, and white
controlling centre and running with freepawn on h-file).
>Crafty, Rebel, Mchess Pro, Genius, and the rest can fight it out for the
>best micro, but DB's playing a different game entirely...
If they improve evaluation, then this is definitely true. If they use
same evaluation PC programs have, then this doesn't make a difference
against prepared Grandmasters. Kasparov doesn't even need to prepare,
like he showed last PR-match.
Vincent
--
+----------------------------------------------------+
| Vincent Diepeveen email: vdie...@cs.ruu.nl |
| http://www.students.cs.ruu.nl/~vdiepeve/ |
+----------------------------------------------------+
Moritz Berger
On Mon, 17 Feb 1997 22:49:42 GMT, Moritz...@msn.com (Moritz Berger)
wrote:
>< snip >
>>: Bob Hyatt wrote:
>>: "How about Rc6 in the Kasparov game where all the micros wanted to
>>: play (either Rc6 or Rc7 whichever DB *didn't* play, and all would
>>: have lost about 20 plies later, because the rook was a critical
>>: piece there, a little later."
>< snip >
>>: Bob Hyatt answered:
>>: "Crafty liked Rc6 forever. I didn't give it forever to search,
>>: however, but Rc6 (it thinks) wins a pawn. Obviously, if DB played
>>: Rc7, it saw something better, but I don't have a clue as to what.
>
>>: You made 2 claims:
>>: 1.. Deep Blue played 32. Rc7 because 32. Rc6 would lose after 20 plies.
>>: 2.. Deep Blue saw it, the micro's did not.
>>: These are your words and not mine.
2.: There are some micros that will see the move, so it might be
indeed the case that it's essential to play Rc7 instead of Rc6 ...
Unfortunately the programs will not show me their 21 ply selective
variants, so I also don't know why exactly they prefer the move ...
Ask Richard Lang or Marty Hirsch for the source code and find out
yourself ;-)
<snip>
>M-Chess 6 prefers the move 32. Rc7 (over the previous Rc6) after 20:37
>(minutes, not hours) at I think 10 ply depth (plus extensions ...). I
>didn't try out other programs and this is a type of position where
>M-Chess certainly plays very well. The score keeps getting better and
>better with each additional ply ...
<snip>
>The above time for M-Chess is certainly sufficient OTB on some faster
>PC hardware, I only used my old Toshiba P5/133 notebook.
<snip>
Genius 5 on my P5/133 notebook finds 32. Rc7 after 05:30
(minutes:seconds) at 9 ply with a score of 2.21 and holds the move for
at least 10 hours (I checked up to 14 ply ...).
Also, Genius 5 beat Kasparov at 25 min. / game 1.5 - 0.5 in London in
1994 ... On P90 hardware ... No, it doesn't have 1024 dedicated
processors and doesn't "know exactly" that it will win 20 ply deep
variants. I wonder if Deep Blue could also have found the right moves
in those games? ;-)))))))
Here are some examples from the London event where a certain
Kasparov,G was on the receiving end of some of Genius' deep variants
;-) Games were taken from the CG 5 CD ... Please note that this was a
regular "Human" event where Kasparov really would have needed to win
the 2nd game ... He certainly was motivated, but obviously this wasn't
enough ... Maybe if they had offered him more money ;-)
[Event "London Intel GP 25' (1)"]
[Date "1994"]
[White "Kasparov,G"]
[Black "ChessGenius 3"]
[Result "0-1"]
1. c4 c6 2. d4 d5
3. Nf3 Nf6 4. Qc2 dxc4 5. Qxc4 Bf5 6. Nc3 Nbd7 7. g3 e6 8.
Bg2
Be7 9. O-O O-O 10. e3 Ne4 11. Qe2 Qb6 12. Rd1 Rad8 13. Ne1
Ndf6
14. Nxe4 Nxe4 15. f3 Nd6 16. a4 Qb3 17. e4 Bg6 18. Rd3
Qb4 19. b3 Nc8 20. Nc2 Qb6 21. Bf4 c5 22. Be3 cxd4
23. Nxd4 Bc5 24. Rad1 e5 25. Nc2 Rxd3 26. Qxd3 Ne7 27. b4
Bxe3+ 28. Qxe3 Rd8 29. Rxd8+ Qxd8 30. Bf1 b6 31. Qc3 f6
32. Bc4+ Bf7 33. Ne3 Qd4 34. Bxf7+ Kxf7 35. Qb3+ Kf8 36. Kg2
Qd2+ 37. Kh3 Qe2 38. Ng2 h5 39. Qe3 Qc4 40. Qd2
Qe6+ 41. g4 hxg4+ 42. fxg4 Qc4 43. Qe1 Qb3+ 44. Ne3 Qd3
45. Kg3 Qxe4 46. Qd2 Qf4+ 47. Kg2 Qd4 48. Qxd4 exd4 49. Nc4
Nc6 50. b5 Ne5 51. Nd6 d3 52. Kf2 Nxg4+ 53. Ke1 Nxh2 54. Kd2
Nf3+ 55. Kxd3 Ke7 56. Nf5+ Kf7 57. Ke4 Nd2+ 58. Kd5 g5
59. Nd6+ Kg6 60. Kd4 Nb3+ 0-1
[Event "London Intel GP 25' (2)"]
[Date "1994"]
[White "ChessGenius 3"]
[Black "Kasparov,G"]
[Result "1/2-1/2"]
1. d4 Nf6 2. c4 e6 3. Nf3 b6 4. a3 Bb7 5. Nc3 d5 6. Bg5 Be7
7. e3 O-O 8. Bd3 Nbd7 9. cxd5 exd5 10. O-O c5 11. Rc1 Ne4
12. Bf4 a6 13. Qc2 Ndf6 14. dxc5 Bxc5 15. Rfd1 Qe8 16. b4 Be7
17. Be2 Rc8 18. Qb2 b5 19. Nd4 Nd6 20. Bd3 Nc4 21. Qb3 Nh5
22. Bf5 Ra8 23. Nde2 Nf6 24. Bg5 Rd8 25. Nf4 d4 26. exd4 h6
27. Bxf6 Bxf6 28. Nce2 Be4 29. Bxe4 Qxe4 30. Qg3 Rfe8 31. Qc3
Rd6
32. Re1 Red8 33. Rcd1 Bxd4 34. Nxd4 Qxf4 35. Ne2 Qe5 36. Rxd6
Rxd6 37. a4 Re6 38. Qc1 Qd6 39. axb5 axb5 40. Ng3 Qxb4 41.
Rxe6
fxe6 42. h3 Qc5 43. Nf1 Qd5 44. Qa1 Qe5 45. Qa7 Kh7 46. Qd7
Qd5
47. Qe7 Qd6 48. Qb7 Qd5 49. Qe7 Qe5 50. Qd7 Nd6 51. Ne3 Nf5
52. Qd3 Kg8 53. Qd8+ Kf7 54. Qd7+ Kg6 55. Qd3 Qd4 56. Qb1
1/2-1/2
-------------
Moritz...@msn.com
Robert Hyatt
: > after a fail high. More details are below, but it was about 26 *plies*
: > later that CB finally reached -2. Notice that... 26 *plies*. It is
: > certainly ugly beyond move 27, but it does illustrate just how powerful that
: > machine is. You can draw your own conclusions of course...
: I see this a lot in computer vs. computer games. One program all of a
: sudden says +2 for no obvious reason, the other program disagrees until
: about a gazillion moves later. Now I just assume that the program with +2
: is searching a bit deeper and makes a lot of moves that don't
: "immediately" result in a win of material, but it can still see the win
: because of an extension or whatever. This annoys me a lot because it makes
: me think I have a bug in my search, but I see it happen to other programs,
: too, so I'm not quite as worried.
: I'm not trying to rip on DT/DTII/DBP/DB/whatever here, but the 26 plies
: may be a bit exaggerated...
: Cheers,
: Tom
Exaggerated how? They reported +2 at move N. CB saw -2 at move N+X where
X was 20+ plies. How can that be exaggeration? CB is no "lightweight" in
tactics, yet DT had +2 for about 10 moves before CB finally found the -2 for
itself. Seemed convincing to me... because the evaluations showed what both
were seeing. DT wasn't stumbling around and fell into this, it searched its
way into this...
Robert Hyatt
mclane (mcl...@prima.ruhr.de) wrote:
: It will be difficult to show evidence that Deep Blue or any other big
: iron is superiour because this IS NOT TRUE.
: Therefore people involved in this business can only claim but not show
: in CHESS-LINES even because they don't play chess or they claim
: without knowing much about.
Again I guess you can't read. I gave the only concrete example of the
calculating power of DT II that I had personal knowledge about, the game
CB vs DT II posted yesterday. That clearly showed just how deep they were
searching back then... I claim to know about *that* case becaues I could
see their output, and CB's output at the same time. And I *was* there...
: They cannot show us that DB is strong, DB is only strong because of
: the hardware difference. It is not superiour. IMO it plays weaker than
: other programs and only the hardware difference hides this.
: But sometimes games against micros show that this is the truth.
you need to learn how to write also. You say it's *not* strong, then you
say it *is* strong, but only because of the hardware. Which is it? How can
it play weaker, yet the hardware difference hides this? If that's not a
statement based on simple prejudice and ignorance, what is it?
Robert Hyatt
Yes, I can... but and *but* and *BUT*... you claim a > 50 point improvement
in rebel due to a 50x hardware boost. You claim only a 50 point improvement
for DB even though they are 10000X faster. Since their hardware is better,
what's the reason that the extra factor of 200 doesn't give them *anything*
at all? Can't be the hardware. *must* be the software. And it didn't drop
from the sky, they wrote it. Therefore they *have* to be doing something that
is *not very good*. That's what I read in what you posted...
: Sofar I have been enjoying our discussion, one strong opinion against
: another strong opinion. It was a good and honest discussion sofar!
: I love strong discussions it forces both participants to get the most
: out of the subject with just one goal; a better understanding!
: I am not interested about who is wrong and who is right, I am interested
: in how strong Deep Blue is and I always will say "I was wrong" if it
: turns out to be I was wrong. You must know that by now don't you?
: Despite of that you keep on saying that my intensions are different! You
: imply that I have said "Hsu is incompetent". I never said that neither
: it has been my goal. It's your conclusion and your words!
I didn't claim you said it. I claimed you implied it, as explained above.
What other explanation can you offer for the +50 points only point you made
earlier?
: I have written this now 3 times, still you keep on saying this!
Robert Hyatt
Moritz Berger (Moritz...@msn.com) wrote:
: Here are some examples from the London event where a certain
Great games of course. I looked at them right after they were played.
However, on chess.net a few month ago, 5 programd did this same thing to
5 GM's... 5 programs finished with more wins each than the best GM. So
winning game/30 is impreseive, but it's not nearly so impressive as winning
a tourament time control like 40/2/20/1 for example.
I'm not surprised that most programs eventually find Rc7, because clearly it
turned out to be better, and post-game analysis showed it was better. The
point was, of course, DB found it OTB. Most micros last year couldn't. Sure
the micros are faster now with the P6/200, but DB's also 4x faster *again*
due to their hardware. It's difficult to catch up... :)
Robert Hyatt
Vincent Diepeveen (vdie...@cs.ruu.nl) wrote:
No it's not. it's a strong tactical win. That's the whole point. Kasparov
agrees with this in comments he's made as well. a3 was not a "wait and see"
it removed a key square from black's use, so that Bxh7 actually does work as
planned...
: So if a sure win is awarded points for the best move prize,
: then 1.a3 is better, definitely. In all other cases Bxh7 is clearly better.
: If Bxh7 would not have won, then Kasparov would have shown it,
: either immediately after the game or in an analyzes.
he did show it, by not playing it...
: I have so far not seen a serious move that prevents black from loosing.
: Someone reported that after Bxh7,Kxh7...etc. ...,Qe8 is a better defence
: than reported by Berliner in ICCA, but still looses after a move or 10/20
: (black piece up doesn't do a thing, black king in centre, and white
: controlling centre and running with freepawn on h-file).
: >Crafty, Rebel, Mchess Pro, Genius, and the rest can fight it out for the
: >best micro, but DB's playing a different game entirely...
: If they improve evaluation, then this is definitely true. If they use
: same evaluation PC programs have, then this doesn't make a difference
: against prepared Grandmasters. Kasparov doesn't even need to prepare,
: like he showed last PR-match.
I agree with you here. I'm looking forward to see just how they have improved...
but clearly speed isn't the only problem when Kasparov is across the board...
: Vincent
Vincent Diepeveen
Perhaps i can try few matches Diep-Stobor?
I estimate Diep has about 2360 SSDF list currently (both programs
on fast hardware NOT on a P90, but on MMX200 - Pro200).
First need some time to connect Diep to internet.
both on tournament level, Stobor is running on a mainframe?
Give me few weeks... (months if i'm not lucky).
Vincent
>Maybe next time?
>
>- Ed Schroder -
>
>
Vincent Diepeveen
In <330971...@nwlink.com> brucemo <bru...@nwlink.com> writes:
>mclane wrote:
[cutoff]
My German is sufficient, i'll try to translate the
last 2 sentences for the English speaking audience,
that will explain why Marx, sorry mean Mclane tried
not to post in English.
>> Ein Kamel geht nicht durch ein Nadeler. Und kein Reicher kommt in den
No rich man will go to heaven.
>> Himmel. Und kein Materialist wird Qualitaet verstehen.
And no materialist will understand quality.
>Are you trying to make some point by only posting in German these days?
No he's asking asylum from the North-Korean government.
They refused, they didn't have enough food to feed him.
>bruce
Dave Gomboc
In article <5e9l02$e...@juniper.cis.uab.edu>,
#: Yes I am interested as long as Rebel may play on tournament level!
#: Yes I take your challenge, Rebel will win a ten game match... :)
#
#You hve the machines. However, back the trolly up. YOu were claiming
#that 2K nodes per second was better than them in 1992. Why do you suddenly
#want to go 50x faster? IE, give Crafty 250 seconds per move, and Rebel
#5, and that should be a close approximation to what was happening in 1992.
#However, I'll take the odds at 3 mins vs 150 mins. Of course, you do know
#you have to not think on opponent's time... If you can win at that handicap,
#I'd likely go back to writing compilers... :)
Not thinking on opponent's time sounded weird to me at first, but I
understand why now. All the same, maybe it would be better to have
Crafty on the PPro 200 and Rebel on a 386-SX16 (maybe this is about 50
times slower than a PPro 200? Anyway, use a machine about 50 times
slower..) since then you don't have to turn either permanent brain off,
and ponder time happens.
Or, if Ed is happier with running Rebel on a PPro 200 as well, maybe
change it to 45:1 (would really need to depend on Rebel's ability to
guess Crafty's moves..) to compensate. Of course, this ignores
Crafty's pondering as well.. :/ Hmm.
Ponder-wise, if Rebel guessed 50% of Crafty's moves, it would get
double the time 50% of the time, and regular time 50% of the time. We
could just make this one and a half times normal all the time, so (now
with uniform platform:) 3 minutes vs. 150 minutes becomes 4.5 minutes
vs. 150 minutes (perhaps better is to have the match 3 minutes vs 100
minutes, I suppose.) At any rate, there's still a very large "Total
Nodes Search" difference to replicate the NPS difference if Rebel was
running on much slower hardware than Crafty.
If Crafty's guess rate of Rebel's moves was also 50% then this
argument can be similarly applied.. but Crafty would go from 3 minutes
vs 150 minutes to 3 minutes vs 225 minutes! Applying both adjustments
gives 3 minutes vs 150 minutes, which is what is expected, because my
estimate for each program's "opponent's next move guess rate" is 50%.
If the two were different then things would slide one way or another.
So it seems that roughly speaking, no permanent brain for either is
just fine. Caveat: it should be clear to you by now that I am
stumbling through this.. you may deduce that I have not yet had my
morning 2 litres of java :)
#: KK, were are you?
#: We need you!
#
#: I will make it even more interesting, the latest version of Crafty
#: against the latest version of Rebel on a PP-200. Rebel at tournament
#: level and Crafty 100 x more time, so 100:1 instead of 50.
#
#: The KK NPS Kup...
#: NPS 70,000 against NPS 7,000,000, I love your idea!
#: NPS rules... your chance to prove it!
#
#Somewhat broken test, of course... unless someone has a P6/200 with something
#like 64 gigs if memory, otherwise hashing is going to effectively turned off
#in Crafty. However, I'd still take the odds... You should talk to Lonnie
#first, however. They are playing reasonably closely at equal time controls in
#the games I've seen on ICC and Chess.net...
Well, Ed has a few PPros, I think.. he can probably dump in some RAM
from other machines into the Crafty machine, and take some out of the
Rebel machine, but assuming he has 8Mb and 16Mb chips available, and
the motherboard requires pairs, it's unlikely he can do better than
Rebel machine 16Mb, Crafty machine 64Mb.. would be much better for
Crafty machine to have 512Mb or whatever, but c'est la vie.
#: [ snip ]
#
#
#: : The only conclusion I can draw is you think the DB guys are
#: : incompetent...
#
#: Hey Bob don't put words in my mouth I never said!!
#
#but it's clearly implied. You want to play Crafty, but you want to use
#the full factor of 50 we've gotten in the past 5 years. And you claim it
#is important. And then you claim that the much larger performance gains
#made by the DB guys is worth only 50 rating points. The only conclusion
#I can see is that they must be incompetent, if you can get so much from a
#factor of 50, but they get nothing from an additional factor of 2,000...
#Right?
#
#
#: We have a discussion about ply depth / NPS versus playing strength, you
#: state "NPS rules..." which is speculative, I only counter attack all
#: your "maybe", "most likely", "probably" with the naked truth, the FACTS.
#
#No, reread what I wrote. I said that for two programs with an adequate
#search, and an adequate evaluation, that NPS *does* rule. Otherwise, why
#are you running on a P6/200 instead of a 386/16? Our major point of
#disagreement is the following:
#
#I claim their search is *better* than yours. That it makes fewer tactical
#errors than yours (note that you can take "yours" and replace with "mine" and
#it doesn't change my opinion here.) I claim their evaluation is better than
#yours, because I know what you can do, and run at the speed you are running
#at, while their hardware eval can do as much as they care to without their
#worrying about slowing things down at all...
Rebel has a pretty darn good selective search.. Ed has stated here (in
r.g.c.c) that he doesn't use null-move. His work in this area
probably will help some.. whether it will help enough is the question
we're trying to answer, of course. :)
#You've done well with Rebel, and it plays well, but if you *really* believe
#you would win 4 out of every 10 games against them, your judgement is really
#clouded. Perhaps by not understanding what they are doing, but I can read
#your mind.
Did you mean "...can't read your mind" ?
#However, see my post from last night. Let's see Rebel make any progress on
#that position, because that was solved by DT II in 1991, when you are lucky
#to do 1K nodes per second.
#
#
#
#: The current facts are:
#: - Madrid 1992, a 2,000 NPS micro won against 150,000 NPS;
#
#one game remember, just one game. Cray Blitz lost to "Bobby" in Cologne
#Germany due to a bug I introduced. It won the rest and the tourament, but
#*one game* doesn't mean a lot. Try 'em in a match if we can ever get Hsu
#to play on ICC or whatever.
#
#: - Hong Kong 1995, a 100,000 NPS micro won against 3,000,000 NPS;
#: - 16.c4??, just a 12 ply tactics.
#
#: I don't say things I can't prove, fact is you did with 32.Rc6
#: I don't fight Hsu, I fight against YOUR CLAIMS!
#: Quite a difference.
#
#But you fight with feathers. You claim 50 points. Can you *prove* that?
#of course not. Relevant questions: (1) when was the last time you beat
#Kasparov in a reasonable time control? (2) when was the last time you drew
#Kasparov in a reasonable time control? (3) can you show an even record against
#GM's in any sort of standard time control to come anywhere near justifying that
#you are within 50 points of them? (4) have you ever beaten them OTB yourself?
#(5) do you have any test suites that suggest you are close? Nolot? Can you
#approach what DT did?
#
#And with 5 out of 5 "no" answers, you would still claim to be "right up there
#with them." That's difficult for me, a scientist, to fathom. I've been able
#to observe them, I've been able to observe Rebel, Crafty, Mchess Pro, Hiarcs,
#and the rest. and in my opinion, based on observed facts, none are close.
Dave Gomboc
drgo...@a.stu.athabascau.ca
Moritz Berger
On 18 Feb 1997 02:30:29 GMT, hy...@crafty.cis.uab.edu (Robert Hyatt)
wrote:
>Moritz Berger (Moritz...@msn.com) wrote:
>: Chessmaster 5000 play 27. .. c5 after about 2 or 3 minutes on my
>: P5/166 ... It needs about 2 or 3 million nodes to see it. Of course,
>: this is not enough to justify a +2 eval (too much speculation or
>: intelligence involved, too less exact "knowledge", therefore only a
>: meagre 0.14 eval ...), but my point is, that micros can play the best
>: moves without calculating zillions of nps. Point taken?
>
>It's not hard to play c5, as the following shows:
<snip>
> 9-> 57.25 -0.141 c5 Be4 d5 Bc2 c4 bxc4 dxc4 Nc5 Bxc5
> Bxc5 Rd7
<snip>
>Of course, a big point here is that Crafty also does bishop mobility which
>makes c5 inteesting. However, note that after a minute (no big hash or
>anything however) it has no clue. And it won't have a clue for 20+
>more plies that this is actually winning.
>
>that was the point of my DB post, that Deep Thought actually saw the whole
>forcing line that none of us could figure out at the time because it was too
>deep. I remember Valvo looking at it too, and scratching his head.
Here are the eval from CM 5000:
9 ply: +0.14
10 ply: +0.39
11 ply: +0.45
(other moves slightly below 0, this means that c5 is worth half a pawn
for Chessmaster!)
-------------
Moritz...@msn.com
Robert Hyatt
Dave Gomboc (drgo...@a.stu.athabascau.ca) wrote:
: In article <5e9l02$e...@juniper.cis.uab.edu>,
: #: Yes I am interested as long as Rebel may play on tournament level!
: #: Yes I take your challenge, Rebel will win a ten game match... :)
: #
: #You hve the machines. However, back the trolly up. YOu were claiming
: #that 2K nodes per second was better than them in 1992. Why do you suddenly
: #want to go 50x faster? IE, give Crafty 250 seconds per move, and Rebel
: #5, and that should be a close approximation to what was happening in 1992.
: #However, I'll take the odds at 3 mins vs 150 mins. Of course, you do know
: #you have to not think on opponent's time... If you can win at that handicap,
: #I'd likely go back to writing compilers... :)
: Not thinking on opponent's time sounded weird to me at first, but I
: understand why now. All the same, maybe it would be better to have
: Crafty on the PPro 200 and Rebel on a 386-SX16 (maybe this is about 50
: times slower than a PPro 200? Anyway, use a machine about 50 times
: slower..) since then you don't have to turn either permanent brain off,
: and ponder time happens.
: Or, if Ed is happier with running Rebel on a PPro 200 as well, maybe
: change it to 45:1 (would really need to depend on Rebel's ability to
: guess Crafty's moves..) to compensate. Of course, this ignores
: Crafty's pondering as well.. :/ Hmm.
Problem with this is that if the "fast search time" program predicts,
it will search just as deep as the "long search time" program. It
makes the results more complicated to interpret. Much easier to
simply turn it off, which is almost mandatory in time-odds games...
: Ponder-wise, if Rebel guessed 50% of Crafty's moves, it would get
: double the time 50% of the time, and regular time 50% of the time. We
: could just make this one and a half times normal all the time, so (now
: with uniform platform:) 3 minutes vs. 150 minutes becomes 4.5 minutes
: vs. 150 minutes (perhaps better is to have the match 3 minutes vs 100
: minutes, I suppose.) At any rate, there's still a very large "Total
: Nodes Search" difference to replicate the NPS difference if Rebel was
: running on much slower hardware than Crafty.
: If Crafty's guess rate of Rebel's moves was also 50% then this
: argument can be similarly applied.. but Crafty would go from 3 minutes
: vs 150 minutes to 3 minutes vs 225 minutes! Applying both adjustments
: gives 3 minutes vs 150 minutes, which is what is expected, because my
: estimate for each program's "opponent's next move guess rate" is 50%.
: If the two were different then things would slide one way or another.
Note that if crafty predicts right, it gets nothing really, because it
would normall search for (using Ed's numbers) for 7.5 hours, plus the
time it used while Rebel is thinking (another 4-5 minutes) which is
basically noise in the overall total... For the side moving quickly,
however, it'd be a huge help and would distort the outcome at times...
: So it seems that roughly speaking, no permanent brain for either is
: just fine. Caveat: it should be clear to you by now that I am
: stumbling through this.. you may deduce that I have not yet had my
: morning 2 litres of java :)
: Well, Ed has a few PPros, I think.. he can probably dump in some RAM
: from other machines into the Crafty machine, and take some out of the
: Rebel machine, but assuming he has 8Mb and 16Mb chips available, and
: the motherboard requires pairs, it's unlikely he can do better than
: Rebel machine 16Mb, Crafty machine 64Mb.. would be much better for
: Crafty machine to have 512Mb or whatever, but c'est la vie.
Actually, he's volunteered to run the entire thing there using the
autoplay feature in crafty that works with no pondering allowed just
fine (it fails miserably however, if you don't disable pondering, for
reasons not yet isolated.)
The only other issue I've thought of is how to set the time controls.
There are two ways in crafty, but they operate much differently. The
easiest would be to set rebel to 40/120 for the primary control, and
set Crafty for 40/12000, which is 100 longer, and let them use that
time any way they want.
The second is to say st=N in crafty, but this behaves differently,
in that it will never step over N, even should it fail low and want
to. I did this to make problem suites consistent. Every now and then
someone reports a WAC result in some time per move N, but the program
actually goes over the limit of N before it finds the solution. That
can be confusing or misleading.
As a result, I don't use st=N for *anything* but test suites since
it disables a lot of the timing decisions Crafty applies when it is
searching. 40/12000 sounds like the right idea, as in
time 40/12000/20/6000
which would work fine and not break any time allocation code. I
might (were I Ed) scale both down from that, because it's likely they
will both go 20+ moves in book in most cases, assuming he uses the
normal big book for crafty, which means that the next 20 moves are
going to be ssllooooowwwwwwww.... But I'll leave that to his judgement.
: #No, reread what I wrote. I said that for two programs with an adequate
: #search, and an adequate evaluation, that NPS *does* rule. Otherwise, why
: #are you running on a P6/200 instead of a 386/16? Our major point of
: #disagreement is the following:
: #
: #I claim their search is *better* than yours. That it makes fewer tactical
: #errors than yours (note that you can take "yours" and replace with "mine" and
: #it doesn't change my opinion here.) I claim their evaluation is better than
: #yours, because I know what you can do, and run at the speed you are running
: #at, while their hardware eval can do as much as they care to without their
: #worrying about slowing things down at all...
: Rebel has a pretty darn good selective search.. Ed has stated here (in
: r.g.c.c) that he doesn't use null-move. His work in this area
: probably will help some.. whether it will help enough is the question
: we're trying to answer, of course. :)
I agree.. let me rewrite what you said however... Rebel has a pretty darn
good *selective* search. DB doesn't have a selective search at all, which
means there's never a chance that it throws out a move that is important to
the search. DB searches 4-6 plies deeper than Rebel or Crafty, yet it uses
*no* unsafe tricks to get that depth. So within it's search horizon, which
is really deep, it doesn't ever exclude a move that might win, or overlook
a move that might cause it to lose. I see games where *every* program that
uses selective forward pruning makes a tactical mistake here and there. I
won't begin to claim that Crafty doesn't, because the null move has it's own
unique set of problems. However, there are plenty of games vs Genius, Rebel,
WchessX, and so forth where crafty does a 10 ply search and returns an eval of
+.002, the opponent makes a move, and at 5 plies the eval is already up to
+3.x... I send these to Dave (WchessX) when I see them, so that he can
figure out what he pruned and perhaps fix it. One thing's for sure, however,
You won't find DB making that kind of mistake. It may well play a weak
positional move, but the bad thing is, it will see that it is losing many many
plies before Crafty/whatever would. If they don't make the weak moves, they
aren't going to fall for tactical mistakes. And it's very much like trying
to push the Chrysler building (in New York) around by yourself... it won't
budge because of what you do, although it might decide to fall down all by
itself...
: #You've done well with Rebel, and it plays well, but if you *really* believe
: #you would win 4 out of every 10 games against them, your judgement is really
: #clouded. Perhaps by not understanding what they are doing, but I can read
: #your mind.
: Did you mean "...can't read your mind" ?
arg... yep... :)
: #However, see my post from last night. Let's see Rebel make any progress on
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: About Crafty...
: Is Crafty now fully compatible with AUTO232?
: If so we can also keep the permament brain and set Rebel on 40/2:00
: and Crafty on 40 in 200 hours. In this case we will have 4-5 moves
: a day, 2.5 games a month.
> If you can do this, I believe crafty will work. It *will not work* with
> thinking on the opponent's time due to some quirk I have not found, but
> with thinking on the opponent's time disabled, I've had reports that it
> works well. Doing it like that is fine by me... I can send detailed
> instructions (and the latest version which appears to work with auto232
> without pondering) if you'd like...
Ok, I will try to see if Crafty will work with AUTO232 and PB=OFF.
I need some time because I have to finish my current matches first. Also
I want to have some kind of control that the match is played honestly
since it will be played in my office. I suggest that Bob and I will
agree on two volonteers who will check the moves and games.
: I am not interested in a 0:10 / 0:02 match.
: I don't like your idea because I am only interested in tournament level
: play. Fritz, Genius will win at blitz-level against Rebel but on
: tournament level... You know what I mean.
> no problem as you've proposed it. Assuming you don't mind the lost
> machine time there...
I do mind that I will miss one of my Autoplayers for months however the
match or expiriment is too interesting to let it pass.
- Ed Schroder -
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: Ed:
: Here's the game I mentioned between Cray Blitz and Deep Thought II. I
: believe (if my file naming convention is correct) that this was played
: in 1991. I think in Orlando, but don't remember exactly.
: In any case, when blacked played 27. ... c5, they did so with a +2 eval
: after a fail high. More details are below, but it was about 26 *plies*
: later that CB finally reached -2. Notice that... 26 *plies*. It is
: certainly ugly beyond move 27, but it does illustrate just how powerful
: that machine is. You can draw your own conclusions of course...
Very impressive of DT.
I analyzed it all.
Head off for DT/DB!
My findings are as follows:
Rebel (and other chess programs) will win this game too because all moves
after 27... c5 are very easy positional moves for black. No problem for
todays chess programs. Rebel8 on a PP-200 just finds 27... c5 on
tournament level, but it was close. After 27... c5 there are no problems
and Rebel8 find all the good moves very easily.
However no +2.xx for any Pc program I tested. The fact that DT announced
+2.xx for 27... c5 is really unbelievable. No Pc program will ever see
the 2.xx score. It's shows that DT to the extreme has used the singular
extension algorithm in these days.
Thanks Bob for showing us this game, you scored.
Well done.
I also saw somebody has posted the Nolut results of DT-II.
Very good, the more information the better.
- Ed Schroder -
Don Fong
Bob Hyatt statement (a):
|The only conclusion I can draw is you think the DB guys are incompetent...
Bob Hyatt statement (b):
|Therefore they *have* to be doing something that is *not very good*.
Bob, do you agree that (a) and (b) are far far different?
there is a semantic difference, and a connotational difference.
note that (a) is what you claimed at first, and is what Ed Schroder
objected to. but your supporting arguments only support (b) at
most. note that (b) is a much less inflammatory claim. do you disagree?
IMHO your supporting arguments don't even prove (b)
conclusively. i wish you'd drop this tack. whatever Ed Schroder
thinks of the DB team is his private business and is scientifically
irrelevant. sheesh, what a stupid thing to argue about. if you
are really so fascinated by the "what A thinks of B" type question,
why don't you watch the soaps or read the hollywood gossip columns.
if i didn't know better, i'd think you were trying to
duck Ed's challenge.
now, Ed didn't say it, but i will. IMHO the DB team is indeed
doing something that is *not very good*: they are hiding their machine.
why are we all having to speculate about it? because there are
practically no publicly available results. you're a believer,
and Ed's a skeptic. the empirical evidence is inconclusive so far.
(this isn't helped by the cryptic nature of some of Hsu's comments.
i've watched with amusement as you continually try to decipher
them and debate their meaning as if he were some kind of mystic.)
Ed didn't say this, but i will. it seems to me that the DB team
is attempting to maximize their surprise factor. i think that's
*not very good*. i don't think that's scientifically valid nor
is it sportsmanlike.
Ed didn't say this, but i will. if it is true that the
DB chips were only delivered shortly before the Kasparov match,
i think that is *not very good*. it suggests poor planning.
oddly, the fans of DB seem to regard this as further evidence
of the machine's power, the fact that it did so well despite
the implied lack of preparation.
the game is not won by having a higher IQ, or a higher ELO,
or by having a higher NPS. the game is won by playing the right moves.
until DB-REBEL actually happens, the score will be in doubt.
and until we see DB fight its way thru the candidates, as a
human challenger would have to, i won't believe they even deserve a
shot at Kasparov. even if they managed to beat Kasparov (which i
hope they won't), i would claim that a properly prepared Kasparov
would whip DB's ass. let a few good super GM's play it and i bet
they would find plenty of weaknesses. does anyone really believe
DB -would- make it thru the candidates?
the DB team has thousands of Kasparov games to examine and
prepare against. the Kasparov team has virtually zilch of DB
to prepare against. that gives DB an unfair advantage. furthermore,
it is an advantage that would never be given to a human challenger.
the human's greatest strength is the ability to learn. by
keeping DB under wraps, and by keeping the match short, they are
in effect minimizing this strength. why not DB-Nunn, DB-Browne,
DB-Kamsky, etc. thought experiment: why not give Kasparov a
chance to experiment with DB and train against it? ans: coz
he'd end up learning how to beat it every time, just like your
nephew learned to beat GNUCHESS every time.
if you believe the above, then it follows that DB team's *only*
hope is the surprise factor. that would explain why they keep it under
wraps like some kind of perpetual motion machine, or a magician's stunt.
IMHO that is *not very good*.
Robert Hyatt
Moritz Berger (Moritz...@msn.com) wrote:
: On 18 Feb 1997 02:30:29 GMT, hy...@crafty.cis.uab.edu (Robert Hyatt)
Just for fun, have you ever seen that thing think it was worse, unless it
is material down? It seems (to me) that it is always somewhat happy in the
games it plays... Most programs are, in fact, for whatever reasons. It's
fun in an ACM event to see both sides at +.5 in a game... you'd never guess
that chess is "zero-sum" if you wander around looking at the evals. :)
BTW, yes, I've also seen both at -.5 or so as well, particularly an early
game with Deep Thought. They had failed low seeing we had a really strong
move, but we couldn't search deep enough on that strong move and failed low
ourselves thinking it was bad. Funny. Maybe. If you aren't an author,
anyway.. :)
Tom C. Kerrigan
Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
> No. not what I meant at all. If you play the "right" move but don't see anything
> good as a result, you just flipped a coin N times, and on the right move it happened
> to turn up heads. Which do you want to do, play the right move for a random
> positional reason, or play the right move because you see it solves a tactical
> (and very deep) problem? There's one WAC position that Crafty solves positionally
> and which is about .06 better than the previous best move. A couple of more plies
> and it sees a +2.7 or so, which is the right solution.
Okay, good, I know exactly what you mean, but I still have a "moral"
problem with it, which is that if a program selects a move, it has its
reasons. Perhaps it doesn't find the mate in 500, but it plays the right
move because, say, it undoubles a pawn. That's basically the point of
eval(). It's just a cute little shortcut we use so we don't have to search
the entire bloody game tree before moving. Considering the quality of
eval() functions these days (i.e., awful) I too would trust the search
more, but I don't discount moves my program makes simply because they
don't win material...
Cheers,
Tom