Deep Blue vs Micros -- an interesting result just available
Robert Hyatt
Here's an interesting bit of Deep Blue trivia, that Hsu reported to a Newsweek
reporter, and later to the NY Times. I asked if he minded if I repeated it, he
responded "it is public knowledge". Chris, Ed, Mclane, Vincent, and the other
Deep Blue bashers had best tune this out at this point.
Please tune out *NOW*.
You were warned. :)
Early this year, Hsu and group were testing the new hardware evaluation they
have. Hsu didn't report this, but I have heard from a couple of reliable
sources that the DB "sorry evaluation" now has over 1,000 different adjustable
parameters. I wonder how that compares with Rebel, Genius, Hiarcs, Nimzo and
the like? I *know* how it compares to Crafty, and it is at least 10X more
than I'm doing, maybe more.
Hsu used a special version of deep blue, that used *exactly* one chess
processor, and this chess processor was clocked down to 1/10th of its normal
execution speed for this test. The idea was to play a match against two top
commercial programs on a "equal platform". Mclane? You still here? You
*really* should leave at this point. :)
So, he held a 10 game match, normal time control, 5 games vs Genius, 5 games
vs Rebel 8. I'd really love to stop at this point and play "guess the result".
Remember, this is "shallow blue" at about 100K nodes per second or so, vs Rebel8
and Genius on a P6/200, which is just what some of the Deep Blue bashers have
been screaming for, for a couple of years. The results of the match... 10-0.
And for those with learning impairment, DB (SB?) did *not* lost a single game.
I've ask Hsu to see if he can find the game scores so they can be posted and
analyzed. I know, only 10 games. What does that show? I'd say at 10-0 it
certainly shows that DB's eval isn't a lightweight eval, since speed was not
an issue. Hsu reported that some of the games were positional slaughters,
the some were tactical kingside attacks that ended in slaughters.
I hope we will get more data like this from the DB guys. I think it at least
puts to rest the noise from Mclane, and Ed, and most anyone else that thinks
that DB is fast, but that's all. Remember, this was DB at roughly equal hardware
speed with Rebel/Genius. They normally are 2,000 times faster than this. This
simply lends more credibility to the result against Kasparov. As I've said before,
they are *strong*.. *very strong*.
The commercial guys have been waiting to get a chance at them on a server. I'd
not be quite so interested in seeing this were I them, now... However, I think
we are going to see "DB Junior" show up before long, at "only" 10M nodes per
second or so. Wouldn't it be something if it doesn't lose a game? For those
not aware, log on to ICC and "hist scratchy". It still has the best win/loss
ratio of any player on the server... at blitz, 141 wins, 1 loss, 4 draws. At
Standard, 374 wins, 14 losses, 37 draws, mostly against strong GM players with
a few against strong IM's...
The evidence is mounting that this is not only the best computer player in the
world, it is probably the best computer player in the world by a huge amount.
I now revise my 200 Elo to 400. And I'm afraid that even 400 might be low.
Discussion or questions? I can only answer what I know, and will prompt
Hsu and group for specifics as needed. Please don't deluge them with queries
about this. I'd rather help field questions and pass on those that are of
interest, rather than running them into a deep hole with hundreds of messages.
Adam Sundor
David Purdy (dpu...@fas.harvard.edu) wrote:
: The date of the issue was Tuesday, May 15, 1997, and has a copy of the printout
: showing a little of what happened when DB chose 36. axb5 over Qb6. This is
: fairly interesting.
May I suggest that Tuesday was May 13, 1997 and it's on Page B2, Column 1?
-snipped the rest but
: strengthen DB, not to mention hardware improvements. The human brain isn't
: developing quite so spectacularly...
But some have and can. The limitations may be external as opposed to
internal. The opposite of computers.
Adam Sundor
Pete Nielsen
Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
> have. Hsu didn't report this, but I have heard from a couple of reliable
> sources that the DB "sorry evaluation" now has over 1,000 different adjustable
> parameters.
> Discussion or questions? I can only answer what I know, and will prompt
> Hsu and group for specifics as needed. Please don't deluge them with queries
> about this. I'd rather help field questions and pass on those that are of
> interest, rather than running them into a deep hole with hundreds of messages.
>
There's been rumors of marketing a "Baby Blue" PC card with one or more
of the special purpose chess processors. If they do that, will they also
market a "tweekers kit" so that we can tweek those 1000 parameters?
brucemo
Don Getkey wrote:
> I doubt any GM could go 10-0-0 against these two very strong programs!
> With this kind of performance, IBM could try going on tour with DB Jr.
> defeating GM after GM, in effect telling the humans beat our "little guy"
> in match play first, if you can, and prove yourself eligible to play
> the/our World Champion Deep Blue. Then they could take the pool of
> players who beat DB Jr. have a Interzonal of sorts, with the winner
> getting a shot at Deep Blue! This could be an entirely seperate
> competition that could suplant all other high level human vs human
> tournaments, as being THE deciding measuring stick of who is the best
> chess player in the world. Kasparov would be furiously ill. But that's
> too bad because, after all, it would be out of his impious remarks that
> something like this would fall upon him.
I think this is interesting but I would still like to see the game scores,
and know more about this, for instance the time control.
I'm surprised that anything going 100K nps can beat a professional program,
also going about 100K nps, 10-0.
bruce
David Purdy
The date of the issue was Tuesday, May 15, 1997, and has a copy of the printout
showing a little of what happened when DB chose 36. axb5 over Qb6. This is
fairly interesting.
Another question:
There's been discussion that DB is for science and discussion that it's for
chess domination. For a little while, humans may have the ability to eke out a
few wins; and, true, this time may be extended if it is widely known what DB
thinks about. But, with the reality that public, scientific discourse can only
help to improve DB, and the reality that it or a successor will be the best
chess player there is, why not just open up all information on it for public
review? The only negatives would be in the short term, as GMs could use the
information in games against DB, but countering their patterns can only
strengthen DB, not to mention hardware improvements. The human brain isn't
developing quite so spectacularly...
--David Purdy
ShaktiFire
Thanks Bob for your post.
Wow,,, unbelievable...10 - 0 pastings against the commercials
at the same NPS rate. Very impressive. Lets see all the details.
Don Getkey
In article <5m26mj$1...@juniper.cis.uab.edu>, hy...@crafty.cis.uab.edu
(Robert Hyatt) writes:
>And for those with learning impairment, DB (SB?) did *not* lost a single
>game.
>I've ask Hsu to see if he can find the game scores so they can be posted
and
>analyzed. I know, only 10 games. What does that show? I'd say at 10-0
it
>certainly shows that DB's eval isn't a lightweight eval, since speed was
not
>an issue. Hsu reported that some of the games were positional
slaughters,
>the some were tactical kingside attacks that ended in slaughters.
>
>
Awesome! Finally the answer we all have waited for!
Oh, and BTW, good to see you posting again Bob.
I doubt any GM could go 10-0-0 against these two very strong programs!
With this kind of performance, IBM could try going on tour with DB Jr.
defeating GM after GM, in effect telling the humans beat our "little guy"
in match play first, if you can, and prove yourself eligible to play
the/our World Champion Deep Blue. Then they could take the pool of
players who beat DB Jr. have a Interzonal of sorts, with the winner
getting a shot at Deep Blue! This could be an entirely seperate
competition that could suplant all other high level human vs human
tournaments, as being THE deciding measuring stick of who is the best
chess player in the world. Kasparov would be furiously ill. But that's
too bad because, after all, it would be out of his impious remarks that
something like this would fall upon him.
yours in chess,
Don
Ramsey MN USA
Dan Thies
On 22 May 1997 19:27:15 GMT, hy...@crafty.cis.uab.edu (Robert Hyatt)
wrote:
>Early this year, Hsu and group were testing the new hardware evaluation they
>have. Hsu didn't report this, but I have heard from a couple of reliable
>sources that the DB "sorry evaluation" now has over 1,000 different adjustable
>parameters. I wonder how that compares with Rebel, Genius, Hiarcs, Nimzo and
>the like? I *know* how it compares to Crafty, and it is at least 10X more
>than I'm doing, maybe more.
I've got about 1300 positional concepts encoded in my program, but to
be able to do it with that kind of speed......
I GIVE UP. ;-)
Dan
Robert Hyatt
ShaktiFire (shakt...@aol.com) wrote:
: Thanks Bob for your post.
: Wow,,, unbelievable...10 - 0 pastings against the commercials
: at the same NPS rate. Very impressive. Lets see all the details.
I agree. I've asked Hsu to see if he can track down the games. He
wasn't super-optimistic. I'm lazy myself, but can rely on icc/chess.net
to email me pgn games from crafty's play, and I just collect them in a
huge pgn file. I don't keep any of the many test games I play vs other
programs, however... I don't consider them reliable in any sense since
I am generally testing something and want to see how it looks against a
known opponent before letting it rip on ICC...
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: Here's an interesting bit of Deep Blue trivia, that Hsu reported to a
: Newsweek reporter, and later to the NY Times. I asked if he minded
: if I repeated it, he responded "it is public knowledge". Chris, Ed,
: Mclane, Vincent, and the other Deep Blue bashers had best tune this
: out at this point.
: Please tune out *NOW*.
: You were warned. :)
: Early this year, Hsu and group were testing the new hardware evaluation
: they have. Hsu didn't report this, but I have heard from a couple of
: reliable sources that the DB "sorry evaluation" now has over 1,000
: different adjustable parameters. I wonder how that compares with
: Rebel, Genius, Hiarcs, Nimzo and the like? I *know* how it compares
: to Crafty, and it is at least 10X more than I'm doing, maybe more.
: Hsu used a special version of deep blue, that used *exactly* one chess
: processor, and this chess processor was clocked down to 1/10th of its
: normal execution speed for this test. The idea was to play a match
: against two top commercial programs on a "equal platform". Mclane?
: You still here? You *really* should leave at this point. :)
: So, he held a 10 game match, normal time control, 5 games vs Genius,
: 5 games vs Rebel 8. I'd really love to stop at this point and play
: "guess the result".
No you won't, this is your victory Bob... :))
You are most welcome, it's your day!
: Remember, this is "shallow blue" at about 100K nodes per second or
: so, vs Rebel8 and Genius on a P6/200, which is just what some of the
: Deep Blue bashers have been screaming for, for a couple of years.
: The results of the match... 10-0.
Awfull!!!!!!!!!!!
Tell Hsu to forbid the commercial plans for Deep Blue!
: And for those with learning impairment, DB (SB?) did *not* lost a
: single game. I've ask Hsu to see if he can find the game scores so
: they can be posted and analyzed. I know, only 10 games. What does
: that show? I'd say at 10-0 it certainly shows that DB's eval isn't
: a lightweight eval, since speed was not an issue. Hsu reported that
: some of the games were positional slaughters, the some were tactical
: kingside attacks that ended in slaughters.
I am not sure if I still like Hsu at this moment... :)
: I hope we will get more data like this from the DB guys. I think it
: at least puts to rest the noise from Mclane, and Ed, and most anyone
: else that thinks that DB is fast, but that's all.
The "noise" you mention was caused by the silence of IBM.
So speculation started, what can you expect?
Well, it was fun most of the time, no?
: Remember, this was DB at roughly equal hardware speed with
: Rebel/Genius. They normally are 2,000 times faster than this.
: This simply lends more credibility to the result against Kasparov.
Yes.
: As I've said before, they are *strong*.. *very strong*.
: The commercial guys have been waiting to get a chance at them on
: a server. I'd not be quite so interested in seeing this were I
: them, now... However, I think we are going to see "DB Junior" show
: up before long, at "only" 10M nodes per second or so. Wouldn't it
: be something if it doesn't lose a game? For those not aware, log on
: to ICC and "hist scratchy". It still has the best win/loss
: ratio of any player on the server... at blitz, 141 wins, 1 loss,
: 4 draws. At Standard, 374 wins, 14 losses, 37 draws, mostly against
: strong GM players with a few against strong IM's...
: The evidence is mounting that this is not only the best computer
: player in the world, it is probably the best computer player in the
: world by a huge amount.
: I now revise my 200 Elo to 400. And I'm afraid that even
: 400 might be low.
So you were wrong after all and all the time... :))
: Discussion or questions? I can only answer what I know, and will
: prompt Hsu and group for specifics as needed. Please don't deluge
: them with queries about this. I'd rather help field questions and
: pass on those that are of interest, rather than running them into a
: deep hole with hundreds of messages.
Thanks for info Bob, you (unfortunately) won the discussion...
So now we finally know that DB is the strongest program.
10-0 says enough.
One question remains, you mentioned "normal time control".
I assume that means 40/2:00?
This ends the discussion between expert-1 and expert-2.
Expert-1 won, who ever that may be... :)
Now who should I congratulate?
Hsu and co or Bob?
Can't make up my mind as I am unable to congratulate both... :)
- Ed Schroder -
Ingo Althoefer
Robert Hyatt wrote:
: Here's an interesting bit of Deep Blue trivia, that Hsu reported to a
: Newsweek reporter, and later to the NY Times. ....
: Hsu used a special version of deep blue, that used *exactly* one chess
: processor, and this chess processor was clocked down to 1/10th of its
: normal execution speed for this test. The idea was to play a match
: against two top commercial programs on a "equal platform". So, he held a
: 10 game match, normal time control, 5 games vs Genius, 5 games vs Rebel 8.
: Remember, this is "shallow blue" at about 100K nodes per second or
: so, vs Rebel8 and Genius on a P6/200, which is just what some of the
: Deep Blue bashers have been screaming for, for a couple of years.
: The results of the match... 10-0.
Congratulations to the Deep Blue team ! That is really impressive.
Is it possible to publish the notations of these ten games ?
: .... Hsu reported that
: some of the games were positional slaughters, the some were tactical
: kingside attacks that ended in slaughters.
: ... However, I think we are going to see "DB Junior" show
: up before long, at "only" 10M nodes per second or so. ....
: The evidence is mounting that this is not only the best computer
: player in the world, it is probably the best computer player in the
: world by a huge amount.
I am eager to try against DB-Junior with my 3-Hirns and Double-Fritzes and
Triple-Hiarcs' and Multi-Shredders and List-3-Hirns.
: Discussion or questions? I can only answer what I know, and will
: prompt Hsu and group for specifics as needed. Please don't deluge
: them with queries about this. I'd rather help field questions and
: pass on those that are of interest, rather than running them into a
: deep hole with hundreds of messages.
I would like to have available on internet a DB or DB-Junior with a 2-best
mode where the internet user may have the final choice amongst the proposals
of DB (2-best). And I would like to play with such a Double-Deep-Blue & Boss
against the pure Deep Blue.
Lets go for chess on an Elo 3000 level !
Men and Women and Mechine together !
Ingo Althoefer.
chrisw
--
http://www.demon.co.uk/oxford-soft
brucemo <bru...@seanet.com> wrote in article <338531...@seanet.com>...
> Don Getkey wrote:
>
> > I doubt any GM could go 10-0-0 against these two very strong programs!
> > With this kind of performance, IBM could try going on tour with DB Jr.
> > defeating GM after GM, in effect telling the humans beat our "little
guy"
> > in match play first, if you can, and prove yourself eligible to play
> > the/our World Champion Deep Blue. Then they could take the pool of
> > players who beat DB Jr. have a Interzonal of sorts, with the winner
> > getting a shot at Deep Blue! This could be an entirely seperate
> > competition that could suplant all other high level human vs human
> > tournaments, as being THE deciding measuring stick of who is the best
> > chess player in the world. Kasparov would be furiously ill. But
that's
> > too bad because, after all, it would be out of his impious remarks that
> > something like this would fall upon him.
>
> I think this is interesting but I would still like to see the game
scores,
> and know more about this, for instance the time control.
>
> I'm surprised that anything going 100K nps can beat a professional
program,
> also going about 100K nps, 10-0.
I'm more than surprised. In fact I challenge these figures and challenge
that the games either exist or gave these results.
Seems Zeus has been going hubristic ever since the DB win, and his posting
style announcing these results was so arrogant, aggressive and overbearing,
that I smell a large rat.
Where's the bloody evidence, then ?
I say bollocks.
Chris Whittington
>
> bruce
>
Robert Hyatt
Ed Schroder (rebc...@xs4all.nl) wrote:
: No you won't, this is your victory Bob... :))
: You are most welcome, it's your day!
It is not *my* day. I'm on *your* side of the fence, here, remember?
If they beat you 5-0, they will beat me 5-0. I don't like it either. :)
: : Remember, this is "shallow blue" at about 100K nodes per second or
: : so, vs Rebel8 and Genius on a P6/200, which is just what some of the
: : Deep Blue bashers have been screaming for, for a couple of years.
: : The results of the match... 10-0.
: Awfull!!!!!!!!!!!
Yes. But not unexpected from my perspective. Remember, this group is
filled with bright minds, as I've said before. They are all Ph.D.'s
which means they not only know how to hack programs, but they know how
to do structured research and make intelligent decisions about what to
do next. And there are several of them. The rest of us are like
small islands, each working on our own program. They are like a
small continent...
: Tell Hsu to forbid the commercial plans for Deep Blue!
Somehow I doubt this will happen. Although I'd encourage him to send me
a sample of one of his 8=processor boards for a VME bus machine. :)
I think a PC version would be a nightmare with all the different processor
speeds, bus timings, etc.
: : And for those with learning impairment, DB (SB?) did *not* lost a
: : single game. I've ask Hsu to see if he can find the game scores so
: : they can be posted and analyzed. I know, only 10 games. What does
: : that show? I'd say at 10-0 it certainly shows that DB's eval isn't
: : a lightweight eval, since speed was not an issue. Hsu reported that
: : some of the games were positional slaughters, the some were tactical
: : kingside attacks that ended in slaughters.
: I am not sure if I still like Hsu at this moment... :)
I like Hsu. I don't think much of DB right now. :)
: : I hope we will get more data like this from the DB guys. I think it
: : at least puts to rest the noise from Mclane, and Ed, and most anyone
: : else that thinks that DB is fast, but that's all.
: The "noise" you mention was caused by the silence of IBM.
: So speculation started, what can you expect?
: Well, it was fun most of the time, no?
at times, yes. Just that the "assumption" about almost "no evaluation"
was way off-base as I had suspected, and then as I had been told, although
more or less "in private."
: : Remember, this was DB at roughly equal hardware speed with
: : Rebel/Genius. They normally are 2,000 times faster than this.
: : This simply lends more credibility to the result against Kasparov.
: Yes.
: : As I've said before, they are *strong*.. *very strong*.
: : The commercial guys have been waiting to get a chance at them on
: : a server. I'd not be quite so interested in seeing this were I
: : them, now... However, I think we are going to see "DB Junior" show
: : up before long, at "only" 10M nodes per second or so. Wouldn't it
: : be something if it doesn't lose a game? For those not aware, log on
: : to ICC and "hist scratchy". It still has the best win/loss
: : ratio of any player on the server... at blitz, 141 wins, 1 loss,
: : 4 draws. At Standard, 374 wins, 14 losses, 37 draws, mostly against
: : strong GM players with a few against strong IM's...
: : The evidence is mounting that this is not only the best computer
: : player in the world, it is probably the best computer player in the
: : world by a huge amount.
: : I now revise my 200 Elo to 400. And I'm afraid that even
: : 400 might be low.
: So you were wrong after all and all the time... :))
Actually, I've been more wrong that you might guess. Remember that
I predicted 4-2 kasparov. How wrong that turned out to be. They were
much better than I expected, even though Hsu had sent me lots of tidbits
that should have warned me. :)
: : Discussion or questions? I can only answer what I know, and will
: : prompt Hsu and group for specifics as needed. Please don't deluge
: : them with queries about this. I'd rather help field questions and
: : pass on those that are of interest, rather than running them into a
: : deep hole with hundreds of messages.
: Thanks for info Bob, you (unfortunately) won the discussion...
No... again, I lose here too. I just suspected what that particular
group of people were capable of...
: So now we finally know that DB is the strongest program.
: 10-0 says enough.
: One question remains, you mentioned "normal time control".
: I assume that means 40/2:00?
I believe so, yes...
: This ends the discussion between expert-1 and expert-2.
: Expert-1 won, who ever that may be... :)
: Now who should I congratulate?
: Hsu and co or Bob?
: Can't make up my mind as I am unable to congratulate both... :)
Not me. I'm unhappy as well. I can't possible do what they do on their
eval without dropping to 1K nodes per second or so. And at 1K, crafty
might be a positional genius, but it would also be a tactical idiot...
: - Ed Schroder -
Robert Hyatt
chrisw (chr...@cpsoft.demon.co.uk) wrote:
: --
: http://www.demon.co.uk/oxford-soft
: I say bollocks.
: Chris Whittington
It is not *my* story. I simply asked if it was ok to post that here and
got the response "it's been given to the NY times and Newsweek" already so
it is public.
As far as evidence, I'm not sure what would suffice. A game score? could
be a GM playing the DB moves. My only comment? The DB guys have *never*
fabricated data and made it public. I obviously can't do this as it is
trivial to confirm what I say Crafty can/can't do, since everyone has or
can get a copy.
To date, the *only* cases of odd behavior, fabricating results, fudging
test results, and fudging match results has been from the "commercial side."
*not* from the academic or research side. The reasons are likely very
obvious. But the idea of fabrication is unpleasant to those of us that do
research and report results...
If you don't believe, that's your right. I haven't been going "hubristic"
either. I felt they were the best in the world before they lost to Fritz
in Hong Kong, and nothing has changed my mind nor my posting style. All
the way back to that fateful day in 1995. My story hasn't changed. My
opinion of DB has, since it did much better than I thought against
Kasparov.
Let's see if Hsu can find the games. Or if he can play a 10 game match
using "junior" as a next-best case...
Jouni Uski
>
> : I now revise my 200 Elo to 400. And I'm afraid that even
> : 400 might be low.
>
This is complete nonsense - I will bet ANY sum of
money, that difference is lower than 400 points. Do You
Bob think, that DB plays with over 3000 rating ?????
Jouni
mclane
hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>Here's an interesting bit of Deep Blue trivia, that Hsu reported to a Newsweek
>reporter, and later to the NY Times. I asked if he minded if I repeated it, he
>responded "it is public knowledge". Chris, Ed, Mclane, Vincent, and the other
>Deep Blue bashers had best tune this out at this point.
>Please tune out *NOW*.
>You were warned. :)
>Early this year, Hsu and group were testing the new hardware evaluation they
>have. Hsu didn't report this, but I have heard from a couple of reliable
>sources that the DB "sorry evaluation" now has over 1,000 different adjustable
>parameters.
But Bob - what do they count for parameters ? If I change the
graphical clock into the digital clock in chris program, it changes
the playing style-. This is not a feature, it is a bug. But is is also
a parameter... counted HOW ?
> I wonder how that compares with Rebel, Genius, Hiarcs, Nimzo and
>the like? I *know* how it compares to Crafty, and it is at least 10X more
>than I'm doing, maybe more.
I don't know how long the source code of Rebel/Genius/Hiarcs... is,
but I know how big Chris' engine-code is.
But - I have not overread that this is HARDWARE evaluation.
If they would open their work some more, we can discuss about.
But this is too rough.
>Hsu used a special version of deep blue, that used *exactly* one chess
>processor, and this chess processor was clocked down to 1/10th of its normal
>execution speed for this test. The idea was to play a match against two top
>commercial programs on a "equal platform". Mclane?
Right ! I am still here. Have not used any word you forbid me to use.
Also my pulse is 60 or 50. Decreasing.
I will drink a tea, with milk - of course.
> You still here? You
>*really* should leave at this point. :)
No. Don't worry. I have worked at Schach Niggemann, I was with Friedel
in Cologne 1986, I spoke with Jaap in Den Haag Aegon 1997, I had to
operate CSTal against Sofia Polgar, had to operate Hiarcs-CSTal in
Paderborn. I can stand stress situations very well. And cool.
Ask Bruce if HE was MORE in stress than I was in Aegon97.
>So, he held a 10 game match, normal time control, 5 games vs Genius, 5 games
>vs Rebel 8. I'd really love to stop at this point and play "guess the result".
>Remember, this is "shallow blue" at about 100K nodes per second or so, vs Rebel8
>and Genius on a P6/200, which is just what some of the Deep Blue bashers have
>been screaming for, for a couple of years. The results of the match... 10-0.
>And for those with learning impairment, DB (SB?) did *not* lost a single game.
Show me the games ! I cannot work with rough data like: 10:0.
10:0 means: you (genius) tried 10 times to make her a baby. But failed
10 times. Now the games show: the reason.
Maybe she used a condom. Maybe he was not errected. Maybe it was...
Whatever.. Show the games.
>I've ask Hsu to see if he can find the game scores so they can be posted and
>analyzed. I know, only 10 games. What does that show? I'd say at 10-0 it
>certainly shows that DB's eval isn't a lightweight eval, since speed was not
>an issue. Hsu reported that some of the games were positional slaughters,
>the some were tactical kingside attacks that ended in slaughters.
OKOKOK. GAMES. If we replay the games, we can find out about it's
strengh.
>I hope we will get more data like this from the DB guys. I think it at least
>puts to rest the noise from Mclane, and Ed, and most anyone else that thinks
>that DB is fast, but that's all.
Maybe. But maybe not. EVIDENCE. FACTS said this man called Hyatt.
>Remember, this was DB at roughly equal hardware
>speed with Rebel/Genius.
How can he measure that the hardware platforms were the same.
How did the measure that both CPUs are as fast as the other ?
There is no need to find it out 100%.
30% would be enough for me. But how can we be sure 30% that the
hardware was the same ?
>They normally are 2,000 times faster than this. This
>simply lends more credibility to the result against Kasparov. As I've said before,
>they are *strong*.. *very strong*.
IF THEY HAVE SUCH A HARDWARE, they could also do some games against
Ferret or Rebel or Hiarcs here on the net in public. That we can all
see it.
But of course - with the same platform hardware.
BTW: you always said: this is not possible. Now you quote that it is
possible to test it on the same hardware . Point taken ?
>Discussion or questions? I can only answer what I know, and will prompt
>Hsu and group for specifics as needed. Please don't deluge them with queries
>about this. I'd rather help field questions and pass on those that are of
>interest, rather than running them into a deep hole with hundreds of messages.
You are the master of moderation and talks with them. We dont have the
background to talk with them.
Why not posting the 10 games first.
I am normally used to find out with one game.
And - although Chris normally shouts 4 days: it was only one game
mclane...! - I am in general right about the strength.
Vincent Diepeveen
>Here's an interesting bit of Deep Blue trivia, that Hsu reported to a Newsweek
>reporter, and later to the NY Times. I asked if he minded if I repeated it, he
>responded "it is public knowledge". Chris, Ed, Mclane, Vincent, and the other
>Deep Blue bashers had best tune this out at this point.
>
>Please tune out *NOW*.
>
>You were warned. :)
>
>Early this year, Hsu and group were testing the new hardware evaluation they
>have. Hsu didn't report this, but I have heard from a couple of reliable
>sources that the DB "sorry evaluation" now has over 1,000 different adjustable
>parameters. I wonder how that compares with Rebel, Genius, Hiarcs, Nimzo and
>the like? I *know* how it compares to Crafty, and it is at least 10X more
>than I'm doing, maybe more.
Evaluation parameters are something different than adjustable
parameters.
I have at the time of writing about 1500+ evaluation patterns (could be
some more, i just counted number of lines in header file
with the average number of patterns, and i did this few weeks ago), although
this is growing fast (few patterns every 2 weeks), as lately my biggest
tester: Jan Louwman, played another bunch of games with Diep against
other programs.
If i count the number of ADJUSTABLE patterns, then my first estimate
is: few tens of thousands, i must include the combinations (a pattern
can be used for a part depending on a different pattern); this
kind of adjustable patterns are growing terribly fast, and here is
where most of the evaluation bugs of Diep are in. I add new patterns,
before having tested previous patterns.
For example i try to find out whether there are open files, and then
have rules about whether an open file is a good file (for example
when there are heavy pieces of the opponent on it, doubled, tripled etc,
or when it is stopped by light pieces).
So pattern = open file, extra conditions can make it stronger
or lighter.
This does not include preprocessor, which i use for strategem purposes,
and where i so far did not have done much about. I suspect however
that preprocessor is not very important any more in Diep. It just
initialize a small part of the patterns.
Also i could increase the number of some adjustable parameters too if i
multiply the bonuses by the number of array elements (bonus depending on
the square file or line where piece remains)... :)
It still doesn't say much 1000, but it probably is more than Genius has
during search (i'm not talking about the preprocessor where Genius probably
has many thousands of adjustable patterns).
>Hsu used a special version of deep blue, that used *exactly* one chess
>processor, and this chess processor was clocked down to 1/10th of its normal
>execution speed for this test. The idea was to play a match against two top
>commercial programs on a "equal platform". Mclane? You still here? You
>*really* should leave at this point. :)
>So, he held a 10 game match, normal time control, 5 games vs Genius, 5 games
>vs Rebel 8. I'd really love to stop at this point and play "guess the result".
>Remember, this is "shallow blue" at about 100K nodes per second or so, vs Rebel8
>and Genius on a P6/200, which is just what some of the Deep Blue bashers have
>been screaming for, for a couple of years. The results of the match... 10-0.
5-0 against Rebel 8 is huge.
With that far over thousand patterns i don't get 100k nodes a second
however, but only 10k-12k pro second. From which only a small part is
evaluations. MOre important is that at important moments Diep only searches
8/9 ply, and less important moments 9-11 ply.
It probably does however indicate that it plays with PCprograms at the
current hardware better at tournament level.
>And for those with learning impairment, DB (SB?) did *not* lost a single game.
>I've ask Hsu to see if he can find the game scores so they can be posted and
>analyzed. I know, only 10 games. What does that show? I'd say at 10-0 it
>certainly shows that DB's eval isn't a lightweight eval, since speed was not
>an issue. Hsu reported that some of the games were positional slaughters,
>the some were tactical kingside attacks that ended in slaughters.
I'm indeed surprised they did something on the evaluation, and even
in hardware!
My compliments for Hsu!
Seems they slowly begin to realize that depth is not the only thing
that matters!
>I hope we will get more data like this from the DB guys. I think it at least
>puts to rest the noise from Mclane, and Ed, and most anyone else that thinks
>that DB is fast, but that's all.
Just like most of us think: let's first meet em before we think they
are really well. Remember Deep Blue - Fritz?
You can call it lucky, just like i call most losses of Diep unlucky where Diep
is having a fine position and the gets mated somehow in auto232play,
let's see how their performance will develop. They can always join using
a different name Deep Thought III. Or Diep II (in Aegon tournament daily
i was asked whether my program was Deep Blue, people don't see the
difference between Diep and Deep Blue it appears, i should ask some
sponsoring of IBM!).
> Remember, this was DB at roughly equal hardware
>speed with Rebel/Genius. They normally are 2,000 times faster than this. This
>simply lends more credibility to the result against Kasparov. As I've said before,
>they are *strong*.. *very strong*.
When playing other programs probably, although i doubt it when
they search 17 ply and i 12 ply, then the difference becomes less than
when they search 15 ply and i search 8-10 ply (currently done by Diep,
Rebel at tournament level 1 ply more).
A problem against humans you must not forget
but something where i haven't done
enough about currently: closed positions.
I don't see how to solve this
problem when playing humans.
I lost 1.5 points because of this against low rated humans
(Jan Joost Lindner and Peter van Wermeskerken, both journalists).
Fritz simply tries to open files (also when it is bad) to prevent
this. Most use a good book.
>The commercial guys have been waiting to get a chance at them on a server. I'd
>not be quite so interested in seeing this were I them, now... However, I think
>we are going to see "DB Junior" show up before long, at "only" 10M nodes per
>second or so. Wouldn't it be something if it doesn't lose a game? For those
>not aware, log on to ICC and "hist scratchy". It still has the best win/loss
>ratio of any player on the server... at blitz, 141 wins, 1 loss, 4 draws. At
>Standard, 374 wins, 14 losses, 37 draws, mostly against strong GM players with
>a few against strong IM's...
>The evidence is mounting that this is not only the best computer player in the
>world, it is probably the best computer player in the world by a huge amount.
>I now revise my 200 Elo to 400. And I'm afraid that even 400 might be low.
This would mean that world top pc-programs have 1900 in your eyes?
Against some humans this is of course true, but most this is not.
>Discussion or questions? I can only answer what I know, and will prompt
>Hsu and group for specifics as needed. Please don't deluge them with queries
Please ask whether they have some testpositions where Deep Blue finds moves
we are not gonna find for a long time in their eyes. Please ask some of these
positions. Benjamin must have some of these positions. Would be nice
to test micros on it.
BTW how do they do on the few positions on my homepage?
>about this. I'd rather help field questions and pass on those that are of
>interest, rather than running them into a deep hole with hundreds of messages.
Vincent
--
+----------------------------------------------------+
| Vincent Diepeveen email: vdie...@cs.ruu.nl |
| http://www.students.cs.ruu.nl/~vdiepeve/ |
+----------------------------------------------------+
Vincent Diepeveen
Well let's not overestimate it; the facts:
a) Kasparov's moves were bad, it will remain a mystery why he played
so much beginners moves. On the other hand sometimes he plays also
unexpected moves which appear worldchampionlevel (thanks to v/d Herik
for this last otherhand suggestion...)
b) the moves played by Deep Blue were not great, but are reproduced
(also moves which i think bad, and really computerish) by pc-programs.
>: : Discussion or questions? I can only answer what I know, and will
>: : prompt Hsu and group for specifics as needed. Please don't deluge
>: : them with queries about this. I'd rather help field questions and
>: : pass on those that are of interest, rather than running them into a
>: : deep hole with hundreds of messages.
>: Thanks for info Bob, you (unfortunately) won the discussion...
Let's be concrete, a 1000 is probably more than Fritz and Genius have,
and for the time being this will probably crush with their speed the
programs. I'm far over this 1000, so are some others (i suspect most
commercial programs and most strong amateurs),
which only makes Deep Blue program a program like we have, with
the big difference that they are 2000 times faster.
Now winning from Deep Blue will mean that search depth is not the
only thing.
If this is done within few years, then we will know definitely
that
a) there is a tactical barrier
b) what that depth is.
Hopefully there will be some event they play with this machine, or
slower version of it. Hopefully for the PC. Would buy it immediately
to play my program (with a pc-card i can do it under Memphis/95 without
problems, as the interface of a PC-card would leave Diep
with almost the full systemtime of the PC).
>No... again, I lose here too. I just suspected what that particular
>group of people were capable of...
>: So now we finally know that DB is the strongest program.
>: 10-0 says enough.
>
>: One question remains, you mentioned "normal time control".
>: I assume that means 40/2:00?
>I believe so, yes...
>: This ends the discussion between expert-1 and expert-2.
>: Expert-1 won, who ever that may be... :)
>
>: Now who should I congratulate?
>: Hsu and co or Bob?
>: Can't make up my mind as I am unable to congratulate both... :)
>Not me. I'm unhappy as well. I can't possible do what they do on their
>eval without dropping to 1K nodes per second or so. And at 1K, crafty
>might be a positional genius, but it would also be a tactical idiot...
You are right... ...my evaluation has dropped
to 10k-12k nodes a second in middlegame. That's quite little.
To speed up evaluation for some pieces i have hashtable to speed it up.
Pawn hashtable speeds up a part of the pawn-evaluation considerably.
I also use bitboards for pawn evaluation. Other pieces i cannot use
bitboards however that would slow it down again. Perhaps you should
strip some of the bitboards, or add a second datastructure?
A lot more in endgame however, did not program that much for endgame.
Different evaluation for rook-endings and pawnendings are coming,
another few tens new patterns which within few years will probably
be extended to hundreds of adjustable parameters too, for these 2 evaluation
functions, which at the time of writing are primitive so that
search speed in these endings is about 30k-40k.
>
>: - Ed Schroder -
>
>
Vincent Diepeveen
Andreas De Troy
Bob,
I remember your furious reaction earlier this year when someone (was it
Tom K.?) claimed that his program had beaten Crafty 10 (?) times in a row.
You said this was impossible; that this would mean that this program would
have an impossible ELO-rating etc. .
You were absolutely right, but I think you are doing the same thing now. A
scaled-down version of Deep Blue would have beaten Genius and Rebel 10
times in 10 games? To say it with your own words:
-I do not believe this
-Publish the games
-You have *no* right to make these 10-0 claims if you "cannot find
the games anymore".
[Apart from that, I liked your recent post about move-ordening, like so
many others. It's only when Deep Thought is involved that you... maybe you
ARE a computer indeed.]
--
Beste groeten vanwege
Andreas De Troy
Andreas...@ped.kuleuven.ac.be
Don Fong
In article <5m3m4v$d9b$1...@thor.wirehub.nl>,
Ed Schroder <rebc...@xs4all.nl> wrote:
>From: hy...@crafty.cis.uab.edu (Robert Hyatt)
>: Remember, this is "shallow blue" at about 100K nodes per second or
>: so, vs Rebel8 and Genius on a P6/200, which is just what some of the
>: Deep Blue bashers have been screaming for, for a couple of years.
>
>: The results of the match... 10-0.
>
>Awfull!!!!!!!!!!!
>
>This ends the discussion between expert-1 and expert-2.
>Expert-1 won, who ever that may be... :)
but kibitzer-1 says, "i want to see games, not rumors".
i want to see the games PLAYED, not just REPORTED third hand.
>Now who should I congratulate?
>Hsu and co or Bob?
i'd hold those congratulations until we have more than
third hand rumors.
>Can't make up my mind as I am unable to congratulate both... :)
Ed, in my book *you* deserve the congratulations.
*if* it turns out Bob was correct, it may not have been because
of superior insight, but because he had "insider" knowledge that
he didn't share fully. (perhaps confidential info from Hsu.)
Bob was continually trying to steamroller people by handwaving,
blustering, insulting, and rumor mongering. but the facts that were
publicly available at the time did not justify Bob's claims.
you did an admirable job of pointing out the holes in his "reasoning".
actually, the facts that are publicly available *still* don't
justify Bob's claims. and of course the folks who were being so
secretive in the first place don't deserve your congratulations.
just my 2c.
Tord Kallqvist Romstad
vdie...@cs.ruu.nl (Vincent Diepeveen) writes:
> To speed up evaluation for some pieces i have hashtable to speed it up.
> Pawn hashtable speeds up a part of the pawn-evaluation considerably.
Are pawn hash tables really a good idea? I haven't experimented with them
myself, but I don't really like the idea. A pawn structure which is bad in
one position, might be good if the pieces are played differently, or if a
few pieces are traded off. A pawn structure evaluation that doesn't consider
piece placement seems too unaccurate to me. I regard pawn hash tables as
something similar to piece square tables --- a cheap substitute to a
knowledgable eval. I am surprised that a programmer like Vincent, who believes
in knowledge, uses pawn hash tables.
I am very interested in hearing what other programmers think about this.
Who uses pawn hash tables? No flames please, what I have written above is
just my own very humble opinion, and I have not tried pawn hash tables myself.
Tord
chrisw
--
http://www.demon.co.uk/oxford-soft
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
<5m455o$q...@juniper.cis.uab.edu>...
Fine, but what time controls ?
Who says its at 100,000 nps ?
And and and
Where are the games. Lets see it. Lets see the data. Lets see the logs.
Lets see the EVIDENCE.
Lets NOT have any more Hyattian secret weapons like "I know the data, but I
can't tell you, but it kills you, even though you can't see it, but its
brilliant even though you can't see it"
These are stone tablets passed on by god and seen only by the true believer
and you expect the rest of us to just swallow the idea ? Huh ?
Bollocks.
First Zeus, now you're being Moses. I suppose god gave you the 11th
commandment all written down on stone, but you dropped it, eh ?
Chris Whittington
>
>
Robert Hyatt
mclane (mcl...@prima.ruhr.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
: >Here's an interesting bit of Deep Blue trivia, that Hsu reported to a Newsweek
: >reporter, and later to the NY Times. I asked if he minded if I repeated it, he
: >responded "it is public knowledge". Chris, Ed, Mclane, Vincent, and the other
: >Deep Blue bashers had best tune this out at this point.
: >Please tune out *NOW*.
: >You were warned. :)
: >Early this year, Hsu and group were testing the new hardware evaluation they
: >have. Hsu didn't report this, but I have heard from a couple of reliable
: >sources that the DB "sorry evaluation" now has over 1,000 different adjustable
: >parameters.
: But Bob - what do they count for parameters ? If I change the
: graphical clock into the digital clock in chris program, it changes
: the playing style-. This is not a feature, it is a bug. But is is also
: a parameter... counted HOW ?
Here is how I would count them. This is what I believe *they* would
count as well. Values stuck in piece square tables don't count, although
you might count a complete piece/square table as "1" if you like. Crafty
has 12. They currently don't change during the course of a game, they are
primarily centralization mechanisms that are very fast to access for that
info. If you look at evaluate.h for Crafty, you will find a bunch of
different evaluation terms that can be modified, and then the engine
recompiled, and you have a new crafty personality. I count about 100
such terms. Bruce says that I have more than he does. I'm sure others
have more than I do. But > 1,000 is a truly large number. And then they
have developed tools to help tune this against hundreds of thousands of
GM vs GM games...
: > I wonder how that compares with Rebel, Genius, Hiarcs, Nimzo and
I've asked. Whether they are available or not is another story. If not,
maybe we can get a repeat of this, maybe even live on the net somewhere
so we can watch...
: >I've ask Hsu to see if he can find the game scores so they can be posted and
: >analyzed. I know, only 10 games. What does that show? I'd say at 10-0 it
: >certainly shows that DB's eval isn't a lightweight eval, since speed was not
: >an issue. Hsu reported that some of the games were positional slaughters,
: >the some were tactical kingside attacks that ended in slaughters.
: OKOKOK. GAMES. If we replay the games, we can find out about it's
: strengh.
: >I hope we will get more data like this from the DB guys. I think it at least
: >puts to rest the noise from Mclane, and Ed, and most anyone else that thinks
: >that DB is fast, but that's all.
: Maybe. But maybe not. EVIDENCE. FACTS said this man called Hyatt.
: >Remember, this was DB at roughly equal hardware
: >speed with Rebel/Genius.
: How can he measure that the hardware platforms were the same.
: How did the measure that both CPUs are as fast as the other ?
: There is no need to find it out 100%.
: 30% would be enough for me. But how can we be sure 30% that the
: hardware was the same ?
Hsu reported that DB was searching about 100K nodes per second. The
hardware *was not* the same of course... because they were doing their
full eval in hardware still, and still running that fast. I'd suspect
they would slow down by a factor of 10X if they had to do it in software
like we have to. Maybe even slower than that, maybe a factor of 50X.
But what he had, was two programs searching about the same number of
nodes per second. *something* made a huge difference. if their eval
is lousy, then something is wrong. If their eval is very strong, this
result makes perfect sense. I'd hate to play Crafty (no eval) vs
Crafty (full eval) at long time controls. I'm sure it would end up at
0-10 as well. I simply conclude, as I've always said, that their eval
is not dumb, just because their search is fast. That's the magic of
hardware. There are *no* limits to what can be done in the eval, and
it doesn't have to cost any search time at all... *we* don't have that
option and have to weigh depth vs smarts all the time. That's why in
one game Rebel smothers Crafty, even though it is way faster (the NPS
match) because his eval seems better, but then in another game (the
Korrespondence Kup) Crafty finds a tactical shot to win material, so that
my faster search won out over a smarter eval. It's give and take. With
DB it is *all* take... :)
: >They normally are 2,000 times faster than this. This
: >simply lends more credibility to the result against Kasparov. As I've said before,
: >they are *strong*.. *very strong*.
: IF THEY HAVE SUCH A HARDWARE, they could also do some games against
: Ferret or Rebel or Hiarcs here on the net in public. That we can all
: see it.
: But of course - with the same platform hardware.
: BTW: you always said: this is not possible. Now you quote that it is
: possible to test it on the same hardware . Point taken ?
No. You wanted DB on a PC. I said that was impossible. They *have* to
have their hardware. This was probably a test-bed for Hsu to check out
a single chip's functionality. He divided the clock by 10, and only used
one processor. But it is still DB hardware through and through. It sounds
like what you wanted, but it isn't... If you moved *everything* to the
PC, it would take a huge performance hit, because their eval is nearly free
time wise right now. It wouldn't be on a PC...
: >Discussion or questions? I can only answer what I know, and will prompt
Robert Hyatt
chrisw (chr...@cpsoft.demon.co.uk) wrote:
: --
: http://www.demon.co.uk/oxford-soft
: Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
Standard.
: Who says its at 100,000 nps ?
Hsu. I suppose he can count and measure time somehow... :)
: And and and
: Where are the games. Lets see it. Lets see the data. Lets see the logs.
: Lets see the EVIDENCE.
: Lets NOT have any more Hyattian secret weapons like "I know the data, but I
: can't tell you, but it kills you, even though you can't see it, but its
: brilliant even though you can't see it"
: These are stone tablets passed on by god and seen only by the true believer
: and you expect the rest of us to just swallow the idea ? Huh ?
: Bollocks.
: First Zeus, now you're being Moses. I suppose god gave you the 11th
: commandment all written down on stone, but you dropped it, eh ?
I suppose we have to agree to disagree. I only reported what I "know".
perhaps we'll get to see them play some before long. How many games
before you become convinced? Or will we have to somehow suffer though
the "GM in a box" discussion as well?
Robert Hyatt
Don Fong (df...@cse.ucsc.edu) wrote:
: In article <5m3m4v$d9b$1...@thor.wirehub.nl>,: >: Remember, this is "shallow blue" at about 100K nodes per second or
: >: Deep Blue bashers have been screaming for, for a couple of years.
: >
: >: The results of the match... 10-0.
: >
: >
: [...]
: >This ends the discussion between expert-1 and expert-2.
: >Expert-1 won, who ever that may be... :)
: but kibitzer-1 says, "i want to see games, not rumors".
: i want to see the games PLAYED, not just REPORTED third hand.
: >Now who should I congratulate?
: >Hsu and co or Bob?
: i'd hold those congratulations until we have more than
: third hand rumors.
: >Can't make up my mind as I am unable to congratulate both... :)
: Ed, in my book *you* deserve the congratulations.
: *if* it turns out Bob was correct, it may not have been because
: of superior insight, but because he had "insider" knowledge that
: he didn't share fully. (perhaps confidential info from Hsu.)
: Bob was continually trying to steamroller people by handwaving,
: blustering, insulting, and rumor mongering. but the facts that were
: publicly available at the time did not justify Bob's claims.
: you did an admirable job of pointing out the holes in his "reasoning".
: actually, the facts that are publicly available *still* don't
: justify Bob's claims. and of course the folks who were being so
: secretive in the first place don't deserve your congratulations.
: just my 2c.
Hmm.. fact... DB and predecessors have won 9 out of every 10 games
against contemporary (at the time of each game) micro programs.
fact... DB just beat the best human player in the world in a 6 game
match.
fact... no micro program has ever beaten a player of that class at
a long time control, much less in a match.
weaker fact... Hsu reported a 10-0 score against Rebel/Genius when
DB was running at 100K nodes per second. Weaker because I didn't witness
it. But I happen to know the group well. They *don't* fabricate data,
unlike some have. *I* don't report incorrect data intentionally. It's
too easy to be exposed.
So all the "facts" seem to support what I've said all along... It's
strong. *very* strong. Stronger than any other program by what now
appears to be a *huge* margin, of perhaps >400 Elo. More facts will
be forthcoming as they (if they) start popping in to the chess servers.
All will be just as extraordinary...
Robert Hyatt
Vincent Diepeveen (vdie...@cs.ruu.nl) wrote:
: This would mean that world top pc-programs have 1900 in your eyes?
No... DB would clearly be a 2700+ player now, although it's actual
TPR must be around 2850 or so with the Kasparov match. So that leaves
the micros right where I've always said they are.. on the lower end of
the IM scale.. at 2300 or so...
: Against some humans this is of course true, but most this is not.
: >Discussion or questions? I can only answer what I know, and will prompt
: >Hsu and group for specifics as needed. Please don't deluge them with queries
: Please ask whether they have some testpositions where Deep Blue finds moves
: we are not gonna find for a long time in their eyes. Please ask some of these
: positions. Benjamin must have some of these positions. Would be nice
: to test micros on it.
I agree. I'll see if they might release any, although they do whack on the
"Nolot" positions and get most or all all pretty quickly, based on results
by the "old" hardware.
: BTW how do they do on the few positions on my homepage?
No idea...
Robert Hyatt
Jouni Uski (Jouni...@semitechturku.com) wrote:
: >
: > : 400 might be low.
: >
: This is complete nonsense - I will bet ANY sum of
: money, that difference is lower than 400 points. Do You
: Bob think, that DB plays with over 3000 rating ?????
: Jouni
Absolutely not. I think the micros play at 2300 or 2350.
Most humans would agree. On the lower end of the IM scale
at present.
Robert Hyatt
Vincent Diepeveen (vdie...@cs.ruu.nl) wrote:
: You are right... ...my evaluation has dropped
: to 10k-12k nodes a second in middlegame. That's quite little.
: To speed up evaluation for some pieces i have hashtable to speed it up.
: Pawn hashtable speeds up a part of the pawn-evaluation considerably.
: I also use bitboards for pawn evaluation. Other pieces i cannot use
: bitboards however that would slow it down again. Perhaps you should
: strip some of the bitboards, or add a second datastructure?
When I started Crafty, I made a promise with myself that I was going
to "stay the course" for five years. To give me enough time to learn how
to use bitmaps in the most effective way, and to give the hardware a chance
to grow up into the bitmap-friendly world. That is happening. Bitmaps don't
take a performance hit at present, with the rotated stuff I developed for
Crafty. When the 64bit machines hit the street in quantity (not counting
the alpha/ new HP chip and so forth) the bitmaps pick up even more steam.
And with new alpha instructions for counting bits and finding the first one
bit (done for Cray Research now SGI I believe) there is nothing wrong with
them. I have yet to see a better way to generate captures. No loops for
sliding pieces, etc.
Robert Hyatt
Andreas De Troy (Andreas...@ped.kuleuven.ac.be) wrote:
: Bob,
: I remember your furious reaction earlier this year when someone (was it
: Tom K.?) claimed that his program had beaten Crafty 10 (?) times in a row.
: You said this was impossible; that this would mean that this program would
: have an impossible ELO-rating etc. .
: You were absolutely right, but I think you are doing the same thing now. A
: scaled-down version of Deep Blue would have beaten Genius and Rebel 10
: times in 10 games? To say it with your own words:
: -I do not believe this
: -Publish the games
: -You have *no* right to make these 10-0 claims if you "cannot find
: the games anymore".
Note that my argument about Stobor was based on about 50 games I have
in my database here. So there I had strong evidence that something was
simply wrong about the testing. No knowledge of whether it was deliberate
or accidental, but something was wrong. I've personally watched too many
games vs Stobor to accept that result.
With DB, it's a different matter. Notice that I was also a little careful
about the conclusions to draw from this. My main point is that at equal
speed, they won all 10 games. So *clearly* they aren't worse. Whether they
are better (which this seems to support) is difficult to say after only 10
games, but going 10-0 and having a 2,000X speed increase in their pocket,
the conclusion is pretty obvious to me...
: [Apart from that, I liked your recent post about move-ordening, like so
: many others. It's only when Deep Thought is involved that you... maybe you
: ARE a computer indeed.]
My greatest wish is that some biologist, working somewhere, ultimately
finds out that our DNA and everything about us is actually nothing more
complex than 0101110001... Wouldn't it be great if *we* were ultimately
proven to be "digital"? :) Then we'd *know* what computers are really
capable of, since we'd be the "example case."
: --
chrisw
--
http://www.demon.co.uk/oxford-soft
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
<5m4peo$l...@juniper.cis.uab.edu>...
Several, against varied opponents. And some evidence that its at 100,000
nps at EQUIVALENT nodal content.
Saying 100K is fine and righty-right, would be ok if DB was a Fritz. But if
it has expemsive knowledge, then it needs to be at the Nps of an expensive
knowledge machine. That could be 10K to 20K.
Lets make this test fair and reasonable, not just picking numbers from the
sky.
> Or will we have to somehow suffer though
> the "GM in a box" discussion as well?
Uh, Uh, tricky comment. people might think I said there was a GM between DB
and the move output. Now you know I didn't - so why suggest it ?
Chris Whittington
>
>
Robert Hyatt
Tord Kallqvist Romstad (tor...@refil.ifi.uio.no) wrote:
: vdie...@cs.ruu.nl (Vincent Diepeveen) writes:
: > To speed up evaluation for some pieces i have hashtable to speed it up.
: > Pawn hashtable speeds up a part of the pawn-evaluation considerably.
: Are pawn hash tables really a good idea? I haven't experimented with them
: myself, but I don't really like the idea. A pawn structure which is bad in
: one position, might be good if the pieces are played differently, or if a
: few pieces are traded off. A pawn structure evaluation that doesn't consider
: piece placement seems too unaccurate to me. I regard pawn hash tables as
: something similar to piece square tables --- a cheap substitute to a
: knowledgable eval. I am surprised that a programmer like Vincent, who believes
: in knowledge, uses pawn hash tables.
They are nearly critical... but think about the evaluation a little
differently:
Part 1: This is simple pawn structure. Open files, backward or weak
pawns, isolated pawns, pawn majorities, mobile pawns, passed pawns,
connected pawns, and so forth. Evaluate all of that and hash it. If
you want to do like I do, and have the isolated pawns score more dynamic
(for example, an isolated pawn on an open file is weaker if there are
rooks on the board) then you simply hash store the number of such pawns
in the hash record, and after looking it up, fold it in when you evaluate
a rook. Much of the pawn structure is independent of pieces, and this
will become essentially free to compute, because 99% pawn hash hits is
a normal result. If you want to do detailed analysis for king safety,
do it in the pawn evaluation. Figure out which files are open, on which
king. and even do it for both sides of the board since you won't do it
twice for the same position. Now, when you evaluate king safety, just look
where your king is and decide whether to use the kingside pawn shelter
analysis, the queenside analysis, or the central analysis.
All in all, it simply lets you evaluate anything that is piece independent
*completely free*, since one evaluation and you will use that result from
then on... Passed pawns can only be partially handled like this because
they are typically based on the number of pieces on the board, whether they
are blockaded or not, and so forth. But at least you can hash the file
numbers where you have passed pawns, so that in the positional eval you
can avoid evaluating anything but passed pawns, which will be very fast
when there are none of course...
: I am very interested in hearing what other programmers think about this.
: Who uses pawn hash tables? No flames please, what I have written above is
: just my own very humble opinion, and I have not tried pawn hash tables myself.
Glad you said that... Had my torch lit. :) (just kidding).
I've used them forever. From before Cray Blitz to the present. If you want
to do some cute majority analysis, you can. And it won't slow you done one
tiny bit, so long as what you analyze is piece-independent. *pay* attention
to that. I've screwed it up more than once. If you factor in a piece's
influence, you will get wrong results if your hash key only considers the
locations of pawns, which is what most of us do...
: Tord
Robert Hyatt
chrisw (chr...@cpsoft.demon.co.uk) wrote:
: Several, against varied opponents. And some evidence that its at 100,000
: nps at EQUIVALENT nodal content.
That is not going to happen I don't think. I already know there eval is
quite complex, as I explained. Many wanted to see how it would do at
equal speed. This is about as equal as it gets, and as I pointed out
already, it is *not* at PC speeds, because their eval is still done at
1/10th light-speed here. So no, it is not comparable. But it still gives
a lot of insight into what it can do, and then when you factor in a
speedup of 2,000 more, it becomes pretty serious as a chess player.
: Saying 100K is fine and righty-right, would be ok if DB was a Fritz. But if
: it has expemsive knowledge, then it needs to be at the Nps of an expensive
: knowledge machine. That could be 10K to 20K.
or maybe even 1K with what they are doing. But then again, they wouldn't
do what they are doing if they had to live within the confines of a PC,
would they?
: Lets make this test fair and reasonable, not just picking numbers from the
: sky.
This wasn't a test for us. Just something run in the lab, and then reported
as being of general interest. From this, I conclude that their eval is much
more sophisticated than Rebel or Genius, because at the same NPS rate, they
ran away with the match. The only thing is, at the same NPS. So what's
different is the "content" of those nodes, which, as you suspect along with
me is quite a bit more than what we would see in a "rebel node" or a "genius
node" or whatever.
We have to take what data we can get about the machine. If you study this
from all angles, you can at least make deductions about the complexity, if
not the content of their eval. *something* has to be working to produce
that result...
: > Or will we have to somehow suffer though
: > the "GM in a box" discussion as well?
: Uh, Uh, tricky comment. people might think I said there was a GM between DB
: and the move output. Now you know I didn't - so why suggest it ?
Didn't mean to imply you did, and if it was taken as such, I simply state
that was not the intent. The intent was that even with 10 game listings,
there'd be no concrete evidence it was DB playing and not a GM. And we'd
see that argument surface all over again.
: Chris Whittington
: >
: >
Athanor IX
hy...@crafty.cis.uab.edu (Robert Hyatt) said:
>>My greatest wish is that some biologist, working somewhere, ultimately
>>finds out that our DNA and everything about us is actually nothing more
>>complex than 0101110001... Wouldn't it be great if *we* were ultimately
>>proven to be "digital"? :) Then we'd *know* what computers are really
>>capable of, since we'd be the "example case."
Crick and Watson discovered exactly this in the 1950s. The
difference is that the DNA/RNA code is base 4, not base 2. For
interesting comparisons of DNA and computer data storage, see
"Blind Watchmaker" and "The Selfish Gene" both by Richard
Dawkins.
Athanor IX.
Moritz Berger
On 22 May 1997 19:27:15 GMT, hy...@crafty.cis.uab.edu (Robert Hyatt)
wrote:
>
>Here's an interesting bit of Deep Blue trivia, that Hsu reported to a Newsweek
>reporter, and later to the NY Times. I asked if he minded if I repeated it, he
>responded "it is public knowledge". Chris, Ed, Mclane, Vincent, and the other
>Deep Blue bashers had best tune this out at this point.
Please count me among them :-)
>Please tune out *NOW*.
nope.
>You were warned. :)
>
>Early this year, Hsu and group were testing the new hardware evaluation they
>have. Hsu didn't report this, but I have heard from a couple of reliable
>sources that the DB "sorry evaluation" now has over 1,000 different adjustable
>parameters. I wonder how that compares with Rebel, Genius, Hiarcs, Nimzo and
>the like? I *know* how it compares to Crafty, and it is at least 10X more
>than I'm doing, maybe more.
It's like counting eval-nodes, nobody knows e.g. how many of those
terms are 1st order, 2nd order etc. chess knowledge. You're back to
materialistic counting parameters (instead of counting NPS like
before).
>Hsu used a special version of deep blue, that used *exactly* one chess
>processor, and this chess processor was clocked down to 1/10th of its normal
>execution speed for this test. The idea was to play a match against two top
>commercial programs on a "equal platform". Mclane? You still here? You
>*really* should leave at this point. :)
Show us the games ! Until then, your results mean nothing to me,
sorry.
>So, he held a 10 game match, normal time control, 5 games vs Genius, 5 games
>vs Rebel 8. I'd really love to stop at this point and play "guess the result".
>Remember, this is "shallow blue" at about 100K nodes per second or so, vs Rebel8
>and Genius on a P6/200, which is just what some of the Deep Blue bashers have
>been screaming for, for a couple of years. The results of the match... 10-0.
10-0 for whom? ;-) sorry, couldn't resist ...
>And for those with learning impairment, DB (SB?) did *not* lost a single game.
>I've ask Hsu to see if he can find the game scores so they can be posted and
>analyzed. I know, only 10 games. What does that show? I'd say at 10-0 it
>certainly shows that DB's eval isn't a lightweight eval, since speed was not
>an issue. Hsu reported that some of the games were positional slaughters,
>the some were tactical kingside attacks that ended in slaughters.
PGN files are welcome.
>I hope we will get more data like this from the DB guys. I think it at least
>puts to rest the noise from Mclane, and Ed, and most anyone else that thinks
>that DB is fast, but that's all. Remember, this was DB at roughly equal hardware
>speed with Rebel/Genius. They normally are 2,000 times faster than this. This
>simply lends more credibility to the result against Kasparov. As I've said before,
>they are *strong*.. *very strong*.
Agreed, they are strong. But just how strong still remains to be seen,
at least for me.
>The commercial guys have been waiting to get a chance at them on a server. I'd
>not be quite so interested in seeing this were I them, now... However, I think
>we are going to see "DB Junior" show up before long, at "only" 10M nodes per
>second or so. Wouldn't it be something if it doesn't lose a game? For those
>not aware, log on to ICC and "hist scratchy". It still has the best win/loss
>ratio of any player on the server... at blitz, 141 wins, 1 loss, 4 draws. At
>Standard, 374 wins, 14 losses, 37 draws, mostly against strong GM players with
>a few against strong IM's...
I'm really looking forward to see it on the servers, *THIS* will
really end all speculation.
>The evidence is mounting that this is not only the best computer player in the
>world, it is probably the best computer player in the world by a huge amount.
>I now revise my 200 Elo to 400. And I'm afraid that even 400 might be low.
This is idle speculation.
>Discussion or questions? I can only answer what I know, and will prompt
>Hsu and group for specifics as needed. Please don't deluge them with queries
>about this. I'd rather help field questions and pass on those that are of
>interest, rather than running them into a deep hole with hundreds of messages.
Game scores, PVs and evals, description of eval terms (e.g. the most
unusual things they do in the eval?), information whether they do a
full eval at all nodes etc. ...
Moritz
-------------
Moritz...@msn.com
"The truth will always come out in the end."
(Komputer Korner, 28/03/1997 08:22 on rec.games.chess.computer)
Howard Exner
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
<5m4q7a$l...@juniper.cis.uab.edu>...
> Vincent Diepeveen (vdie...@cs.ruu.nl) wrote:
>
> : This would mean that world top pc-programs have 1900 in your eyes?
>
> No... DB would clearly be a 2700+ player now, although it's actual
> TPR must be around 2850 or so with the Kasparov match. So that leaves
> the micros right where I've always said they are.. on the lower end of
> the IM scale.. at 2300 or so...
I wonder if Hergott and the Aegon participants would agree with this
estimate of the commercial programs. Those events provided us with
some usefull data. The 2300 estimate doesn't fit with these results.
Lonnie
On Thu, 22 May 1997 22:55:30 -0700, brucemo <bru...@seanet.com> wrote:
|Don Getkey wrote:
|
|> I doubt any GM could go 10-0-0 against these two very strong programs!
|> With this kind of performance, IBM could try going on tour with DB Jr.
|> defeating GM after GM, in effect telling the humans beat our "little guy"
|> in match play first, if you can, and prove yourself eligible to play
|> the/our World Champion Deep Blue. Then they could take the pool of
|> players who beat DB Jr. have a Interzonal of sorts, with the winner
|> getting a shot at Deep Blue! This could be an entirely seperate
|> competition that could suplant all other high level human vs human
|> tournaments, as being THE deciding measuring stick of who is the best
|> chess player in the world. Kasparov would be furiously ill. But that's
|> too bad because, after all, it would be out of his impious remarks that
|> something like this would fall upon him.
|
|I think this is interesting but I would still like to see the game scores,
|and know more about this, for instance the time control.
|
|I'm surprised that anything going 100K nps can beat a professional program,
|also going about 100K nps, 10-0.
|
I agree with Bruce. I find this *fact* a hard one to swallow.
같같같
Lonnie
So far as the laws of mathematics refer to reality,
they are not certain. And so far as they are certain, they
do not refer to reality. -- Albert Einstein,
**Put a "c" before the "ook" to respond to proper email address**
mclane
atha...@aol.com (Athanor IX) wrote:
>hy...@crafty.cis.uab.edu (Robert Hyatt) said:
>>>My greatest wish is that some biologist, working somewhere, ultimately
>>>finds out that our DNA and everything about us is actually nothing more
>>>complex than 0101110001...
Here speaks the materialist.
> Wouldn't it be great if *we* were ultimately
>>>proven to be "digital"? :) Then we'd *know* what computers are really
>>>capable of, since we'd be the "example case."
It is not that easy. The DATA info that is being done by particles (so
it doesn't matter if there is life involved or dead materia) is not
only working using 101010101. Particles exchange their whole
information pool, in an instant, without need of time, at any
distance, exchanging information about past and future.
Of course the same stuff happens in the brain. Because also here we
have particles (energy) switching arround and exchanging information.
But what a primitive computer is a latest computer in relation to the
human-body (or much better working: a wale-body/brain) ?
In each DNS we have a surrounding that makes the particles possible to
swtich and exchange information. The whole DNS in each unit of the
body is a little machine itself.
The human body is nothing more than a machine that should collect and
sort and manage DATA. Mainly finding higher quality of data. Because
quantity is cold. Only quality makes heat.
>Crick and Watson discovered exactly this in the 1950s. The
>difference is that the DNA/RNA code is base 4, not base 2. For
>interesting comparisons of DNA and computer data storage, see
> "Blind Watchmaker" and "The Selfish Gene" both by Richard
>Dawkins.
Their is a similarity. But we have one problem: It is not the storage,
it is the LOSE of data. Humans lose / forget / focus data meanwhile
computers do not qualify data, they can only quantify data.
If computers have a program, they can SIMULATE qualification
processes. But they still have no life themselfes. The key is not in
the similar data, the similar coding, it is in the 2 splitted brain,
the fact that humans manage to forget unimportant (quantity) of data
meanwhile they remember very good important stuff. But humans do not
remember important stuff by WILL, they access different.
>Athanor IX.
brucemo
Adam Sundor wrote:
>
> David Purdy (dpu...@fas.harvard.edu) wrote:
> : The date of the issue was Tuesday, May 15, 1997, and has a copy of the printout
> : showing a little of what happened when DB chose 36. axb5 over Qb6. This is
> : fairly interesting.
>
> May I suggest that Tuesday was May 13, 1997 and it's on Page B2, Column 1?
>
> -snipped the rest but
>
> : strengthen DB, not to mention hardware improvements. The human brain isn't
> : developing quite so spectacularly...
>
> But some have and can. The limitations may be external as opposed to
> internal. The opposite of computers.
>
> Adam Sundor
Do you guys know for sure which day this was? I went to the library thinking "May
13", and there was some interesting discussion, but no "copy" of anything. There's
a picture of Murry and Tan sitting next to a terminal, but the text on the terminal
is unreadable, and also the position on the terminal doesn't seem to be this
position.
I didn't check May 15. If there is really a "copy" of something in there, I'll go
back.
bruce
David Purdy
Bruce, et al.:
I was wrong. It is Tuesday May 13th. In my paper it is on p. A22. The small
copy of the printout is a little difficult to read, and someone more familiar
with what's going on in computer chess might have a better time deducing the
meanings of the data.
So, the main caveat here is that the page numbers might differ depending on if
you get the NYC version, the national version, the Boston version, etc. of the
Times. What's interesting is that they cite the source of the info as IBM,
rather than a particular person, so I wonder if this was a big release by IBM &
that more of the printout was released but the Times chose not to use it. Did
any other papers print it? Any journalists or folks w/ the Times who could
check?
--David
Daryl Lakes
In article <5m4qiv$l...@juniper.cis.uab.edu>, hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>Jouni Uski (Jouni...@semitechturku.com) wrote:
>: >
>: > : I now revise my 200 Elo to 400. And I'm afraid that even
>: > : 400 might be low.
>: >
>
>: This is complete nonsense - I will bet ANY sum of
>: money, that difference is lower than 400 points. Do You
>: Bob think, that DB plays with over 3000 rating ?????
>
>: Jouni
>
>Absolutely not. I think the micros play at 2300 or 2350.
>
>at present.
Mr Hyatt...if this is true how would you explain the Micros recent result at
Aegon? Over the last two years Rebel (and others) performance ratings have
been no way near 2300...And could a 2300 knock out an IM 4-2?
Marcel van Kervinck
chrisw (chr...@cpsoft.demon.co.uk) wrote:
: Who says its at 100,000 nps ?
: And and and
: Where are the games. Lets see it. Lets see the data. Lets see the logs.
: Lets see the EVIDENCE.
Yeah. And after that you'll demand print-outs of the grown
game trees. And after that you'll demand print outs including
evaluations in the leafes. And after that you'll demand these
evals factored. And after that you'll demand VHDL sources
(if any ofcourse).
Can't you people just don't wait 'till they publish?
Marcel
-- _ _
_| |_|_|
|_ |_ Marcel van Kervinck
|_| mar...@stack.nl
chrisw
--
http://www.demon.co.uk/oxford-soft
Marcel van Kervinck <mar...@stack.nl> wrote in article
<5m7str$g...@turtle.stack.nl>...
> chrisw (chr...@cpsoft.demon.co.uk) wrote:
>
> : Who says its at 100,000 nps ?
>
> : And and and
>
> : Where are the games. Lets see it. Lets see the data. Lets see the logs.
> : Lets see the EVIDENCE.
>
> Yeah. And after that you'll demand print-outs of the grown
> game trees. And after that you'll demand print outs including
> evaluations in the leafes. And after that you'll demand these
> evals factored. And after that you'll demand VHDL sources
> (if any ofcourse).
>
> Can't you people just don't wait 'till they publish?
Publishing would be fine.
It's the 'data' handed out via Moses on the tablets of stone that's the
problem.
Chris Whittington
Robert Hyatt
chrisw (chr...@cpsoft.demon.co.uk) wrote:
: --
: http://www.demon.co.uk/oxford-soft
: Publishing would be fine.
: Chris Whittington
Tell you what. From this point forward, you get *nothing* from me. I
simply relayed information that I believe to be factual and interesting.
You want to prattle about "Moses" and that crapola, feel free. If you
like living in a complete vacuum, you may now do so.
Sheesh... I've never seen such a bunch of babies... wanting information,
then trying to pick it apart when some is given. Is it any mystery why
the DB guys won't post here directly?
Sorry I fed your fire. It won't happen again...
Joe McCaughan
Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
: chrisw (chr...@cpsoft.demon.co.uk) wrote:
: : --
: : http://www.demon.co.uk/oxford-soft
: : Publishing would be fine.
: : Chris Whittington
Keep posting Bob. Although I find the 100 NPS arg a little
hard to believe I still think the posts are interesting.
All we have so far is hearsay. I guess we'll have to wait
until baby blue gets on a server and starts mopping the floor
with the micros... :)
Komputer Korner
Robert Hyatt wrote:
>
> Discussion or questions? I can only answer what I know, and will prompt
> Hsu and group for specifics as needed. Please don't deluge them with queries
> about this. I'd rather help field questions and pass on those that are of
> interest, rather than running them into a deep hole with hundreds of messages.
What is the full width search of Deep Blue in the middlegame.
I am guessing that it is 15-18 plies depending on the position. Is
this correct?
--
Komputer Korner
The inkompetent komputer.
mclane
>Yeah. And after that you'll demand print-outs of the grown
>game trees.
No - I don't need this information. I only need main-lines and evals
and search depth. I am used to these few parameters. More is not more
helpful.
mclane
hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>Tell you what. From this point forward, you get *nothing* from me. I
>simply relayed information that I believe to be factual and interesting.
THAT I BELIEVE TO BE FACTUAL !!!!
Bob - what you post is interesting. No doubt.
Also you post mass !
Also ok. Nice indeed.
But our discussion was about : HOW ACCURATE is the data you post.
Because you always claim it is 100% fact or it is proven long time
before whoever was born or these kind of stuff.
My only point was, that a "realist" like you, is not more than a
believer e.g. of the catholic-church.
You are beleving in what you define as FACTS. Others believe in UFO
and aliens from outer space. Others believe in catholics commandments
etc.
We are onle making our internal jokes because you do it alway and
always without seeing it, and many others don't see it . That is funny
for us, to see you present your BELIEVED data like a priest in the
church is open the golden shrine, raising his voice and quoting out of
the golden book.
We are not attacking you for this. It only amuses us, therefore the
images.
>Sheesh... I've never seen such a bunch of babies... wanting information,
>then trying to pick it apart when some is given.
A german princess has one day said, when she heard about the wishes of
the people in her country , starving because they had no bread/food:
Sollen sie doch Kuchen essen!
They shall eat cookies! What is their problem ?!
We do want information. Otherwise we would not subscribe this
newsgoup. Also we want to share this with other computerchess-freaks.
But we don't want wrong information or filtered information.
Is that so difficult to understand?
> Is it any mystery why
>the DB guys won't post here directly?
Do you really believe we would ATTACK them in person only because they
are that nice to present DATA ?
I have e.g. never said that THEY have cheated by manually making other
moves than DB has made. I would never claim this. I have respect for
the persons behind, although I don't like what they do. Why should I
believe that well-honored people cheat. With Kasparov it is different.
He has cheated in past.
>Sorry I fed your fire. It won't happen again...
NONO. You always misunderstand. It is not YOU. It is what you do.
mclane
shi...@best.com (Joe McCaughan) wrote:
>All we have so far is hearsay. I guess we'll have to wait
>until baby blue gets on a server and starts mopping the floor
>with the micros... :)
We do not complain about hearsaid or data that is NOT FACT or NOT
EVIDENT or roumor.
We are complaining about Bob saying what he says is more fact, better
data, and earlier theory.
That was the point of the whole discussion , that I wanted to explain
him that his posts are not COMMANDMENTS but also just cool or
spontaneous reactions and not scientific theories.
Rolf Tueschen
hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>chrisw (chr...@cpsoft.demon.co.uk) wrote:
>: --
>: http://www.demon.co.uk/oxford-soft
>: Marcel van Kervinck <mar...@stack.nl> wrote in article
>: <5m7str$g...@turtle.stack.nl>...
>: > chrisw (chr...@cpsoft.demon.co.uk) wrote:
>: >
>: > : Who says its at 100,000 nps ?
>: >
>: > : And and and
>: >
>: > : Where are the games. Lets see it. Lets see the data. Lets see the logs.
>: > : Lets see the EVIDENCE.
>: >
>: > Yeah. And after that you'll demand print-outs of the grown
>: > game trees. And after that you'll demand print outs including
>: > evaluations in the leafes. And after that you'll demand these
>: > evals factored. And after that you'll demand VHDL sources
>: > (if any ofcourse).
>: >
>: > Can't you people just don't wait 'till they publish?
>: Publishing would be fine.
>: It's the 'data' handed out via Moses on the tablets of stone that's the
>: problem.
>: Chris Whittington
>Tell you what. From this point forward, you get *nothing* from me. I
>simply relayed information that I believe to be factual and interesting.
>You want to prattle about "Moses" and that crapola, feel free. If you
>like living in a complete vacuum, you may now do so.
>Sheesh... I've never seen such a bunch of babies... wanting information,
>then trying to pick it apart when some is given. Is it any mystery why
>the DB guys won't post here directly?
>Sorry I fed your fire. It won't happen again...
----------------------------------------------
First of all folks it's sunday. <sniff> and now this. My heart nearly crashed.
Bob, think of the times of prohibition.
You _are_ the bar-tender. In front of all these drug addicts.
Did they never grab for their knives when you had to go to sleep??
The next day all was forgotten again. No?
Bob, we're prepared to exploit you till the last piece of evidence you hide
behind your wooden roof. I tell you.
Try to see it from _our_ side. _You_ write mails with Hsu. We're jealous guys.
_You_ write I can't tell _all_. We're drug addicts. Remember that.
And we're dangerous if we don't get enough. And enough is never enough.
I mean we all have these damned genetics, no? And hate prohibition. :)
Robert Hyatt
Joe McCaughan (shi...@best.com) wrote:
: : : --
: : : http://www.demon.co.uk/oxford-soft
: : : Publishing would be fine.
: : : Chris Whittington
: Keep posting Bob. Although I find the 100 NPS arg a little
: hard to believe I still think the posts are interesting.
: All we have so far is hearsay. I guess we'll have to wait
: until baby blue gets on a server and starts mopping the floor
: with the micros... :)
Sorry, but it is simply not worth the hassle. I'd personally rather have
third-hand information that is factual, than nothing. Others don't agree,
however. I'll constrain my posts to Crafty and such, which will be
"first-hand" and not subject to claims of "false! lies! exaggerations!"
I'll answer private email of course...
Chris Carson
I agree with Bob here. Bob only passed along information. If you
do not like the message, don't shoot the messanger.
The DB team did a 10 game experiment at 100K nps. Take or leave the
results. 10-0 speaks for itself, Bob said (my interpretation) that
10 games does not prove anything, but does show that the DB eval
is no weakling (ask Kasparov, he should agree since we just had
another set of relevant data).
Best regards,
Chris Carson
Robert Hyatt wrote:
[Big snip]
Robert Hyatt
Komputer Korner (kor...@netcom.ca) wrote:
: Robert Hyatt wrote:
: The inkompetent komputer.
Speculation. If you don't like speculation, please tune out. If you
want an educated opinion, read on...
Cray Blitz, searching about 5M nodes per second last time I tried it
on a T90, was searching barely 10 plies. It was *very heavy* on
search extensions.
Since they are 40x faster, I'd guess 13-15 plies, depending on the position,
when it is wild, 13, when it is quiet, 15 or more...
a guess only...
Robert Hyatt
mclane (mcl...@prima.ruhr.de) wrote:
: shi...@best.com (Joe McCaughan) wrote:
: >All we have so far is hearsay. I guess we'll have to wait
: >until baby blue gets on a server and starts mopping the floor
: >with the micros... :)
: We do not complain about hearsaid or data that is NOT FACT or NOT
: EVIDENT or roumor.
: We are complaining about Bob saying what he says is more fact, better
: data, and earlier theory.
: That was the point of the whole discussion , that I wanted to explain
: him that his posts are not COMMANDMENTS but also just cool or
: spontaneous reactions and not scientific theories.
Or, perhaps, simply data that was reported by one of the develpers, and
which is assumed (by me) to be factual.
Robert Hyatt
mclane (mcl...@prima.ruhr.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
: >Tell you what. From this point forward, you get *nothing* from me. I
: >simply relayed information that I believe to be factual and interesting.
: THAT I BELIEVE TO BE FACTUAL !!!!
: Bob - what you post is interesting. No doubt.
: Also you post mass !
: Also ok. Nice indeed.
: But our discussion was about : HOW ACCURATE is the data you post.
: Because you always claim it is 100% fact or it is proven long time
: before whoever was born or these kind of stuff.
I only claimed that it was reported by Hsu, to two different journalists,
and to me. I simply trust his data. If you don't, there's nothing I can
do about it. I don't have a "hidden deep blue" at my office to run some
convincing tests on.
With DB, you either have to take what you get, or take nothing. There
is no other choice we have at present...
: My only point was, that a "realist" like you, is not more than a
: believer e.g. of the catholic-church.
: You are beleving in what you define as FACTS. Others believe in UFO
: and aliens from outer space. Others believe in catholics commandments
: etc.
: We are onle making our internal jokes because you do it alway and
: always without seeing it, and many others don't see it . That is funny
: for us, to see you present your BELIEVED data like a priest in the
: church is open the golden shrine, raising his voice and quoting out of
: the golden book.
: We are not attacking you for this. It only amuses us, therefore the
: images.
: >Sheesh... I've never seen such a bunch of babies... wanting information,
: >then trying to pick it apart when some is given.
: A german princess has one day said, when she heard about the wishes of
: the people in her country , starving because they had no bread/food:
: Sollen sie doch Kuchen essen!
: They shall eat cookies! What is their problem ?!
: We do want information. Otherwise we would not subscribe this
: newsgoup. Also we want to share this with other computerchess-freaks.
: But we don't want wrong information or filtered information.
: Is that so difficult to understand?
: > Is it any mystery why
: >the DB guys won't post here directly?
: Do you really believe we would ATTACK them in person only because they
: are that nice to present DATA ?
: I have e.g. never said that THEY have cheated by manually making other
: moves than DB has made. I would never claim this. I have respect for
: the persons behind, although I don't like what they do. Why should I
: believe that well-honored people cheat. With Kasparov it is different.
: He has cheated in past.
: >Sorry I fed your fire. It won't happen again...
: NONO. You always misunderstand. It is not YOU. It is what you do.
brucemo
Chris Carson wrote:
>
> I agree with Bob here. Bob only passed along information. If you
> do not like the message, don't shoot the messanger.
>
> The DB team did a 10 game experiment at 100K nps. Take or leave the
> results. 10-0 speaks for itself, Bob said (my interpretation) that
> 10 games does not prove anything, but does show that the DB eval
> is no weakling (ask Kasparov, he should agree since we just had
> another set of relevant data).
Ten games does prove something. If X beat Y ten in a row, it is true
almost beyond a shadow of a doubt that X would continue to beat Y a
majority of the time. Note that I don't say that X would beat Y all
of the time, just more than 50%. Which is a way of saying that X is
better than Y.
Problem with this example is that we have little idea about why X
beats Y, or what the contest is itself.
I'm interested in this, but I'm not dying to know. There are things
I'd be more interested in knowing.
If I beat Rebel ten times in a row, I don't think I'd publish the
result or the games, because I'd like to make it 11 at a WMCCC or
something.
bruce
chrisw
--
http://www.demon.co.uk/oxford-soft
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
<5m9l24$q...@juniper.cis.uab.edu>...
> Joe McCaughan (shi...@best.com) wrote:
> : Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
> : : chrisw (chr...@cpsoft.demon.co.uk) wrote:
>
> : : : --
> : : : http://www.demon.co.uk/oxford-soft
>
> : : : Marcel van Kervinck <mar...@stack.nl> wrote in article
> : : : <5m7str$g...@turtle.stack.nl>...
> : : : > chrisw (chr...@cpsoft.demon.co.uk) wrote:
> : : : >
> : : : > : Who says its at 100,000 nps ?
> : : : >
> : : : > : And and and
> : : : >
> : : : > : Where are the games. Lets see it. Lets see the data. Lets see
the logs.
> : : : > : Lets see the EVIDENCE.
> : : : >
> : : : > Yeah. And after that you'll demand print-outs of the grown
> : : : > game trees. And after that you'll demand print outs including
> : : : > evaluations in the leafes. And after that you'll demand these
> : : : > evals factored. And after that you'll demand VHDL sources
> : : : > (if any ofcourse).
> : : : >
> : : : > Can't you people just don't wait 'till they publish?
>
> : : : Publishing would be fine.
>
> : : : It's the 'data' handed out via Moses on the tablets of stone that's
the
> : : : problem.
>
> : : : Chris Whittington
>
> : : Tell you what. From this point forward, you get *nothing* from me.
I
> : : simply relayed information that I believe to be factual and
interesting.
> : : You want to prattle about "Moses" and that crapola, feel free. If
you
> : : like living in a complete vacuum, you may now do so.
>
> : : Sheesh... I've never seen such a bunch of babies... wanting
information,
> : : then trying to pick it apart when some is given. Is it any mystery
why
> : : the DB guys won't post here directly?
>
> : : Sorry I fed your fire. It won't happen again...
>
> : Keep posting Bob. Although I find the 100 NPS arg a little
> : hard to believe I still think the posts are interesting.
>
> : All we have so far is hearsay. I guess we'll have to wait
> : until baby blue gets on a server and starts mopping the floor
> : with the micros... :)
>
>
> Sorry, but it is simply not worth the hassle. I'd personally rather have
> third-hand information that is factual, than nothing.
You just don't see it, do you ?
You don't have FACTS, you have MYTHS.
The IBM-DB-Hyatt axis is busy generating not facts and true stories, but
myths and dreams.
Myth 1. DB and Kasparov was a realistic match.
Myth 2. DB has some fantastic score against 'top' micros. A score, which if
extrapolated via ELO calculations, and then 1000x speed up, suggests a full
ELO which doesn't relate to (a) reality and (b) its performance against
Kasparov and (c) its oen makers view of its own strength.
Myth 3. DB is doing something that nobody else has thought of, and is doing
it so well that it wipes the floor with everything. And that this something
is known to nobody else and hasn't leaked out.
Myth 1. has at least the advantage of being measured in a match for all to
see.
Myths 2 and 3 are just (sorry Bob for the analogy, but it seems apt) handed
out on stone tabletes that only the high priest with special access to the
mountain gets to read.
If this argument were just about how strong is Crafty or Stobor or
soemthing, then no matter. But its about the major and fundamental element
of what all of us in computer chess are involved in.
I, for one, am not taking the high priests stone tablets at face value.
If Bob wants to go all prima-donna again and say he's not going to talk -
he 'knows' but he's not gonna talk, ok. I've seen it in the little boys
playground and I've seen it in the big boys playground. Its prima-donna
crap either way.
Chris Whittington
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: chrisw (chr...@cpsoft.demon.co.uk) wrote:
: : Publishing would be fine.
: : It's the 'data' handed out via Moses on the tablets of stone
: : that's the problem.
: : Chris Whittington
: Tell you what. From this point forward, you get *nothing* from me. I
: simply relayed information that I believe to be factual and interesting.
: You want to prattle about "Moses" and that crapola, feel free. If you
: like living in a complete vacuum, you may now do so.
Remember it was you who in triumph told us about the 10-0.
Now I am a good loser and I at first have congratulate you winning an
old discussion. Right?
Other people were more sceptical. Having thought it over I think they
are simply right. The claim is simply too high as I have pointed out
in another posting.
4-6, 5-5, 6-4 is acceptable to accept without evidence, but 10-0 without
any facts needs proof otherwise it can't be taken serious.
: Sheesh... I've never seen such a bunch of babies... wanting
: information, then trying to pick it apart when some is given. Is it
: any mystery why the DB guys won't post here directly?
Smoke Bob!
We want facts and no smoke!
- The 10 games;
- The main lines, scores and depths;
Remember you were going completely nuts when somebody posted a lot of
lost Crafty games here in rgcc. And who was the one yelling for the
games? Remember?
This is the same situation Bob. Don't try to patronize people calling
them "a bunch of babies". Better bring us the 10 babies if they exist.
To avoid any cherry picking please provide the 10 babies with the
main lines, scores and depths.
Or even better arrange a 10 game match live on the ICC so everybody can
see. Should be no problem as the DB at 100,000 NPS seems to exist.
I can't believe the 10-0 without proof.
The claim is too high.
To take the 10-0 claim without evidence is naive.
And not very scientific either.
- Ed Schroder -
chrisw
--
http://www.demon.co.uk/oxford-soft
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
<5m8alq$b...@juniper.cis.uab.edu>...
> chrisw (chr...@cpsoft.demon.co.uk) wrote:
>
> : --
> : http://www.demon.co.uk/oxford-soft
>
> : Marcel van Kervinck <mar...@stack.nl> wrote in article
> : <5m7str$g...@turtle.stack.nl>...
> : > chrisw (chr...@cpsoft.demon.co.uk) wrote:
> : >
> : > : Who says its at 100,000 nps ?
> : >
> : > : And and and
> : >
> : > : Where are the games. Lets see it. Lets see the data. Lets see the
logs.
> : > : Lets see the EVIDENCE.
> : >
> : > Yeah. And after that you'll demand print-outs of the grown
> : > game trees. And after that you'll demand print outs including
> : > evaluations in the leafes. And after that you'll demand these
> : > evals factored. And after that you'll demand VHDL sources
> : > (if any ofcourse).
> : >
> : > Can't you people just don't wait 'till they publish?
>
> : Publishing would be fine.
>
> : It's the 'data' handed out via Moses on the tablets of stone that's the
> : problem.
>
> : Chris Whittington
>
> Tell you what. From this point forward, you get *nothing* from me. I
> simply relayed information that I believe to be factual and interesting.
> You want to prattle about "Moses" and that crapola, feel free.
I thought it was a rather neat and aposite analogy. So sorry you disagree.
> If you
> like living in a complete vacuum, you may now do so.
>
> Sheesh... I've never seen such a bunch of babies... wanting information,
> then trying to pick it apart when some is given.
Picking apart information is the scientific process, no ?
Or have you moved way beyond that now ? Into the realms of myths and dreams
?
> Is it any mystery why
> the DB guys won't post here directly?
Brilliant attempt at democratising your position by using, no, not posts,
but not-posts :)
I like it. Goebbels would have loved it.
Look, if DB and co can't won't post because of critics, then their position
can't be so sound.
If they/you are right, their exposition will stand against ANYTHING.
Of perhaps I should democratise the counter-position by pointing to
a) your endless threats to drop out of the discussion
b) their failure to enter it ?
>
> Sorry I fed your fire. It won't happen again...
>
What, you're quitting again ?
Don't believe you, Bob.
Chris Whittington
>
Robert Hyatt
chrisw (chr...@cpsoft.demon.co.uk) wrote:
: You just don't see it, do you ?
: You don't have FACTS, you have MYTHS.
: The IBM-DB-Hyatt axis is busy generating not facts and true stories, but
: myths and dreams.
: Myth 1. DB and Kasparov was a realistic match.
It's a myth more believe than don't however, which includes me..
: Myth 2. DB has some fantastic score against 'top' micros. A score, which if
: extrapolated via ELO calculations, and then 1000x speed up, suggests a full
: ELO which doesn't relate to (a) reality and (b) its performance against
: Kasparov and (c) its oen makers view of its own strength.
This is not a myth, however much you'd like to believe it so. Have you
ever done the experiment to verify the speed of light (C)? I haven't, but
I trust the 186K+ figure I was taught in various physics courses. Have you
seen *any* false data from the DB guys in their 12+ years of working on that
hardware? I haven't...
: Myth 3. DB is doing something that nobody else has thought of, and is doing
: it so well that it wipes the floor with everything. And that this something
: is known to nobody else and hasn't leaked out.
Never said this. I said their eval is simply *good* and they can do anything
they've seen, and do it instantly. They want actual mobility, they can do it
and not worry about the cost. They want square-control, and they do it at no
cost. They want a fast tree search in the hardware for the last 4 plies of
the tree, they do it *fast*... I didn't claim they have "new and never-before
heard ideas, although they might. I simply claim that those that think that
DB is a dumb evaluator/fast searcher are not just wrong, but badly wrong.
: Myth 1. has at least the advantage of being measured in a match for all to
: see.
: Myths 2 and 3 are just (sorry Bob for the analogy, but it seems apt) handed
: out on stone tabletes that only the high priest with special access to the
: mountain gets to read.
: If this argument were just about how strong is Crafty or Stobor or
: soemthing, then no matter. But its about the major and fundamental element
: of what all of us in computer chess are involved in.
: I, for one, am not taking the high priests stone tablets at face value.
And I, for one, don't care. Anyone here is free to take what I supply and
read it, or ignore it, or discuss it. But this "high priest and stone
tablets" is just a tad stupid. Make that a big tad... However, as the
saying goes, if you want to live in ignorance of what few details I know,
suits me. I only thought *any* reliable information was better than *no*
information. The info *is* reliable, but I was mistaken in posting it. I'll
stick to topics where I can show log files or source code to support results,
to avoid this knee-jerk "anything about DB is myth and bullshit" sort of
response...
: If Bob wants to go all prima-donna again and say he's not going to talk -
: he 'knows' but he's not gonna talk, ok. I've seen it in the little boys
: playground and I've seen it in the big boys playground. Its prima-donna
: crap either way.
I didn't say that. I said I simply would not pass along anything interesting
I learn about them. Seems to be little point in it. I suspect that there is
more to this than meets the eye, maybe the old "the academics just can't do
it like us professional, they only have more resources..." argument?
A Prima-donna is someone that believes he is better than everyone else.
That seems to fit others much better than it does me.
: Chris Whittington
However, before ending this point, for those of you interested in DB, please
don't flood me with email. I've already seen about 40 queries this morning.
This is a lose-lose situation. I can post results here where everyone can
see them and then put up with this minority view of "bullocks, that didn't
hapen" or else respond to a large number of email queries about DB. Either
way is a losing proposition. Which is sad, when you think about it... because
there are lots of losers either way it is handled...
Robert Hyatt
Ed Schroder (rebc...@xs4all.nl) wrote:
: From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: : chrisw (chr...@cpsoft.demon.co.uk) wrote:
: : : Publishing would be fine.
: : : It's the 'data' handed out via Moses on the tablets of stone
: : : that's the problem.
: : : Chris Whittington
: : Tell you what. From this point forward, you get *nothing* from me. I
: : simply relayed information that I believe to be factual and interesting.
: : You want to prattle about "Moses" and that crapola, feel free. If you
: : like living in a complete vacuum, you may now do so.
: Remember it was you who in triumph told us about the 10-0.
: Now I am a good loser and I at first have congratulate you winning an
: old discussion. Right?
: Other people were more sceptical. Having thought it over I think they
: are simply right. The claim is simply too high as I have pointed out
: in another posting.
: 4-6, 5-5, 6-4 is acceptable to accept without evidence, but 10-0 without
: any facts needs proof otherwise it can't be taken serious.
You have to believe what you have to believe. If you are suspcious of
their results, maybe you could explain why? If I told you that the current
Crafty seems to be beating gnuchess 85% of the time, would you take that at
face value, demand proof, or get a copy of both and run it yourself? If you
did (3) you'd probably get the same result one of my "cohorts" has gotten.
There's been no examples of where I have fabricated test results on my on,
and posted them here, because it would be exposed within 24 hours. Another
one: Crafty 12.3 solve 297/300 of wac in 1 min or less. 295 in 30 seconds
or less. Most are < 1 second in fact. After I post this, if you don't believe
it, you can run the test, and get the same results. The point is, I have
never fabricated data. Hsu and group are *exactly* the same. They once
drew an erroneous conclusion about what singular extensions did to their
rating, but only after analyzing the very thing you hold so dear, computer
vs computer testing. They discovered that a small change would exaggerate
the rating difference, when the testing is between the have and the have-not
program. They published a newer result later.
I have a hard tim swallowing "it can't be taken serious." Of course it
can. They are reputable scientists. Don't let your ego outweigh your
better judgement here. There is one of you. There are several of them
They all have Ph.D.'s, they are all *good* computer scientists. It's not
a big surprise they can blast you, me, or anyone else. They have more
horsepower in the hardware than we do, they have more people (not
necessarily any better, but *more* of them) than we do, and yet you and
others are simply "certain" they can't win 10 games in a row...
: : Sheesh... I've never seen such a bunch of babies... wanting
: : information, then trying to pick it apart when some is given. Is it
: : any mystery why the DB guys won't post here directly?
: Smoke Bob!
smoke is what I see here all the time. Smoke in the form of "that can't
be right" "I *know* it can't". Of course, having beaten Kasparov doesn't
lend any credibility to the fact that something remarkable has been done
by those guys?
: We want facts and no smoke!
I gave you *all* the facts I had. 10 games, 5 vs rebel,5 vs genius,
10 wins for 100K DB. I don't have any more information. I'd like to
see scores/variations for lots of reasons... but knowing they won 10
is important...
: - The 10 games;
: - The main lines, scores and depths;
: Remember you were going completely nuts when somebody posted a lot of
: lost Crafty games here in rgcc. And who was the one yelling for the
: games? Remember?
I had contrary data. If someone was "breaking even" with deep blue, and
they had given me this data, I'd have questioned it already. However,
We already knew they were winning 9 out of every 10 over the past 12
years against the best of the micros. So is going from 9 out of 10, to
10 out of 10 such a big jump? You keep forgetting their past results,
simply because you now have a P6/200, and will have a P6/300 later this
year. Of course, they are just a wee bit faster themselves this year
than last.
: This is the same situation Bob. Don't try to patronize people calling
: them "a bunch of babies". Better bring us the 10 babies if they exist.
: To avoid any cherry picking please provide the 10 babies with the
: main lines, scores and depths.
Sorry, you've gotten all I have. It's an interesting result. It is *not*
outside of what I'd expect, based on watching them smash micros for 12 years
now.
: Or even better arrange a 10 game match live on the ICC so everybody can
: see. Should be no problem as the DB at 100,000 NPS seems to exist.
I think we might see Junior play there at 10M nodes per second. I don't
think anyone would really want to play them however, if "marketing" was the
goal...
: I can't believe the 10-0 without proof.
: The claim is too high.
: To take the 10-0 claim without evidence is naive.
: And not very scientific either.
what about the 9-1 claim. Is that too high? I can produce game
scores, opponents, and everything else for those, all the games at
40/2, played in the ACM and WCCC events. 10-0 is hardly an improvement
at all...
: - Ed Schroder -
Simon Read
mclane, most of the words you post here do come from an English
dictionary. The trouble is, if you connect them together they
don't always form understandable sentences. If you connect the
sentences together, you get paragraphs which read like you were
hallucinating, or your brain is scrambled, or just very very
confused.
mclane (mcl...@prima.ruhr.de) wrote:
>: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>
>
>: >Tell you what. From this point forward, you get *nothing* from me. I
>: >simply relayed information that I believe to be factual and interesting.
>
>: THAT I BELIEVE TO BE FACTUAL !!!!
>
>: Bob - what you post is interesting. No doubt.
>: Also you post mass !
>: Also ok. Nice indeed.
>
>: But our discussion was about : HOW ACCURATE is the data you post.
>: Because you always claim it is 100% fact or it is proven long time
>: before whoever was born or these kind of stuff.
>
>I only claimed that it was reported by Hsu, to two different journalists,
>and to me. I simply trust his data. If you don't, there's nothing I can
>do about it. I don't have a "hidden deep blue" at my office to run some
>convincing tests on.
>
>With DB, you either have to take what you get, or take nothing. There
>is no other choice we have at present...
>
>: My only point was, that a "realist" like you, is not more than a
>: believer e.g. of the catholic-church.
>: You are beleving in what you define as FACTS. Others believe in UFO
>: and aliens from outer space. Others believe in catholics commandments
>: etc.
Let me see if I understand what you're saying, mclane:
"Bob believing that he heard Hsu say something about DB's score
against Rebel or Genius is equivalent to someone believing in UFOs."
You really do talk a complete load of rubbish.
>: We are onle making our internal jokes because you do it alway and
>: always without seeing it, and many others don't see it . That is funny
>: for us, to see you present your BELIEVED data like a priest in the
>: church is open the golden shrine, raising his voice and quoting out of
>: the golden book.
>
>: We are not attacking you for this. It only amuses us, therefore the
>: images.
>
>: >Sheesh... I've never seen such a bunch of babies... wanting information,
>: >then trying to pick it apart when some is given.
>
>: A german princess has one day said, when she heard about the wishes of
>: the people in her country , starving because they had no bread/food:
>: Sollen sie doch Kuchen essen!
>: They shall eat cookies! What is their problem ?!
>
>: We do want information. Otherwise we would not subscribe this
>: newsgoup. Also we want to share this with other computerchess-freaks.
>: But we don't want wrong information or filtered information.
>: Is that so difficult to understand?
>
>: > Is it any mystery why
>: >the DB guys won't post here directly?
>: Do you really believe we would ATTACK them in person only because they
>: are that nice to present DATA ?
>: I have e.g. never said that THEY have cheated by manually making other
>: moves than DB has made. I would never claim this. I have respect for
>: the persons behind, although I don't like what they do. Why should I
>: believe that well-honored people cheat. With Kasparov it is different.
>: He has cheated in past.
>
>
>: >Sorry I fed your fire. It won't happen again...
>: NONO. You always misunderstand. It is not YOU. It is what you do.
Listen, mclane, you post the biggest load of rubbish on this newsgroup.
Bob posted something factual. You have two choices:
(1) Say "I believe it."
(2) Say "I don't believe it becaue Hsu and company are a bunch of liars."
mclane, your pile of festering puke you call a posting can't even make
up its mind if you believe it or not.
Let me quote the confused ramblings of your mind:
>: We do want information. Otherwise we would not subscribe this
>: newsgoup. Also we want to share this with other computerchess-freaks.
>: But we don't want wrong information or filtered information.
>: Is that so difficult to understand?
What the hell are you saying here? It's like saying "We want facts, but
not wrong information." So are you saying that Bob lied? Or are you saying
that Hsu lied? You seem to be implying that somebody is not telling
the truth.
You're also implying that "filtered information" ie via Bob, is not
as good as hearing the same thing from Hsu's mouth. If Bob is only
reporting what he heard Hsu say, I don't see what the difference is.
If Hsu says it it's true, but you're implying that if Bob repeats
THE SAME THING it suddenly becomes "filtered" which you don't want
to hear.
In other words mclane, you just don't make sense. You don't even know
what you believe and your postings really don't make any sense at all.
I stopped reading you months ago and only occasionally I am forced to
see some of your garbage which someone else has quoted.
Simon wwwwwwwwwwwwwwwww dot cranfield dot ac dot uk/~me944p
brucemo
Robert Hyatt wrote:
>
> chrisw (chr...@cpsoft.demon.co.uk) wrote:
> : Myth 2. DB has some fantastic score against 'top' micros. A score, which if
> : extrapolated via ELO calculations, and then 1000x speed up, suggests a full
> : ELO which doesn't relate to (a) reality and (b) its performance against
> : Kasparov and (c) its oen makers view of its own strength.
Estimates of computer Elo are really strange because we constantly mix apples and
oranges.
If you take program A and have it play against a bunch of humans, and have
program B play against the same bunch, you'd get a rating delta between A and B
that might not have anything to do with how well A plays against B.
Likewise you could play A against a pool of other programs, and the rating it
gets from the pool might not have anything to do with how well it would do
against a human or a bunch of humans.
You could have a program that performed at a rating of 2500 against a bunch of
humans, and another program that beats this first program 9 out of 10, which is
382 Elo points, but it wouldn't necessarily follow that the rating of this other
program is 2882. Maybe the rating of this second program against the human pool
would be 2550, or maybe even 2400, who knows.
I have seen arguments that DB couldn't possibly be hugely better than a micro,
because if it were, it would be way better than Kasparov.
I have no evidence as to how good DB is, and it's not like the truth of my own
personal religions depends upon how good DB is. I just think the argument in the
previous paragraph could easily be fallacious. One possible explanation may be
found in what I wrote prior to that. It is also possible that the Swedish list
has suffered from some inflation or drift for the same reasons, so there is
another explanation for inaccuracy.
It is also possible to conclude from my arguments that Rebel could be better than
DB against Kasparov. I have no opinion about this, I just point it out in an
effort to show that I'm not attempting to prove a specific religious viewpoint.
If I were Bob I would not post info he hears about DB. I can sympathize with his
frustration about what he reads in here. A lot of what is written is based upon
religious viewpoints, people are threatned by DB, because its strength or
weakness will give authority to arguments of people who are on one side or the
other of this stupid and artifically imposed speed/knowledge dichotomy.
It is always hard to argue against religous viewpoints by presenting scientific
evidence, and it is especially hard or impossible when the evidence presented is
incomplete, second-hand, or otherwise flawed like this. I think Bob should stop
trying to do so.
I hope that the DB folks publish, in the mean time if they want to say something
here, they all have internet access I expect. I'm interested in knowing about
this 10-0 thing, but I can't conclude anything from it, it's too fragmentary.
> This is not a myth, however much you'd like to believe it so. Have you
> ever done the experiment to verify the speed of light (C)? I haven't, but
> I trust the 186K+ figure I was taught in various physics courses. Have you
> seen *any* false data from the DB guys in their 12+ years of working on that
> hardware? I haven't...
He has a point though, Bob. There would still be the requisite shrieking,
accusation of cheating, complaints about ICCA events, death threats, etc., if
they published something, but at least something would exist in black and white,
and if people weren't able to try to reproduce the data, at least it would be
from a primary source, and at least people could see enough that they'd have a
chance of understanding it.
> : Myth 3. DB is doing something that nobody else has thought of, and is doing
> : it so well that it wipes the floor with everything. And that this something
> : is known to nobody else and hasn't leaked out.
>
> Never said this. I said their eval is simply *good* and they can do anything
> they've seen, and do it instantly. They want actual mobility, they can do it
> and not worry about the cost. They want square-control, and they do it at no
> cost. They want a fast tree search in the hardware for the last 4 plies of
> the tree, they do it *fast*... I didn't claim they have "new and never-before
> heard ideas, although they might. I simply claim that those that think that
> DB is a dumb evaluator/fast searcher are not just wrong, but badly wrong.
They haven't published anything about their eval function. We hear second-hand
comments from a reliable poster (you) that they do a lot in their eval. I take
this at face value.
Maybe they do other stuff, who knows. If they want to publish it, great. If
they don't, well, that's not unprecedented, is it?
> : Myths 2 and 3 are just (sorry Bob for the analogy, but it seems apt) handed
> : out on stone tabletes that only the high priest with special access to the
> : mountain gets to read.
There's a point to this, but admit also that it's not the high priest's fault
that his email gets answered. It's gross to have information you can't publish
about something you are very interested in, information you think proves your
points in a very contentious debate.
> : If this argument were just about how strong is Crafty or Stobor or
> : soemthing, then no matter. But its about the major and fundamental element
> : of what all of us in computer chess are involved in.
>
> : I, for one, am not taking the high priests stone tablets at face value.
Neither am I, about some stuff. If Bob were to tell me many of the Nolot
positions they solve, and how fast they solve them (which he hasn't done, I don't
even know if he knows this information), I'd believe what I heard. But this 10-0
thing, what is that?
Appreciate the uncomfortable position Bob is in though. He can't do much better
than he is already doing.
> And I, for one, don't care. Anyone here is free to take what I supply and
> read it, or ignore it, or discuss it. But this "high priest and stone
> tablets" is just a tad stupid. Make that a big tad... However, as the
> saying goes, if you want to live in ignorance of what few details I know,
> suits me. I only thought *any* reliable information was better than *no*
> information. The info *is* reliable, but I was mistaken in posting it. I'll
> stick to topics where I can show log files or source code to support results,
> to avoid this knee-jerk "anything about DB is myth and bullshit" sort of
> response...
If you can get them to post any of this, I would accept it at face value:
1) Results of test suites. Doesn't have to be detailed, just a vague idea. I'd
even take Bratko-Kopec results, I don't care, it's still interesting, and of
course nobody would fake these.
2) Results of a long match with any commercial program, preferably at standard
time controls with the micro on a P6/200, and using the whole DB configuation,
not just a small chunk, and preferably a decent number of games. If any of this
is absent (like the time control is 30 0 or something), mention this. I don't
have to see the game scores, I don't care, this isn't supposed to be something
that will let the commercial guy bust DB's book or something like that, although
it would be interesting to see one interesting win and one interesting loss (or
at least one interesting draw :-). I don't believe anyone would fake this,
either, and if the match was long enough it would serve as a real bit of
evidence.
3) Anything about search depth achieved in N minutes on some specific position we
all know about. I don't have to see the PV's, I'm just curious how deep they go
in at least one or two specific situations. I can tell someone that my program
goes 10-12 plies in the middlegame in a standard match, but the difference
between 10 and 12 is pretty significant, so it's hard to compare with other
programs, and in any case it's just a guess on my part, so maybe it is wrong.
I'm concerned that if we attack these guys too hard, by demanding much more
information than this, that we might get a blanket "fuck you" back. We don't
have any right for a full test suite dump of anything, nor do we have the right
to the game scores they produce internally for testing.
It's a fascinating machine though, and I think we all would like to know more
about it. Let's not try to get it by telling them to fuck themselves both before
and after demanding that they send us their source code.
> : If Bob wants to go all prima-donna again and say he's not going to talk -
> : he 'knows' but he's not gonna talk, ok. I've seen it in the little boys
> : playground and I've seen it in the big boys playground. Its prima-donna
> : crap either way.
This is too much, Chris. Bob's in a gross position and we don't need to put the
weight of every maniac who reads this group on his head. He's given us
everything he can, we shouldn't bomb him because he reported something that we
have a hard time digesting.
> However, before ending this point, for those of you interested in DB, please
> don't flood me with email. I've already seen about 40 queries this morning.
> This is a lose-lose situation. I can post results here where everyone can
> see them and then put up with this minority view of "bullocks, that didn't
> hapen" or else respond to a large number of email queries about DB. Either
> way is a losing proposition. Which is sad, when you think about it... because
> there are lots of losers either way it is handled...
Exactly. Something happened in this newsgroup a few months back, it wasn't even
related to DB, and Hsu saw fit to email us both about it. I answered him, and
since he sent me mail first, I appended a few questions about DB.
I didn't get any answer, sometimes life is like that. I figured that he had good
reason for not answering.
Both before and during the match I thought about sending him some more questions,
but I figured that since must have good reason for not answering my first batch,
that I wouldn't bother him with more.
I also wasn't going to bomb him with questions right after the match, and this
still qualifies as right after the match, in my opinion. I'd like to get answers
to my questions, but I also don't want to think about how much crap these guys
are getting bombed with from everyone who has a pet idea about how a chess
program should work.
Let's see if they come out with real info.
bruce
Don Getkey
In article <5m9l24$q...@juniper.cis.uab.edu>, hy...@crafty.cis.uab.edu
(Robert Hyatt) writes:
>Sorry, but it is simply not worth the hassle. I'd personally rather have
>third-hand information that is factual, than nothing. Others don't agree,
>however. I'll constrain my posts to Crafty and such, which will be
>"first-hand" and not subject to claims of "false! lies! exaggerations!"
>
>I'll answer private email of course...
>
>
Sign me on your Private News e-mail list Bob. I'm always interested in
the inside scoops you share.
yours in chess,
Don
Ramsey MN USA
Don Getkey
I can't believe how much people have read into this simple face value post
by Bob Hyatt on "Little Blue's" alleged performance against R8 and CG5.
It is highly uncalled for. Bob has nothing to gain except the pleasure of
adding to the pool of
knowledge/speculation/heresay/myth/theory/hypothesis/conjecture/ignorance/
etc., that swirls about on the topic of Deep Blue.
The people here (and you know who you are) who have knocked Bob for
speaking about unconfirmed performance results of "Little Blue" against
the micro elite, (not to say MCP6 and HIARCS6 should not be included in
this), need to back off, and get a life.
Howard Exner
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
<5ma6jb$n...@juniper.cis.uab.edu>...
> what about the 9-1 claim. Is that too high? I can produce game
> scores, opponents, and everything else for those, all the games at
> 40/2, played in the ACM and WCCC events. 10-0 is hardly an improvement
> at all...
... most everything deleted ....
This would be great if you would post these Deep Blue/ Deep Thought game
scores.
Or if you had the URL link that would be much appreciated.
Thanks in advance.
Howard
brucemo
Robert Hyatt wrote:
> what about the 9-1 claim. Is that too high? I can produce game
> scores, opponents, and everything else for those, all the games at
> 40/2, played in the ACM and WCCC events. 10-0 is hardly an improvement
> at all...
It is if you drop node rate to 100K nps. It's been assumed that in the
past they won via depth, this implies that they are winning by eval now.
So going from winning by depth to winning (massively) via eval is not
something that can taken on faith.
bruce
brucemo
Robert Hyatt wrote:
> One thing I can post, is over 50 games they have played at 40/2 against
> micro programs as well as programs that are stronger than micros (Cray
> Blitz for one)... and they are sporting >9 wins for every 10 games they
> played.
The Ed argument against this is that DT vs some cheesy micro is basically
DT vs nothing. The difference between a PC circa 1989 and a current PC is
pretty huge, and this may allow the modern micro to get above tactics that
have killed micros in the past.
That is the Ed argument, if I remember it right, with a small dose of
Diepeveen.
Whether or not you believe the Ed argument or not, there are those who do.
And there's enough substance to it that it'd be more compelling to see
results against current hardware.
bruce
Don Fong
In article <5ma5ka$n...@juniper.cis.uab.edu>,
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote:
>chrisw (chr...@cpsoft.demon.co.uk) wrote:
>: Myth 2. DB has some fantastic score against 'top' micros. A score, which if
>: extrapolated via ELO calculations, and then 1000x speed up, suggests a full
>: ELO which doesn't relate to (a) reality and (b) its performance against
>: Kasparov and (c) its oen makers view of its own strength.
>
>ever done the experiment to verify the speed of light (C)? I haven't, but
>I trust the 186K+ figure I was taught in various physics courses. Have you
>seen *any* false data from the DB guys in their 12+ years of working on that
>hardware? I haven't...
the difference is that you *can* do the experiment(s) to
measure c (the speed of light). there are no secrets. it's all
public. it is not a matter of faith or credibility.
in contrast, we cannot do the experiment with DB. all we have to
go on is hearsay. this *is* a matter of faith and credibility. over
the years you've posted a bunch of stuff like "Hsu says this" or
that. unf either because of your writing style, or because of Hsu's
speaking style, it usually comes off as rather vague and cryptic,
the kind of thing that raises more questions than it answers. that
doesn't do much for the credibility of you or the DB team.
Rolf Tueschen
Ed Schroder <rebc...@xs4all.nl> wrote:
>To take the 10-0 claim without evidence is naive.
>And not very scientific either.
>- Ed Schroder -
I have to agree with Ed. On this point.
I don't know why Ed had congratulated so early.
It's the Sofia syndrom this time. Ed? :)
Rolf Tueschen (The Doctor of Eddie)
Moritz Berger
On Sun, 25 May 1997 10:17:58 -0700, brucemo <bru...@seanet.com>
wrote:
>Chris Carson wrote:
< snip >
>> The DB team did a 10 game experiment at 100K nps. Take or leave the
>> results. 10-0 speaks for itself, Bob said (my interpretation) that
>> 10 games does not prove anything, but does show that the DB eval
>> is no weakling (ask Kasparov, he should agree since we just had
>> another set of relevant data).
>Ten games does prove something. If X beat Y ten in a row, it is true
>almost beyond a shadow of a doubt that X would continue to beat Y a
>majority of the time. Note that I don't say that X would beat Y all
>of the time, just more than 50%. Which is a way of saying that X is
>better than Y.
>Problem with this example is that we have little idea about why X
>beats Y, or what the contest is itself.
>I'm interested in this, but I'm not dying to know. There are things
>I'd be more interested in knowing.
>If I beat Rebel ten times in a row, I don't think I'd publish the
>result or the games, because I'd like to make it 11 at a WMCCC or
>something.
>bruce
Point is that not all of those 10 fabulous games were at 40/120 tc,
some were at game/30 etc. Our "claims" about relative strenght DB vs.
micros were always about 40/120.
Moritz
-------------
Moritz...@msn.com
"The truth will always come out in the end."
(Komputer Korner, 28/03/1997 08:22 on rec.games.chess.computer)
Robert Hyatt
brucemo (bru...@seanet.com) wrote:
<snip>
: He has a point though, Bob. There would still be the requisite shrieking,
: accusation of cheating, complaints about ICCA events, death threats, etc., if
: they published something, but at least something would exist in black and white,
: and if people weren't able to try to reproduce the data, at least it would be
: from a primary source, and at least people could see enough that they'd have a
: chance of understanding it.
One thing I can post, is over 50 games they have played at 40/2 against
micro programs as well as programs that are stronger than micros (Cray
Blitz for one)... and they are sporting >9 wins for every 10 games they
played.
This is available in the old ICCA journals, as well as the Journal of the
ACM, as well as in Monty's new book, which tracks every game they played
in the computer championships.
Going from 9-1 (actually >45-5) to 10-0 is not a huge jump, and is quite
easy for me to believe..
they have released details, just not in places where we hear them. Hsu and
others have given talks around the country and have released a detail here,
a detail there, quite a bit makes it back via that route..
: I'm concerned that if we attack these guys too hard, by demanding much more
: information than this, that we might get a blanket "fuck you" back. We don't
: have any right for a full test suite dump of anything, nor do we have the right
: to the game scores they produce internally for testing.
: It's a fascinating machine though, and I think we all would like to know more
: about it. Let's not try to get it by telling them to fuck themselves both before
: and after demanding that they send us their source code.
<snip again>
: bruce
Robert Hyatt
Howard Exner (hex...@dlcwest.com) wrote:
: Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
: <5ma6jb$n...@juniper.cis.uab.edu>...
: > what about the 9-1 claim. Is that too high? I can produce game
: > scores, opponents, and everything else for those, all the games at
: > 40/2, played in the ACM and WCCC events. 10-0 is hardly an improvement
: > at all...
: ... most everything deleted ....
: This would be great if you would post these Deep Blue/ Deep Thought game
: scores.
: Or if you had the URL link that would be much appreciated.
: Thanks in advance.
: Howard
:
Actually, most of the game scores are available on various ftp sites
around the world, in the place where game collections are kept. They
are typically labelled something like "wccc95.pgn" or "NACC91.pgn" or
something similar. NACC is the ACM events... WCCC is the every three
year world champ. WMCCC is the micro tourney, but you won't find a DB
entry there...
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
[ snip ]
: : Remember it was you who in triumph told us about the 10-0.
: : Now I am a good loser and I at first have congratulate you winning an
: : old discussion. Right?
: : Other people were more sceptical. Having thought it over I think they
: : are simply right. The claim is simply too high as I have pointed out
: : in another posting.
: : 4-6, 5-5, 6-4 is acceptable to accept without evidence, but 10-0
: : without any facts needs proof otherwise it can't be taken serious.
: You have to believe what you have to believe. If you are suspcious of
: their results, maybe you could explain why?
I am not suspicious these are your words...
And your premature conclusion of my words.
I never said it nor implied it.
Deep Blue at 100,000 NPS emulating a PP-200.
Can we simply call it DB-PC ??
Would be a nice abbreviation to shorten our postings.
After all there are rumors for such plans.
Now DB-PC vs R8/G5 = 10-0 implies that DB-PC is by far the strongest
chess program on our beautiful earth. Right?
The information 10-0 implies that if the real DB would be translated
to Pc they will kill every Pc chess program that is available. After
all the initial goal of Hsu and co (creating the NPS 100,000 machine /
version) was to check the state of art between DB and todays Pc chess
programs.
Right?
10-0 implies that DB-PC is about 300-400 ELO points stronger than
Rebel8, Genius5, Hiarcs6 etc. etc.
Right?
Don't you see the HUGE claim that has been made?
300-400 ELO better than Rebel8, Genius5, Hiarcs6 etc. ???
And we the "bunch of bad losers", "crying babies" should eat that cake?
As I said, I can buy scores like 4-6, 5-5, 6-4 without further evidence.
But 10-0 implying 300-400 ELO better than the current top sounds too
heavy to accept and has noting to do with being suspicious.
I would call it common sense.
Eating the 10-0 cake without any other information is naive.
Also you are a university man so you the word "scientific" must mean
something to you.
: If I told you that the current
I get the point...
- I have a too big ego;
- I am not a reputable scientist;
- I am not a Ph.D.;
All true Bob... :)
But I also have common sense, I don't need to be a Ph.D for that... :)
Also the information comes right from the place where the computer chess
power (status) is. There are rumors about a Deep Blue for the Pc. If that
is true will the program be free downloadable or is money a subject here.
What do you think?
10-0
A total (money) smash!
Did you ever thought about this scenario?
: : : Sheesh... I've never seen such a bunch of babies... wanting
: : Smoke Bob!
No, now be honest to yourself.
Suppose I put on my home page the following information:
Rebel - Crafty 10 - 0
Crafty was slaughtered by Rebel in positional play and king attack.
End of info.
Would you accept that information?
: : This is the same situation Bob. Don't try to patronize people calling
: : them "a bunch of babies". Better bring us the 10 babies if they exist.
: : To avoid any cherry picking please provide the 10 babies with the
: : main lines, scores and depths.
: Sorry, you've gotten all I have. It's an interesting result. It is
*not*
: outside of what I'd expect, based on watching them smash micros for 12
years
: now.
: : Or even better arrange a 10 game match live on the ICC so everybody
can
: : see. Should be no problem as the DB at 100,000 NPS seems to exist.
: I think we might see Junior play there at 10M nodes per second. I don't
: think anyone would really want to play them however, if "marketing" was
the
: goal...
: : I can't believe the 10-0 without proof.
: : The claim is too high.
: : To take the 10-0 claim without evidence is naive.
: : And not very scientific either.
: what about the 9-1 claim. Is that too high? I can produce game
: scores, opponents, and everything else for those, all the games at
: 40/2, played in the ACM and WCCC events. 10-0 is hardly an improvement
: at all...
Wasn't this kind of statements not the reason for our NPS game?
Your 9-1 claim (if it is true) is based on information to which year?
1994?
1995?
This is 1997, your info is out dated.
Todays Pc chess programs are 150-200 ELO stronger than 1994/1995 due to
hardware and software improvements.
Also your 9-1 claim looks not right.
I remember at least 2 DT/DB losses against Genius and Fritz.
- Ed Schroder -
chrisw
--
http://www.demon.co.uk/oxford-soft
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
<5mapij$8...@juniper.cis.uab.edu>...
> brucemo (bru...@seanet.com) wrote:
>
> <snip>
>
> : He has a point though, Bob. There would still be the requisite
shrieking,
> : accusation of cheating, complaints about ICCA events, death threats,
etc., if
> : they published something, but at least something would exist in black
and white,
> : and if people weren't able to try to reproduce the data, at least it
would be
> : from a primary source, and at least people could see enough that they'd
have a
> : chance of understanding it.
>
> One thing I can post, is over 50 games they have played at 40/2 against
> micro programs as well as programs that are stronger than micros (Cray
> Blitz for one)... and they are sporting >9 wins for every 10 games they
> played.
>
> This is available in the old ICCA journals, as well as the Journal of the
> ACM, as well as in Monty's new book, which tracks every game they played
> in the computer championships.
This is totally falacious.
1. The 50 x 40/2 games are played by the full version of DB or DT or
whatever, against what was presumably the fastest PC at the time of
playing. No doubt this ranges down to a 386 or 486.
2. The 10-0 is a result by allegedly slowing down DB by 1000-fold.
>
> Going from 9-1 (actually >45-5) to 10-0 is not a huge jump, and is quite
> easy for me to believe..
So going from 9-1 to 10-0 and jumping 1000x in slowness, is not quite easy
for me to believe.
Chris Whittington
>
> they have released details, just not in places where we hear them. Hsu
and
> others have given talks around the country and have released a detail
here,
> a detail there, quite a bit makes it back via that route..
>
>
> : I'm concerned that if we attack these guys too hard, by demanding much
more
> : information than this, that we might get a blanket "fuck you" back. We
don't
> : have any right for a full test suite dump of anything, nor do we have
the right
> : to the game scores they produce internally for testing.
>
> : It's a fascinating machine though, and I think we all would like to
know more
> : about it. Let's not try to get it by telling them to fuck themselves
both before
> : and after demanding that they send us their source code.
>
> <snip again>
>
> : bruce
>
Tord Kallqvist Romstad
hy...@crafty.cis.uab.edu (Robert Hyatt) writes:
> I had contrary data. If someone was "breaking even" with deep blue, and
> they had given me this data, I'd have questioned it already. However,
> We already knew they were winning 9 out of every 10 over the past 12
> years against the best of the micros. So is going from 9 out of 10, to
> 10 out of 10 such a big jump? You keep forgetting their past results,
> simply because you now have a P6/200, and will have a P6/300 later this
> year. Of course, they are just a wee bit faster themselves this year
> than last.
Bob, thanks for posting the results of DB vs. Rebel/Genius. Please continue
to post such information. However, I disagree about one thing. Going from
9 out of 10 to 10 out of 10 is a _very_ big jump. Remember that the 10 game
match was played with a crippled version of Deep Blue, searching only 100K
nps. Deep Thought searched around 10M nps, I think. That is, Deep Blue is
100 times slower that the computer which crushed the micros, and the micros
are at least 10 times faster now. I find the result 10-0 for Deep Blue very
surprising.
I really hope they will play some games on the chess servers soon. This
discussion is boring.
Tord
Enrique Irazoqui
Robert Hyatt <hy...@crafty.cis.uab.edu> escribió en artÃculo
<5ma6jb$n...@juniper.cis.uab.edu>...
[snip]
> We already knew they were winning 9 out of every 10 over the past 12
> years against the best of the micros. So is going from 9 out of 10, to
> 10 out of 10 such a big jump?
Yes it is. This 9-1 was achieved by 20M NPS machines. The 10-0 comes from a
100K machine. The difference is enormous. In the first case it means
outsearching. In the second, a qualitatively better evaluation function.
This 10-0 on equal horsepower would be such a turning point that I don't
find it so outrageous to ask for the proof of existance of these results.
It is outrageous to pretend that we believe in it just because someone says
so, no matter how qualified this someone is.
Enrique
Enrique Irazoqui
[snip]
> We already knew they were winning 9 out of every 10 over the past 12
> years against the best of the micros. So is going from 9 out of 10, to
> 10 out of 10 such a big jump?
Yes it is. This 9-1 was achieved by 20M NPS machines. The 10-0 comes from a
Enrique Irazoqui
(snip)
> Going from 9-1 (actually >45-5) to 10-0 is not a huge jump, and is quite
> easy for me to believe..
It's easy for you to believe what? That a 10-0 happened? In the first
place, as I said in another posting, the 9-1 and 10-0 are not comparable,
since the 10-0 is supposed to have happened on a 100K machine.
Secondly, I question the meaning of this 10-0. If the DB team wanted to
find out the strength of their program compared to the best commercials,
two 5 games matches are not enough to prove much, as we all know. For
example, in my own testing of Hiarcs 6 I played 3 10 games matches where H6
played against Rebel 8, Mchess 6 and Genius 5. Few games, just to have a
rough idea. In the first 7 games against M6, Hiarcs 6 won 7-0. In the first
5 games against Rebel 8, Hiarcs 6 won 4-1. Total, 11-1. Now, would that
allow me to post an 11-1 result and claim that Hiarcs 6 is 400 points
stronger than the average of Mchess and Rebel? I would have been crucified,
and rightly so.
Therefore I question:
1- The meaning of the 10-0
2- The testing procedure of the DB team, since 10 games are not enough
proof.
And in case this 10-0 can be taken as proof (?) of the superior evaluation
of DB's program, then we need the games and the scores before accepting
blindly that a superior creature has been born.
Enrique
Marcel van Kervinck
Ed Schroder (rebc...@xs4all.nl) wrote:
> : : Other people were more sceptical. Having thought it over I think they
> : : are simply right. The claim is simply too high as I have pointed out
> : : in another posting.
> : : 4-6, 5-5, 6-4 is acceptable to accept without evidence, but 10-0
> : : without any facts needs proof otherwise it can't be taken serious.
> : You have to believe what you have to believe. If you are suspcious of
> : their results, maybe you could explain why?
> I am not suspicious these are your words...
> And your premature conclusion of my words.
> I never said it nor implied it.
> Deep Blue at 100,000 NPS emulating a PP-200.
This is where most posters seem to make their basic false assumption:
that DB@100KNPS is somehow equivalent to a software engine on a PP200.
Perhaps this comes from being blinded by years of general purpose processor
programming. Please don't underestimate the power of the enormous
amount of parallellism that one can exploit when using special purpose
hardware. As has been pointed out in this thread before, a PP200 might
be closer to DB@1KNPS. In that light, the 10-0 is not so magical as
people suggest.
> The information 10-0 implies that if the real DB would be translated
> to Pc they will kill every Pc chess program that is available. After
> all the initial goal of Hsu and co (creating the NPS 100,000 machine /
> version) was to check the state of art between DB and todays Pc chess
> programs.
> Right?
Only to compare evals. The platforms are in not equivalent!
> And we the "bunch of bad losers", "crying babies" should eat that cake?
> I would call it common sense.
> Eating the 10-0 cake without any other information is naive.
Just take some time to appreciate the power of VLSI.
And have a little patience to let the DB team prepare their articles.
Regards,
Marcel
-- _ _
_| |_|_|
|_ |_ Marcel van Kervinck
|_| mar...@stack.nl
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: However, before ending this point, for those of you interested in
: DB, please don't flood me with email. I've already seen about 40
: queries this morning.
: This is a lose-lose situation.
Why?
: I can post results here where everyone can see them and then put up
: with this minority view of "bullocks, that didn't hapen" or else
: respond to a large number of email queries about DB.
Can you explain the use of the word "minority" in your context?
How do you know it's a minority?
- Ed Schroder -
: Either way is a losing proposition. Which is sad, when you think
Robert Hyatt
Ed Schroder (rebc...@xs4all.nl) wrote:
: From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: : However, before ending this point, for those of you interested in
: : DB, please don't flood me with email. I've already seen about 40
: : queries this morning.
: : This is a lose-lose situation.
: Why?
Because I (a) get chastised for reporting information here, when it comes from a
source I don't have direct access to (Deep Blue in this case).. even though I believe
that the scientific facts are there to support the data. (b) if I elect
to handle the queries via email, it only adds to my email volume, which I religiously
try to respond to on a daily basis. I am averaging about 30 emails per day, plus the
2-10 "crafty mailing list" inquiries I get per day on average.
I'm still amazed by the "wow, that can't be" when it is only a modest improvement
over what we already know to be "fact"...
: : I can post results here where everyone can see them and then put up
: : with this minority view of "bullocks, that didn't hapen" or else
: : respond to a large number of email queries about DB.
: Can you explain the use of the word "minority" in your context?
: How do you know it's a minority?
Simple. I can count. Quite well, actually. Visit Deja News... and count the
number of people commenting on the DB info I posted, and divide it into two
groups. (a) those that say "no way" and (b) those that say "tell us more.."
if you count, there's more in (b). I can't read minds, so I can't comment on
who thinks what, but I can count...
: - Ed Schroder -
Tord Kallqvist Romstad
hy...@crafty.cis.uab.edu (Robert Hyatt) writes:
> I'm still amazed by the "wow, that can't be" when it is only a modest improvement
> over what we already know to be "fact"...
Come on, Bob. As several people have pointed out, going from 9-1 to
10-0 when your speed advantage decreases by a factor of 1000 is not a
"modest" improvement --- it is a very remarkable improvement, probably
the biggest improvement anyone has ever made in such a short period of
time. Although I believe in your information, I don't think it is strange
that people find this result hard to believe.
Tord
mclane
Simon Read <bl...@blurgh.bleah.retch> wrote:
>mclane, most of the words you post here do come from an English
>dictionary.
Sorry for my bad english!
>Let me see if I understand what you're saying, mclane:
>You really do talk a complete load of rubbish.
To understand images it needs a transfer action. That should be done
by the receivers brain. If there is a lack at the receiving side, the
images are not uderstood. Am I right an you don't understand me ?
>Listen, mclane, you post the biggest load of rubbish on this newsgroup.
>Bob posted something factual. You have two choices:
I have not seen him posting the 10-0 games.
Nor have I seen any log-files of deep-blue.
Where are the facts ?
>(1) Say "I believe it."
>(2) Say "I don't believe it becaue Hsu and company are a bunch of liars."
I say: I have not seen the data. Bob has maybe seen them, or heard.
Whatever. But why should I say Bob is a liar ?
If he believes in something, he is maybe lieing, but a liar KNOWS that
he lies. Somebody believing in something believes he is right.
Do you see the difference?
>>: We do want information. Otherwise we would not subscribe this
>>: newsgoup. Also we want to share this with other computerchess-freaks.
>>: But we don't want wrong information or filtered information.
>>: Is that so difficult to understand?
>What the hell are you saying here? It's like saying "We want facts, but
>not wrong information." So are you saying that Bob lied? Or are you saying
>that Hsu lied? You seem to be implying that somebody is not telling
>the truth.
If somebody tells you he has seen an ufo, would you believe him?
I would not believe him, although I believe in Ufo's myself.
So do you think I would call this guy a liar ?
No - I would say: I don't believe this. Show me evidence!
Maybe I am wrong doing this, maybe you don't understand this. Sorry.
No - filtering means: not all information has come through.
A filter lets e.g. only vertical-polarized light come through, but
stops horicontical polarized light.
Sorry Simon.
mclane
Chris Carson <chris-...@ti.com> wrote:
>I agree with Bob here. Bob only passed along information. If you
>do not like the message, don't shoot the messanger.
Information. Thats right. But i don't believe his information until I
have replayed the games myself with my genius5.
Don't you see that I am not shooting Bob. I only comment that I don't
believe his information.
>The DB team did a 10 game experiment at 100K nps. Take or leave the
>results.
No result interests me. I am interested in the event, the way, not the
result.
> 10-0 speaks for itself,
If genius was broken or the result is not worth anything.
Robert Hyatt
brucemo (bru...@seanet.com) wrote:
: Robert Hyatt wrote:
: > what about the 9-1 claim. Is that too high? I can produce game
: > scores, opponents, and everything else for those, all the games at
: > 40/2, played in the ACM and WCCC events. 10-0 is hardly an improvement
: > at all...
: It is if you drop node rate to 100K nps. It's been assumed that in the
: past they won via depth, this implies that they are winning by eval now.
: So going from winning by depth to winning (massively) via eval is not
: something that can taken on faith.
: bruce
I believe the fastest they've claimed to go is 3M nodes per second, before
DB hardware. The last results I saw from an ACM event had them slower than
that.
Also, I don't consider something given out to various journalists as "faith"
because that'd be *very* dangerous were it proven to be false, just like the
conspiricy theory about "the" match. Everyone would stand to lose much more
than they won.
However, there is evidence of eval as well, in the last match. Something
pretty unusual did happen...
Robert Hyatt
Tord Kallqvist Romstad (tor...@dromi.ifi.uio.no) wrote:
: Tord
you might be right... however, notice that in dropping their speed, they still
didn't drop down to the micro's speed. NPS, yes, but they can do whatever they want
in the hardware eval, things we don't because of the costs, or because of the
complexity. I'd much rather have a group of Ph.D.'s working with a GM or two
than working by myself... progress goes much faster when the group reaches some
critical mass...
None of which changes their result, of course, other than to say that they are
good, and their hardware gives them certain advantages that are impossible to
"handicap out."
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: Ed Schroder (rebc...@xs4all.nl) wrote:
: : From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: : : However, before ending this point, for those of you interested in
: : : DB, please don't flood me with email. I've already seen about 40
: : : queries this morning.
: : : This is a lose-lose situation.
: : Why?
: Because I (a) get chastised for reporting information here, when it
: comes from a source I don't have direct access to (Deep Blue in this
: case).. even though I believe that the scientific facts are there to
: support the data. (b) if I elect to handle the queries via email, it
: only adds to my email volume, which I religiously try to respond to
: on a daily basis. I am averaging about 30 emails per day, plus the
: 2-10 "crafty mailing list" inquiries I get per day on average.
: I'm still amazed by the "wow, that can't be" when it is only a modest
: improvement over what we already know to be "fact"...
: : : I can post results here where everyone can see them and then put up
: : : with this minority view of "bullocks, that didn't hapen" or else
: : : respond to a large number of email queries about DB.
: : Can you explain the use of the word "minority" in your context?
: : How do you know it's a minority?
: Simple. I can count. Quite well, actually. Visit Deja News...
: and count the number of people commenting on the DB info I posted,
: and divide it into two groups. (a) those that say "no way" and (b)
: those that say "tell us more.." if you count, there's more in (b).
: I can't read minds, so I can't comment on who thinks what, but I
: can count...
And were is group (c) "I believe it!"?
Does group (c) exist?
For your information count me in group (b)
*After* the answer I like to decide to move to either (a) or (c)
So better have 3 groups...
Group (a) the noway group.
Group (b) tell me more, evidence before I believe it.
Group (c) I believe it!
Let's vote and after a week or so check DejaNews again... :)
I bet group (b) will be the biggest.
- Ed Schroder -
Ed Schroder
From: mar...@stack.nl (Marcel van Kervinck)
: Ed Schroder (rebc...@xs4all.nl) wrote:
: > : : Other people were more sceptical. Having thought it over I think
they
: > : : are simply right. The claim is simply too high as I have pointed
out
: > : : in another posting.
: > : : 4-6, 5-5, 6-4 is acceptable to accept without evidence, but 10-0
: > : : without any facts needs proof otherwise it can't be taken serious.
: > : You have to believe what you have to believe. If you are suspcious
of
: > : their results, maybe you could explain why?
: > I am not suspicious these are your words...
: > And your premature conclusion of my words.
: > I never said it nor implied it.
: > Deep Blue at 100,000 NPS emulating a PP-200.
: This is where most posters seem to make their basic false assumption:
: that DB@100KNPS is somehow equivalent to a software engine on a PP200.
: Perhaps this comes from being blinded by years of general purpose
: processor programming. Please don't underestimate the power of the
: enormous amount of parallellism that one can exploit when using
: special purpose hardware. As has been pointed out in this thread
: before, a PP200 might be closer to DB@1KNPS. In that light, the
: 10-0 is not so magical as people suggest.
I don't understand.
As Bob has pointed out only *one* chip was used for the experiment. Also
this *one* chip was tuned to 100,000 NPS with as only goal to emulate a
PP-200 as close as possible.
I mean NPS 100,000 vs NPS 100,000 looks ok to me.
: > The information 10-0 implies that if the real DB would be translated
: > to Pc they will kill every Pc chess program that is available. After
: > all the initial goal of Hsu and co (creating the NPS 100,000 machine /
: > version) was to check the state of art between DB and todays Pc chess
: > programs.
: > Right?
: Only to compare evals. The platforms are in not equivalent!
Tell me what I am missing here.
What's wrong with NPS 100,000 vs NPS 100,000 ?
: > And we the "bunch of bad losers", "crying babies" should eat that
cake?
: > I would call it common sense.
: > Eating the 10-0 cake without any other information is naive.
: Just take some time to appreciate the power of VLSI.
: And have a little patience to let the DB team prepare their articles.
Was it promised then?
I mean articles about the 10-0?
- Ed Schroder -
: Regards,
brucemo
Robert Hyatt wrote:
> I'm still amazed by the "wow, that can't be" when it is only a modest improvement
> over what we already know to be "fact"...
Bob! There is no way that this conclusion is true.
In the past we had DT going at millions of NPS with a simple eval, playing against
things that go 5-35K NPS (more toward the lower end, I would guess).
Now we have this baby thing going at 100K NPS with an apparently more complicated
eval, playing against things that go 80-150K NPS.
You can't say that the latter is a natural evolution of the former, and that we
should expect improved results, because you're crippling the thing, and you're
playing against stronger opposition.
If Boeing decided to produce an economy car, and someone reported in here that it
went 700 miles per hour, you couldn't say, "Well, that's natural, seeing as their
747 goes like 550 mph, and it's been built since about 1970."
A car is not a commercial jet, you can't expect an incremental improvement over the
jet.
What is being reported here is NOT a simple evolution, it's a strange result. What
evidence do I have to say that it's not true, but why do you expect that people
won't say, "Wow, I have to understand this better?"
This result did nothing but increase controversy.
bruce
Tom C. Kerrigan
Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
> Note that my argument about Stobor was based on about 50 games I have
> in my database here. So there I had strong evidence that something was
> simply wrong about the testing. No knowledge of whether it was deliberate
> or accidental, but something was wrong. I've personally watched too many
> games vs Stobor to accept that result.
"No knowledge of whether it was deliberate or accidental"? It's good to
know that you have no moral reservations about calling data "bullshit"
when you lack such knowledge...
Hum, based on a few games I have from Crafty last year, I'm convinced your
program can't push wood out of a paper bag...
Stop trying to sugar-coat all of this. The simple fact is that you jumped
all over me, and you were wrong to do so. Don't go "Note,"-ing this and
that to other people.
Cheers,
Tom
Tom C. Kerrigan
Tord Kallqvist Romstad (tor...@refil.ifi.uio.no) wrote:
> Are pawn hash tables really a good idea? I haven't experimented with them
> myself, but I don't really like the idea. A pawn structure which is bad in
> one position, might be good if the pieces are played differently, or if a
> few pieces are traded off. A pawn structure evaluation that doesn't consider
> piece placement seems too unaccurate to me. I regard pawn hash tables as
> something similar to piece square tables --- a cheap substitute to a
> knowledgable eval. I am surprised that a programmer like Vincent, who believes
> in knowledge, uses pawn hash tables.
Right, but now count the programs that don't use piece/square tables. I
bet you'd be hard pressed to fine *1*. Point being, some of your pawn
terms aren't going to involve pieces; you can safely cache those and won't
notice a difference, save speed. It's perfectly "legal" for you to
evaluate the pawn/piece relationships at every node.
Having said that, all my pawn terms can be and are cached. :)
Cheers,
Tom
Tom C. Kerrigan
Vincent Diepeveen (vdie...@cs.ruu.nl) wrote:
> Let's be concrete, a 1000 is probably more than Fritz and Genius have,
> and for the time being this will probably crush with their speed the
> programs. I'm far over this 1000, so are some others (i suspect most
> commercial programs and most strong amateurs),
> which only makes Deep Blue program a program like we have, with
> the big difference that they are 2000 times faster.
What the hell is a 1000? Are you trying to say that Fritz and Genius have
1000 ratings? Absurd. What programs are they going to crush? How do you
know Genius is fast?
Cheers,
Tom
Tom C. Kerrigan
Andreas De Troy (Andreas...@ped.kuleuven.ac.be) wrote:
> Bob,
> I remember your furious reaction earlier this year when someone (was it
> Tom K.?) claimed that his program had beaten Crafty 10 (?) times in a row.
No, I think I posted the actual game scores of 5 games, and Crafty did get
one draw. Interestingly enough, as Bob was ripping on me for "obvious
bullshit," he was posting (to other threads) that win and loss runs of
around 5 games were't unusual at all. More evidence that Bob is 2 people?
Anyway, what would be interesting here is if somebody with The King gave
it a time handicap such that it searched 100k nodes per "second" and then
played a match against Genius and Rebel. Perhaps the results wouldn't be
10-0, but I expect them to be fairly lopsided...
Cheers,
Tom
Robert Hyatt
brucemo (bru...@seanet.com) wrote:
: Robert Hyatt wrote:
: bruce
"Wow, I'd like to understand that better" would be a natural thing. It was *exactly*
what I thought. But "wow, that is impossible, where's the games? where's the PV's"
and so forth is something else. I recall a parachute accident at an airshow last
summer, where a 'chute failed to open, and the person hit the ground, got up, and walked
away. And I thought "wow, I wonder *what* caused that?" when I saw it on the national
news. Not "wow, that's impossible".. I consider the source, and my own judgement, and
conclude for myself whether a result is possible or not. If I don't have data, but the
source seems reliable, I tend to accept it. Hsu's certainly been reliable. I don't
have anything to gain or lose by exaggerating the result... I'd think it more fun to
try and figure out "why" and not "if"...
Robert Hyatt
Ed Schroder (rebc...@xs4all.nl) wrote:
: From: mar...@stack.nl (Marcel van Kervinck)
: : Ed Schroder (rebc...@xs4all.nl) wrote:
: : > : : Other people were more sceptical. Having thought it over I think
: they
: : > : : are simply right. The claim is simply too high as I have pointed
: out
: : > : : in another posting.
: : > : : 4-6, 5-5, 6-4 is acceptable to accept without evidence, but 10-0
: : > : : without any facts needs proof otherwise it can't be taken serious.
: : > : You have to believe what you have to believe. If you are suspcious
: of
: : > : their results, maybe you could explain why?
: : > I am not suspicious these are your words...
: : > And your premature conclusion of my words.
: : > I never said it nor implied it.
: : > Deep Blue at 100,000 NPS emulating a PP-200.
: : This is where most posters seem to make their basic false assumption:
: : that DB@100KNPS is somehow equivalent to a software engine on a PP200.
: : Perhaps this comes from being blinded by years of general purpose
: : processor programming. Please don't underestimate the power of the
: : enormous amount of parallellism that one can exploit when using
: : special purpose hardware. As has been pointed out in this thread
: : before, a PP200 might be closer to DB@1KNPS. In that light, the
: : 10-0 is not so magical as people suggest.
: I don't understand.
: As Bob has pointed out only *one* chip was used for the experiment. Also
: this *one* chip was tuned to 100,000 NPS with as only goal to emulate a
: PP-200 as close as possible.
: I mean NPS 100,000 vs NPS 100,000 looks ok to me.
sort of. But imagine your eval 10x more complicated than it already is. What
would that do to your NPS? Assume you've made that 10x complications things
that are actually very important, and you think you are playing well. Now,
you want to play against a program that is searching 100K nodes per second,
but you'd only search 10K. If you could blink your eyes, and pop your NPS
up to 100K, so that nps is even, what would be the likely result? You'd
really be going 10x faster than them, correct? Because your nodes would do
more "work".
And this is how DB works. They do what they want in the eval, and it takes one
"cycle." They add mobility and it still takes one cycle. They add whatever they
want and it *still* takes one cycle. So 100K vs 100K is not exactly an "even"
comparison, although it is far more even than playing Junior or the real DB...
: : > The information 10-0 implies that if the real DB would be translated
: : > to Pc they will kill every Pc chess program that is available. After
: : > all the initial goal of Hsu and co (creating the NPS 100,000 machine /
: : > version) was to check the state of art between DB and todays Pc chess
: : > programs.
: : > Right?
: : Only to compare evals. The platforms are in not equivalent!
: Tell me what I am missing here.
: What's wrong with NPS 100,000 vs NPS 100,000 ?
: : > And we the "bunch of bad losers", "crying babies" should eat that
: cake?
: : > I would call it common sense.
: : > Eating the 10-0 cake without any other information is naive.
: : Just take some time to appreciate the power of VLSI.
: : And have a little patience to let the DB team prepare their articles.
: Was it promised then?
: I mean articles about the 10-0?
Nothing was promised to me. I was given the result, thought it interesting, and
asked if posting it would be ok. That's *all* the info I have. No secret game
scores, no inside info on how deep they searched, or anything else. Sort of like
the "blind leading the blind" actually..
Tord Kallqvist Romstad
Ed Schroder <rebc...@xs4all.nl> writes:
>
> From: hy...@crafty.cis.uab.edu (Robert Hyatt)
> : Simple. I can count. Quite well, actually. Visit Deja News...
> : and count the number of people commenting on the DB info I posted,
> : and divide it into two groups. (a) those that say "no way" and (b)
> : those that say "tell us more.." if you count, there's more in (b).
> : I can't read minds, so I can't comment on who thinks what, but I
> : can count...
>
> And were is group (c) "I believe it!"?
> Does group (c) exist?
>
> For your information count me in group (b)
> *After* the answer I like to decide to move to either (a) or (c)
>
> So better have 3 groups...
> Group (a) the noway group.
> Group (b) tell me more, evidence before I believe it.
> Group (c) I believe it!
>
> Let's vote and after a week or so check DejaNews again... :)
>
> I bet group (b) will be the biggest.
>
> - Ed Schroder -
>
>
I am somewhere between group (b) and (c). I see no reason why Hsu & co.
would fabricate results, and I feel quite certain that the 10-0 result
is correct. On the other hand, I wonder why the DB team always keeps
everything secret, and never publish any games against micros. And why
don't they ever play on the chess servers any more?
Bob, it seems that you often talk/exchange emails with Hsu. Could you
please ask him to play a 40/2 match with DB (or some crippled version)
against one of the top commercial programs, and *save the games*? If
the DB team is not able to produce a single game score, I understand
very well if lots of people stay in the (a) and (b) groups.
Tord
Simon Read
kerr...@merlin.pn.org (Tom C. Kerrigan) wrote:
>
>What the hell is a 1000? Are you trying to say that Fritz and Genius have
>1000 ratings? Absurd. What programs are they going to crush? How do you
>know Genius is fast?
So many questions! No, of course he isn't saying that Fritz and Genius
have 1000 ratings.
Perhaps if you read some of the other postings on this newsgroup
you would read that DB has over 1,000 adjustable parameters in its
evaluation function. That is the number which Vincent is talking
about.
Silly boy, try paying attention.
Simon wwwwwwwwwwwwwwwww.c.r.a.n.f.i.e.l.d.a.c.u.k/~me944p
Dan Thies
On 23 May 1997 19:31:42 GMT, hy...@crafty.cis.uab.edu (Robert Hyatt)
wrote:
>Tord Kallqvist Romstad (tor...@refil.ifi.uio.no) wrote:
>: vdie...@cs.ruu.nl (Vincent Diepeveen) writes:
>: > To speed up evaluation for some pieces i have hashtable to speed it up.
>: > Pawn hashtable speeds up a part of the pawn-evaluation considerably.
>
>: Are pawn hash tables really a good idea? I haven't experimented with them
>: myself, but I don't really like the idea. A pawn structure which is bad in
>: one position, might be good if the pieces are played differently, or if a
>: few pieces are traded off. A pawn structure evaluation that doesn't consider
>: piece placement seems too unaccurate to me. I regard pawn hash tables as
>: something similar to piece square tables --- a cheap substitute to a
>: knowledgable eval. I am surprised that a programmer like Vincent, who believes
>: in knowledge, uses pawn hash tables.
>
>They are nearly critical... but think about the evaluation a little
>differently:
IMHO.... pawn hash tables (and hash tables in general) become a lot
more critical if you want to use a lot of knowledge. In Knowchess,
pawn analysis can take as much as .5 second on my P66 (yeah, it's slow
but it was free). I don't have an evaluation score, but I do need to
store all the relevant concepts so when I hit that pawn set again, I
can just run the relevant productions right then. You see between
80-100% "hits" on a typical search, and it saves a lot of time. I'm
experimenting right now with storing piece-battery information, but
it's a little hard to decide what to store - probably I'll have to
write another set of rules so the program can decide.
>Part 1: This is simple pawn structure. Open files, backward or weak
>pawns, isolated pawns, pawn majorities, mobile pawns, passed pawns,
>connected pawns, and so forth. Evaluate all of that and hash it. If
>you want to do like I do, and have the isolated pawns score more dynamic
>(for example, an isolated pawn on an open file is weaker if there are
>rooks on the board) then you simply hash store the number of such pawns
>in the hash record, and after looking it up, fold it in when you evaluate
>a rook. Much of the pawn structure is independent of pieces, and this
>will become essentially free to compute, because 99% pawn hash hits is
>a normal result. If you want to do detailed analysis for king safety,
>do it in the pawn evaluation. Figure out which files are open, on which
>king. and even do it for both sides of the board since you won't do it
>twice for the same position. Now, when you evaluate king safety, just look
>where your king is and decide whether to use the kingside pawn shelter
>analysis, the queenside analysis, or the central analysis.
>
>All in all, it simply lets you evaluate anything that is piece independent
>*completely free*, since one evaluation and you will use that result from
>then on... Passed pawns can only be partially handled like this because
>they are typically based on the number of pieces on the board, whether they
>are blockaded or not, and so forth. But at least you can hash the file
>numbers where you have passed pawns, so that in the positional eval you
>can avoid evaluating anything but passed pawns, which will be very fast
>when there are none of course...
>
>
>: I am very interested in hearing what other programmers think about this.
>: Who uses pawn hash tables? No flames please, what I have written above is
>: just my own very humble opinion, and I have not tried pawn hash tables myself.
>Glad you said that... Had my torch lit. :) (just kidding).
>
>I've used them forever. From before Cray Blitz to the present. If you want
>to do some cute majority analysis, you can. And it won't slow you done one
>tiny bit, so long as what you analyze is piece-independent. *pay* attention
>to that. I've screwed it up more than once. If you factor in a piece's
>influence, you will get wrong results if your hash key only considers the
>locations of pawns, which is what most of us do...
>
>
>: Tord
>
Peter W. Gillgasch
Ed Schroder <rebc...@xs4all.nl> wrote:
> I don't understand.
>
> As Bob has pointed out only *one* chip was used for the experiment. Also
> this *one* chip was tuned to 100,000 NPS with as only goal to emulate a
> PP-200 as close as possible.
>
> I mean NPS 100,000 vs NPS 100,000 looks ok to me.
Not at all. Let's assume the following scenario: You have some ideas
since a couple of years for the eval of Rebel but you didn't implement
them because you would slow down Rebel by a factor of 1,000, although
those ideas would win 200 elo points on the knowledge axis. Bad thing is
you loose more than 200 elo points on the search depth axis.
But you implement them anyway and you play a X nodes match between the
two Rebels. Rebel 8 needs 3 minutes to search X nodes. Rebel BIG-EVAL
needs 3,000 minutes. The search depth axis is constant, so Rebel
BIG-EVAL would win. Of course this would be X/180 NPS versus
X/1000*180 NPS.
> Tell me what I am missing here.
You are missing that they get the factor of 1,000 in evaluation speed.
Although this is X/180 NPS versus X/180 NPS the NPS rate doesn't mean
a thing here. It is still not "equal hardware". They don't pay for the
big eval. Not a single clock cycle.
> What's wrong with NPS 100,000 vs NPS 100,000 ?
Fritz type program with 100,000 NPS versus Rebel with 100,000 NPS. That
is the difference, essentially. Who would win ? Rebel 200 MMX versus
Fritz P90 ?
-- Peter
Valavan Manohararajah
Tord Kallqvist Romstad wrote:
>
> vdie...@cs.ruu.nl (Vincent Diepeveen) writes:
> > To speed up evaluation for some pieces i have hashtable to speed it up.
> > Pawn hashtable speeds up a part of the pawn-evaluation considerably.
>
> Are pawn hash tables really a good idea? I haven't experimented with them
> myself, but I don't really like the idea. A pawn structure which is bad in
> one position, might be good if the pieces are played differently, or if a
> few pieces are traded off. A pawn structure evaluation that doesn't consider
> piece placement seems too unaccurate to me. I regard pawn hash tables as
> something similar to piece square tables --- a cheap substitute to a
> knowledgable eval. I am surprised that a programmer like Vincent, who believes
> in knowledge, uses pawn hash tables.
>
> Who uses pawn hash tables? No flames please, what I have written above is
> just my own very humble opinion, and I have not tried pawn hash tables myself.
>
I don't have pawn hash tables or any other piece hash tables in my
program,
and I still manage 40-50k (on a P166) in the middle game positions.
But, they
definitely do help from some preliminary tests I did, I am gonna convert
my
evaluator to use them soon. Although they are kind of a pain since I
had my
evaluator working in a way not suited to a pawn hash table....