Deep Blue and Bob Hyatt...
Ed Schröder
Deep Blue and Bob Hyatt...
Bob is defending Deep Blue with heart and soul, no problem.
But the truth is that Deep Blue *NEVER* proved that they have the
strongest chess engine of the world.
So why don't they show it to us?
Bob's speculation is that DB will beat ALL the best PC programs.
Well let me speculate too!
In a 10 game match at 40/2:00 Deep Blue will lose at least ONE match
against Genius, Fritz, Rebel or CM5000 ^^^
How about a bet of $1000 Bob?
40 games will show the (chess) world the state of art of Deep Blue.
Moreover we can draw conclusions about the future of computer chess.
Speed or knowlegde?
My opinion, the faster the hardware the more chess knowledge is needed.
I mean if you search 5 plies the 6th ply will bring 50 ELO or more.
But when you already search 16 ply, the 17th ply will bring almost NOTHING!
- Ed Schroder -
Robert Hyatt
Ed Schröder (10065...@CompuServe.COM) wrote:
: Deep Blue and Bob Hyatt...
:
: Bob is defending Deep Blue with heart and soul, no problem.
:
: But the truth is that Deep Blue *NEVER* proved that they have the
: strongest chess engine of the world.
:
: So why don't they show it to us?
that's a good question that only Hsu and company can answer. You'd
think it was a "black project" (top secret US GOV type). However, they
showed us Deep Thought II, and that was quite enough for me, thank you
very much. It completely dominated every event it played in, with one
notable loss in Hong Kong, due to an unfortunate opening choice. I've
been on the "crushing" end of that machine too many times already, and
have a full understanding of what the old DT II can do. I've not
played against DB yet, but I speculate that after getting run over by
a small car, I'm sure as hell not going to stand in front of a fast
moving *bus*... :)
:
: Bob's speculation is that DB will beat ALL the best PC programs.
: Well let me speculate too!
:
: In a 10 game match at 40/2:00 Deep Blue will lose at least ONE match
: against Genius, Fritz, Rebel or CM5000 ^^^
:
: How about a bet of $1000 Bob?
We'd be betting on the same side here, I posted that I thought that
DB would win 9 out of 10 games. That was the figure that Hsu posted
a year ago after the Hong Kong game, where he mentioned that he won
about 90% of the games played in his lab. He didn't mention what
happened the other 10%, whether they were losses or draws, but that
is still an impressive result. And don't forget that that was Deep
Thought II, which is at least 25-50 times slower than the old deep
blue, and maybe as much as 100-200 times slower than the new if they
have really produced the 1024 chess chips they had promised.
:
: 40 games will show the (chess) world the state of art of Deep Blue.
: Moreover we can draw conclusions about the future of computer chess.
:
: Speed or knowlegde?
:
: My opinion, the faster the hardware the more chess knowledge is needed.
: I mean if you search 5 plies the 6th ply will bring 50 ELO or more.
I think that when you go from 5 to 6, you are talking about 200-300 ELO
rating points, based on past history. 8-9 might be 50 for the right
program.
: But when you already search 16 ply, the 17th ply will bring almost NOTHING!
:
: - Ed Schroder -
Right, but suppose you don't use the speed to get another ply? Suppose you
use it to add additional selective extensions? Say, maybe instead of their
singular extension idea, they try a "double extension" where they extend if
there are only two good replies, rather than just one.
The fallacy of this argument is that you use more speed to search another
ply, but there's nothing to force anyone to use the speed in that way. As
I mentioned, Cray Blitz saw it's "hardware" go from roughly 80 MIPS on the
first Cray 1, to 16,000 MIPS on the T90 (rough approximation of course.)
On the Cray 1 we were searching 7 plies deep, by the time the XMP came
along we had reached 8-9 plies deep on hardware 4-8x faster, and there
we "stuck." We added knowledge in the evaluation, we added *lots* of
selective extensions. When I ran on the T90 last, we were *still* at
8-9 plies. Whether it was better than the old program is a matter for
conjecture, but humans that played it were impressed.
If you want to pick a smaller number, like a 7-3 split in wins vs losses
by *deep thought II* (not deep blue, I doubt we could get Hsu to turn that
loose at present) I'll take your bet, because I think DT II will win 9
games for every one game it loses. I play too many crafty vs crafty games
at huge time odds (for testing new code) to overlook that on occasion the
shallow searcher will luck into something and win.
I'll even make it easier:
In a match, we throw out all draws. If Deep Thought II wins 10 without a loss,
I win. If it wins only 8 or 9 with 2 (or 1) losses, we call it "even". If it wins
7 or less, you win. It'll have to be *very* good for me to win, and *very* bad
for you to win, chances are it will end up 9-1 or 8-2. Still overwhelming since
this is not DB of course.
Ed Schröder
Did Hsu also mentioned the playing level?
Was this on tournament level?
And howmany games were played?
And results against other programs?
And on which PC type did Hsu play?
I am sure a PP-200 is another story.
: 40 games will show the (chess) world the state of art of Deep Blue.
: Moreover we can draw conclusions about the future of computer chess.
: Speed or knowlegde?
: My opinion, the faster the hardware the more chess knowledge is needed.
: I mean if you search 5 plies the 6th ply will bring 50 ELO or more.
>I think that when you go from 5 to 6, you are talking about 200-300 ELO
>rating points, based on past history. 8-9 might be 50 for the right
>program.
I agree.
*** A C C E P T E D ***
Still ONE minor (?) problem.
I want to see Deep Blue and not DT-II.
I mean we hear SO MANY speculations but I want EVIDENCE of that.
I say the new Rebel8 is 20-30 ELO points stronger which will put
Rebel8 on top of SSDF.
I can say what I want, it still has to be proven!!
Right?
I want to see the current strength of Deep Blue, the best version available,
the maximum number of processors etc. etc. and not an old program.
I mean what's the point to play against a Rebel Decade if there is
a Rebel8 available?
Playing against DT-II would mean to play the World Champion Soccer with
plastic balls.
I want to see the real beast in action, with 10 games we can draw some
careful conclusions and I am prepared to pay a $1000 for it.
I rather would see my 40 games proposal (4 programs), but I understand
from your postings that this is not an option for Hsu?
- Ed Schroder -
Vincent Diepeveen
>Deep Blue and Bob Hyatt...
>
>
>But the truth is that Deep Blue *NEVER* proved that they have the
>strongest chess engine of the world.
>
>So why don't they show it to us?
>
>Well let me speculate too!
>
>In a 10 game match at 40/2:00 Deep Blue will lose at least ONE match
>against Genius, Fritz, Rebel or CM5000 ^^^
>
>How about a bet of $1000 Bob?
>
>Moreover we can draw conclusions about the future of computer chess.
>
>Speed or knowlegde?
>
>My opinion, the faster the hardware the more chess knowledge is needed.
>I mean if you search 5 plies the 6th ply will bring 50 ELO or more.
>
>- Ed Schroder -
I agree with Ed,
In difficult positional (correspondential) positions this is clearly visible.
In Diep i try to put as much knowledge as i can possibly implement.
Every piece of knowledge i implement clearly let it play better.
Of course at tournament level this can misperform (bad parameters?), but
when it searches say 11/12 ply, then the search corrects the parametersetting,
and the knowledge tells him that it is good to play a certain line, where
no knowledge gives an advantage of 0.00
At tournament level however it seems that parameters are more important
(Diep only searches about 8 ply at a P100 after 3 minutes), as in a game
you in the first place may never play very bad moves.
Playing humans (or positions that are typical human) it seems that
programs need all the knowledge-considerations
humans also make. for example: if a human played a line in order to
begin a minority attack, then the computer needs also to know what is a
minority attack, instead of waiting for the opponent to give away a piece.
Vincent Diepeveen
vdie...@cs.ruu.nl
--
+--------------------------------------+
|| email : vdie...@cs.ruu.nl ||
|| Vincent Diepeveen ||
+======================================+
Valavan Manohararajah
In article <511kiq$h82$1...@mhafc.production.compuserve.com>,
Ed Schröder <10065...@CompuServe.COM> wrote:
>Deep Blue and Bob Hyatt...
>
>Bob is defending Deep Blue with heart and soul, no problem.
>
Add me to the list of DB defenders.... Ok, I am no expert, but
I have had some experience with writing chess programs.
>But the truth is that Deep Blue *NEVER* proved that they have the
>strongest chess engine of the world.
>
Why not? Except for the tourney in Hong Kong, they have soundly
beaten the opponents before....
>So why don't they show it to us?
>
This is a really good question! May be they should put DB on ICC, so
GMs can play some games with it....might serve as a great debugger -
Remember that they have had all sorts of problems with Opening Books,
some castling bug etc.
>Bob's speculation is that DB will beat ALL the best PC programs.
>Well let me speculate too!
>
>In a 10 game match at 40/2:00 Deep Blue will lose at least ONE match
>against Genius, Fritz, Rebel or CM5000 ^^^
>
This is really no bet.... the dubious pruning/extensions that commercial
programs do might actually *luck out* once in a while. I have seen
CM5K do some real,real long PVs....
Why don't you give better odds for the micros? I thought you said that
DB really haven't proved that they are better - So methinks that you have
some better micro program than DB .... thats fine, just give better odds
for the match for the micros (since you think so highly of them).
>How about a bet of $1000 Bob?
>
>40 games will show the (chess) world the state of art of Deep Blue.
>Moreover we can draw conclusions about the future of computer chess.
>
>Speed or knowlegde?
>
>My opinion, the faster the hardware the more chess knowledge is needed.
>I mean if you search 5 plies the 6th ply will bring 50 ELO or more.
>But when you already search 16 ply, the 17th ply will bring almost NOTHING!
>
>- Ed Schroder -
Yes you are right, but Bob has already pointed out that the extra speed is
being used to do more selective extensions. According to Web Page, DB does
12-15 ply searches. This is real bad if it is only doing a "brute force"
search to 12-15 ply. The 12-15 ply includes a heck of extensions as well.
I wonder what kind of PVS DB would throw out in end games? 30+ line PVS?
--
-------------------------------------------------------------------------------
man...@ecf.utoronto.ca | 3rd year Comp Eng., University of Toronto
Valavan Manohararajah |
-------------------------------------------------------------------------------
Robert Hyatt
Ed Schröder (10065...@CompuServe.COM) wrote:
: : Bob is defending Deep Blue with heart and soul, no problem.
This was *IBM* so I'd suspect the fastest thing of the day. Note of course,
this was *deep thought II*, *not* Deep Blue, which is maybe 200 times faster,
wiping out any minor speed game with the P6/200. The games were longer than
blitz, but I have no idea how long. Probably 1-3 minutes per move to get a
feel for how they do at real games.
:
: I am sure a PP-200 is another story.
I think this is hopeless. I might could talk hsu into letting us have access
to DT II to play some games, but with the secrecy surrounding DB, I suspect
we are out of luck. I'd like to see it play too, but it seems we have to
wait for the next Kasparov match... :(
:
: I mean we hear SO MANY speculations but I want EVIDENCE of that.
:
: I say the new Rebel8 is 20-30 ELO points stronger which will put
: Rebel8 on top of SSDF.
:
: I can say what I want, it still has to be proven!!
: Right?
:
: I want to see the current strength of Deep Blue, the best version available,
: the maximum number of processors etc. etc. and not an old program.
:
: I mean what's the point to play against a Rebel Decade if there is
: a Rebel8 available?
:
: Playing against DT-II would mean to play the World Champion Soccer with
: plastic balls.
:
: I want to see the real beast in action, with 10 games we can draw some
: careful conclusions and I am prepared to pay a $1000 for it.
:
: I rather would see my 40 games proposal (4 programs), but I understand
: from your postings that this is not an option for Hsu?
I don't think any match is a possibility for DB. If they don't play GM's
in public, there's little chance they will play another program. <sigh>
:
: - Ed Schroder -
Peter W. Gillgasch
Ed Schröder <10065...@CompuServe.COM> wrote:
> I am sure a PP-200 is another story.
If the commercial state of the art programs have a *real* chance then
this comes from superior selectivity, and man years spent on evaluation
and book tuning. 3x speedup will not suffice, the gap is still too wide
[altough it will help a little of course].
Then Bob wrote:
> >Right, but suppose you don't use the speed to get another ply? Suppose you
> >use it to add additional selective extensions? Say, maybe instead of their
> >singular extension idea, they try a "double extension" where they extend if
> >there are only two good replies, rather than just one.
Does anybody know if they are really into this ? I know that they
speculated in a paper on variations of the singular extension scheme but
other researchers and practioneers have tried singular extensions and
after a first euphoric appreciation of the technique everybody has given
up on that idea. Bob tried this a long time ago and gave it up, Hans
tried it and found it to contribute no real improvement for Hitech, I
think Jonathan Schaeffer tried it in Phoenix but wasn't too happy with
it as well...
I tried it myself for DarkThought [took a long time, really] and found
it to be too costly. I was first impressed with it [ probably more due
to the fact that I managed to get around most implementation problems, I
did PV singular extensions only ] but we *never* used it in a tournament
since the overall impact was unclear at best.
Maybe it is ok if you are doing a couple of millions nodes per second,
but on a micro ? Any (success) stories from someone ?
I cannot see how the idea of "double extensions" can be implemented
without introducing a *lot* of additional test searches - I doubt that
this is a win even for a machine as fast as Deep Blue.
> >The fallacy of this argument is that you use more speed to search another
> >ply, but there's nothing to force anyone to use the speed in that way. As
> >I mentioned, Cray Blitz saw it's "hardware" go from roughly 80 MIPS on the
> >first Cray 1, to 16,000 MIPS on the T90 (rough approximation of course.)
That's pretty interesting. Can you quote some numbers how CB evolved
over time in terms of speed / depth reached ? Specifically how much did
you gain by tuning it to the architecture ?
[ snip ]
> *** A C C E P T E D ***
I really hope that we can see this !
> Still ONE minor (?) problem.
> I want to see Deep Blue and not DT-II.
Hm, you need a pretty big SP2 to run Deep Blue with the maximum number
of chips. I doubt that IBM will give them machine time for that. OTOH DT
2 is a simple RS/6000 with some micro channel cards that is probably
bored in the moment. Your chance :-)
> I want to see the real beast in action, with 10 games we can draw some
> careful conclusions and I am prepared to pay a $1000 for it.
Interesting bet, I am impressed, but that will probably not pay for 10
minutes of machine time on a big SP2 configuration...
Peter
--
May God grant me the serenity to accept the things I cannot change,
courage to choke the living shit out of those who piss me off,
and wisdom to know where I should hide the bodies...
Tord Kallqvist Romstad
Peter W. Gillgasch (gil...@ilk.de) wrote:
Richard Lang introduced singular extensions in his program many years
ago (in Mephisto Lyon, I believe). I recall having read somewhere that
Ed Schroder also uses singular extensions.
Tord
Robert Hyatt
:
: this comes from superior selectivity, and man years spent on evaluation
: and book tuning. 3x speedup will not suffice, the gap is still too wide
: [altough it will help a little of course].
:
: Then Bob wrote:
:
: > >use it to add additional selective extensions? Say, maybe instead of their
: > >singular extension idea, they try a "double extension" where they extend if
: > >there are only two good replies, rather than just one.
: speculated in a paper on variations of the singular extension scheme but
: other researchers and practioneers have tried singular extensions and
: after a first euphoric appreciation of the technique everybody has given
: up on that idea. Bob tried this a long time ago and gave it up, Hans
: tried it and found it to contribute no real improvement for Hitech, I
: think Jonathan Schaeffer tried it in Phoenix but wasn't too happy with
: it as well...
I tried it, but discarded it as too costly. However, when I tried it in
1978, it took my from 5 plies to 4 on a univac 1100/42 machine, which was
a *significant* depth reduction.
I want to play with it again, because apparently it's now in Genius (4.0)
and Dave Kittinger's been playing with it. Since I've noticed that Crafty
can now reacy 10-12 plies in 3 minutes, I'd be willing to lose a ply to gain
the extra depth along critical paths if, *if* it makes it play better. *Not*
doing better on WAC or whatever. More after I re-code this again in a few
months. Note that it's a bear, PV-singular extensions are easy, but the
FH-singular extensions are more difficult, but also more useful.
:
: I tried it myself for DarkThought [took a long time, really] and found
: it to be too costly. I was first impressed with it [ probably more due
: to the fact that I managed to get around most implementation problems, I
: did PV singular extensions only ] but we *never* used it in a tournament
: since the overall impact was unclear at best.
:
: Maybe it is ok if you are doing a couple of millions nodes per second,
: but on a micro ? Any (success) stories from someone ?
See above. I don't *know* that Richard does this, but I believe that Dave
Kittinger told me that Richard said he does. Hearsay I know, but from a
pretty reliable source...
:
: I cannot see how the idea of "double extensions" can be implemented
: without introducing a *lot* of additional test searches - I doubt that
: this is a win even for a machine as fast as Deep Blue.
Me either, just a comment on my part, one of the reasons I was not happy
with singular extensions. Knight check at e7, king goes to either h8 or
f8, no extension, yet both might be bad. block f8 and now the knight
check will trigger an extension because there is only one good way to
get out, moving to h8. other moves (capturing the knight, etc, are all
bad, but the one good king move is the trigger. before there were two
good king moves and it failed to trigger.)
:
: > >The fallacy of this argument is that you use more speed to search another
: > >ply, but there's nothing to force anyone to use the speed in that way. As
: > >I mentioned, Cray Blitz saw it's "hardware" go from roughly 80 MIPS on the
: > >first Cray 1, to 16,000 MIPS on the T90 (rough approximation of course.)
:
: That's pretty interesting. Can you quote some numbers how CB evolved
: over time in terms of speed / depth reached ? Specifically how much did
: you gain by tuning it to the architecture ?
I wish I could, but the one reason that Crafty exists today is the following
short story:
In the beginning. :) (couldn't resist a biblical beginning to this parable)
Cray Blitz was designed and developed over a period of about 10 years, from 1981
to roughly 1991, with little work after that to speak of. It was heavily dependent
on the Cray architecture, vectors, lots of registers, parallel processing, etc. I
kept getting the same sort of comments as the DB guys are now getting (thanks, Hsu,
I needed a break. :) ) Why don't you run that thing on a PC and see how it
compares to genius. I didn't have to. I *knew* it was slow as hell, because our
move generator, attack detector, static exchange evaluator, all were vectorized,
and ran like blazes on the Cray, but ran *slowly* otherwise. *very* slowly.
One example. On a vax 11/780 we hit around 100nps, on the Ymp, around 30,000
nodes per second, one cpu. On a PC, it might not have been able to search at
all for all I know. :)
Back to the story. Getting time on a Cray is difficult. Cray Research was
generous enough over the years to give us time for almost every chess tournament
the ACM and ICCA held, but that was it. No time for human events, very little time
for testing, tuning, and evaluation, which made progress nearly impossible. In fact,
about 1/2 of the ACM events exposed serious bugs because that was our only chance to
play whole games on the machine. I got tired of that (still am, in fact.)
So, crafty. Now I have a P6 in my office, no one dares touch it, even remotely, so
I can let crafty play it's 30,000 games per year against very strong opposition.
All because the machine is always available.
On the flip side, to answer your question, I don't have enough data. I've run
tactical tests, like WAC, which it solves trivially due to it's speed. I don't
have the results from the last full run, but with a couple of positions corrected
due to known cooks that were later repaired, it solves everything in under 15 seconds
per move, and, if I remember correctly, did 296 in 1 second or less. This sounds
pretty quick until you take Crafty, and run it on the P6, and watch it hit 295 right
in 60 seconds or less, and then you compare the costs of those two machines, one at
$60 million, one at $5,000. :)
All I can tell you about Cray Blitz is that we did lots of extensions as the years
went by, and were able to find amazing tactics as a result. Of course, we still
made positional mistakes that we worked on, but CB's strong point was simply that
it was very fast, and until deep thought came along, was really on a different
plane tactically from all other programs, excepting HiTech, where HiTech and CB
were pretty close until the C90 hit the street.
The last T90 tests I ran were pretty impressive, 5,000,000 nodes per second using
32 processors and 16 gigabytes of hash tables. The C90 was quick, but nothing like
that. We used one in one ACM event for two or three rounds and were seeing roughly
1,000,000 nodes per second. throughout all of this, our depth remained pretty much
static, because between hardware generations, we were modifying the evaluation or
adding new extensions to try and look down interesting lines to resolve them. We
were also a R=1 non-recursive null-move searcher as well. It'd be interesting to
take crafty to that machine, it might well reach 10-20 million nodes per second with
a lot of tweaking, but it'd clearly be no slower in the worst case, and the depth
would be exciting, if we can do 10-12 at 80K, a factor of 100-200 would be
interesting, since 3x3x3x3=81, that's 4 more plies, we would probably expect
at least 5 or 6 more. Anybody interested in playing crafty at a depth of 16-18
plies (humans, that is?) It'd be "stiff" to be sure.
:
: [ snip ]
:
: > *** A C C E P T E D ***
:
: I really hope that we can see this !
:
: > Still ONE minor (?) problem.
: > I want to see Deep Blue and not DT-II.
:
: Hm, you need a pretty big SP2 to run Deep Blue with the maximum number
: of chips. I doubt that IBM will give them machine time for that. OTOH DT
: 2 is a simple RS/6000 with some micro channel cards that is probably
: bored in the moment. Your chance :-)
:
: > I want to see the real beast in action, with 10 games we can draw some
: > careful conclusions and I am prepared to pay a $1000 for it.
:
: Interesting bet, I am impressed, but that will probably not pay for 10
: minutes of machine time on a big SP2 configuration...
:
I suspect they have all the SP2 time they want, but apparently they don't
want to expose it to public scrutiny for any of a dozen reasons. therefore,
a DB match probably won't happen, although I'd like to see it show up on ICC
like the old Scratchy. It was exciting, and Scratchy still holds the record
for the best won-lost ratio, losing only a couple of games against some very
stiff opposition.
Ed Schröder
[ big snip ]
>I'll even make it easier:
>In a match, we throw out all draws. If Deep Thought II wins 10 without
>a loss, I win. If it wins only 8 or 9 with 2 (or 1) losses, we call
>it "even". If it wins 7 or less, you win. It'll have to be *very* good
>for me to win, and *very* bad for you to win, chances are it will end
>up 9-1 or 8-2. Still overwhelming since this is not DB of course.
: *** A C C E P T E D ***
: Still ONE minor (?) problem.
: I want to see Deep Blue and not DT-II.
>I think this is hopeless. I might could talk hsu into letting us have access
>to DT II to play some games, but with the secrecy surrounding DB, I suspect
>we are out of luck. I'd like to see it play too, but it seems we have to
>wait for the next Kasparov match... :(
[ snip ]
>I don't think any match is a possibility for DB. If they don't play GM's
>in public, there's little chance they will play another program. <sigh>
Ok I did my best.
Deep Thought II - Rebel 8.0 than.
Can you ask Hsu?
- Ed Schroder -
Robert Hyatt
Ed Schröder (10065...@CompuServe.COM) wrote:
: [ big snip ]
:
: >I'll even make it easier:
: >In a match, we throw out all draws. If Deep Thought II wins 10 without
: >a loss, I win. If it wins only 8 or 9 with 2 (or 1) losses, we call
: >it "even". If it wins 7 or less, you win. It'll have to be *very* good
: >for me to win, and *very* bad for you to win, chances are it will end
: >up 9-1 or 8-2. Still overwhelming since this is not DB of course.
:
: : *** A C C E P T E D ***
:
: : Still ONE minor (?) problem.
: : I want to see Deep Blue and not DT-II.
:
: >to DT II to play some games, but with the secrecy surrounding DB, I suspect
: >we are out of luck. I'd like to see it play too, but it seems we have to
: >wait for the next Kasparov match... :(
:
: [ snip ]
:
: >I don't think any match is a possibility for DB. If they don't play GM's
: >in public, there's little chance they will play another program. <sigh>
:
:
: Ok I did my best.
:
: Deep Thought II - Rebel 8.0 than.
:
:
: - Ed Schroder -
Will Do...
Chris Whittington
tor...@ifi.uio.no (Tord Kallqvist Romstad) wrote:
>
> Peter W. Gillgasch (gil...@ilk.de) wrote:
> : Ed Schröder <10065...@CompuServe.COM> wrote:
> : > I am sure a PP-200 is another story.
>
> : If the commercial state of the art programs have a *real* chance then
> : this comes from superior selectivity, and man years spent on evaluation
> : and book tuning. 3x speedup will not suffice, the gap is still too wide
> : [altough it will help a little of course].
> :
> : Then Bob wrote:
>
> : > >Right, but suppose you don't use the speed to get another ply? Suppose you
> : > >use it to add additional selective extensions? Say, maybe instead of their
> : > >singular extension idea, they try a "double extension" where they extend if
> : > >there are only two good replies, rather than just one.
>
> : Does anybody know if they are really into this ? I know that they
> : speculated in a paper on variations of the singular extension scheme but
> : other researchers and practioneers have tried singular extensions and
> : after a first euphoric appreciation of the technique everybody has given
> : up on that idea. Bob tried this a long time ago and gave it up, Hans
> : tried it and found it to contribute no real improvement for Hitech, I
> : think Jonathan Schaeffer tried it in Phoenix but wasn't too happy with
> : it as well...
>
> : it to be too costly. I was first impressed with it [ probably more due
> : to the fact that I managed to get around most implementation problems, I
> : did PV singular extensions only ] but we *never* used it in a tournament
> : since the overall impact was unclear at best.
>
> : Maybe it is ok if you are doing a couple of millions nodes per second,
> : but on a micro ? Any (success) stories from someone ?
>
> ago (in Mephisto Lyon, I believe). I recall having read somewhere that
> Ed Schroder also uses singular extensions.
>
> Tord
>
Chess System Tal uses singular and double extensions (and sometimes
triples). But I think I do it differently, I use the evaluation
function to determine the number of 'good' moves available in
king attack (and other)situations. Then I determine the 'strength' of
the king attack, before deciding how many and which candidates to
look at.
This seems to give good results. But, sometimes, especially when
the position has become very wild, then extension rate
spirals from its usual 25% to 50% of the 'normal' search up to
1000% or 2000%. If the attack is good then this is no problem, but
if not, then ......
Chris Whittington
mark glickman
In article <511kiq$h82$1...@mhafc.production.compuserve.com>,
Ed Schröder <10065...@CompuServe.COM> wrote:
>against Genius, Fritz, Rebel or CM5000 ^^^
>
>How about a bet of $1000 Bob?
If Deep Blue outrates these four machines by about 150 points,
then there's a 70% chance of winning individual games, roughly 85%
chance of winning 10 game matches, and roughly 51% chance of
winning all four 10 game matches.
So even if Deep Blue is better (by less than 150 rating points),
it will not likely defeat all 4 opponents.
Bob - don't make the bet!! :-)
- Mark
Peter W. Gillgasch
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote:
> Ed Schröder (10065...@CompuServe.COM) wrote:
[ razor ]
> : *** A C C E P T E D ***
> :
> : Still ONE minor (?) problem.
> : I want to see Deep Blue and not DT-II.
>
> I think this is hopeless. I might could talk hsu into letting us have access
> to DT II to play some games, but with the secrecy surrounding DB, I suspect
> we are out of luck. I'd like to see it play too, but it seems we have to
> wait for the next Kasparov match... :(
[ snip ]
> I don't think any match is a possibility for DB. If they don't play GM's
> in public, there's little chance they will play another program. <sigh>
Now don't give up so fast ! Do you think that you can pester^H^H^H^H^H^H
ask FHH successfully to make this experiment happen ?
It is fairly obvious that we probably can't have Deep Blue to play a
match against Rebel 8, but hopefully Ed is interested in playing versus
the little brother as well ?! C'mon, r.g.c.c. would love to see this
happening and although the marketing effect would be substantially
higher if you win a game or two versus Deep Blue isn't that interesting
enough to go for it ?
-- Peter
mcl...@prima.ruhr.de
>Deep Blue and Bob Hyatt...
>
>Bob is defending Deep Blue with heart and soul, no problem.
>
>But the truth is that Deep Blue *NEVER* proved that they have the
>strongest chess engine of the world.
>
>So why don't they show it to us?
>
>Bob's speculation is that DB will beat ALL the best PC programs.
>Well let me speculate too!
>
>In a 10 game match at 40/2:00 Deep Blue will lose at least ONE match
>against Genius, Fritz, Rebel or CM5000 ^^^
>
>How about a bet of $1000 Bob?
>
>40 games will show the (chess) world the state of art of Deep Blue.
>Moreover we can draw conclusions about the future of computer chess.
>
>Speed or knowlegde?
>
>My opinion, the faster the hardware the more chess knowledge is needed.
>I mean if you search 5 plies the 6th ply will bring 50 ELO or more.
>But when you already search 16 ply, the 17th ply will bring almost NOTHING!
>
>- Ed Schroder -
I totally agree.
One thing to the speed:
Always and always it is wrongly said:
Why should we reduce DB's speed???
You should not reduce the speed of db to help the slow-pc-programs to win.
But: in sport BOTH sides should have fair and the same conditions.
You cannot let Rebel8 run on PentiumPro200 and DB on it's fast hardware.
Thats unfair.
Both oponents should have the same hardware.
I guess this is not unfair.
Chris mentioned to reduce the nodes per second.
I don't like this idea, it suggest chris would like to decrease deep-blues
strength by force.
If it is not possible to reduce deep-blues speed, or not possible to
increase Rebel8s hardware, than we cannot compare.
But since we cannot compare, we cannot say: DB is stronger!
DB is strong BECAUSE OF ITS HARDWARE AND *Sophisticated software* !
But a fair competition will let them run on the same hardware level.
Otherwise we don't compare chess-strength but hardware-power.
I don't know if you can understand the point, because past has shown
you don't understand it, but you cannot compare things that use
different scales.
Robert Hyatt
mark glickman (glic...@fas.harvard.edu) wrote:
: In article <511kiq$h82$1...@mhafc.production.compuserve.com>,
: Ed Schröder <10065...@CompuServe.COM> wrote:
: >In a 10 game match at 40/2:00 Deep Blue will lose at least ONE match
: >against Genius, Fritz, Rebel or CM5000 ^^^
: >
: >How about a bet of $1000 Bob?
: If Deep Blue outrates these four machines by about 150 points,
: then there's a 70% chance of winning individual games, roughly 85%
: chance of winning 10 game matches, and roughly 51% chance of
: winning all four 10 game matches.
:
: So even if Deep Blue is better (by less than 150 rating points),
: it will not likely defeat all 4 opponents.
:
: Bob - don't make the bet!! :-)
:
: - Mark
I think it's *much* better than 150 points. I think at least 300.
The odds are not bad in that case... :)
Robert Hyatt
:
: > Ed Schröder (10065...@CompuServe.COM) wrote:
: [ razor ]
: > :
: > : Still ONE minor (?) problem.
: > : I want to see Deep Blue and not DT-II.
: >
: > I think this is hopeless. I might could talk hsu into letting us have access
: > to DT II to play some games, but with the secrecy surrounding DB, I suspect
: > we are out of luck. I'd like to see it play too, but it seems we have to
: > wait for the next Kasparov match... :(
:
:
: > I don't think any match is a possibility for DB. If they don't play GM's
: > in public, there's little chance they will play another program. <sigh>
:
: ask FHH successfully to make this experiment happen ?
:
: It is fairly obvious that we probably can't have Deep Blue to play a
: match against Rebel 8, but hopefully Ed is interested in playing versus
: the little brother as well ?! C'mon, r.g.c.c. would love to see this
: happening and although the marketing effect would be substantially
: higher if you win a game or two versus Deep Blue isn't that interesting
: enough to go for it ?
:
: -- Peter
We are going to try to set this up, as I previously posted. I suspect that
if we get anything, it will be Deep Thought II, assuming it is still in one
piece and usable. I suspect it is as it would make a good sparring partner
for DB if nothing else.
BTW, my comments above were regarding the likely hopelessness of getting
DB to play. DT II is a higher probability, but I'm not sure just how "high"
that is...
More when I know more...
Bob
Robert Hyatt
mcl...@prima.ruhr.de wrote:
: In article <511kiq$h82$1...@mhafc.production.compuserve.com>, Ed Schröder
:
: >Deep Blue and Bob Hyatt...
: >
: >Bob is defending Deep Blue with heart and soul, no problem.
: >
: >But the truth is that Deep Blue *NEVER* proved that they have the
: >strongest chess engine of the world.
: >
: >So why don't they show it to us?
: >
: >Bob's speculation is that DB will beat ALL the best PC programs.
: >Well let me speculate too!
: >In a 10 game match at 40/2:00 Deep Blue will lose at least ONE match
: >against Genius, Fritz, Rebel or CM5000 ^^^
: >
: >How about a bet of $1000 Bob?
: >
: >Moreover we can draw conclusions about the future of computer chess.
: >
: >Speed or knowlegde?
: >
: >My opinion, the faster the hardware the more chess knowledge is needed.
: >I mean if you search 5 plies the 6th ply will bring 50 ELO or more.
: >But when you already search 16 ply, the 17th ply will bring almost NOTHING!
: >
: >- Ed Schroder -
:
:
: I totally agree.
:
: One thing to the speed:
:
: Always and always it is wrongly said:
: Why should we reduce DB's speed???
:
: You should not reduce the speed of db to help the slow-pc-programs to win.
: But: in sport BOTH sides should have fair and the same conditions.
:
: You cannot let Rebel8 run on PentiumPro200 and DB on it's fast hardware.
: Thats unfair.
: Both oponents should have the same hardware.
: I guess this is not unfair.
:
: Chris mentioned to reduce the nodes per second.
: I don't like this idea, it suggest chris would like to decrease deep-blues
: strength by force.
:
: If it is not possible to reduce deep-blues speed, or not possible to
: increase Rebel8s hardware, than we cannot compare.
Aha, you *finally* "see the light"... I've only said this a hundred times.
:
: But since we cannot compare, we cannot say: DB is stronger!
:
: DB is strong BECAUSE OF ITS HARDWARE AND *Sophisticated software* !
And then you backtrack and come up with this conclusion. All we have to do
is play rebel (or any program, take Crafty for example to avoid having Ed think
I'm picking on *his* program which is quite good) vs DB. The winner is better.
*period*. Why the qualification? The hardware is *part* of DB.
:
: But a fair competition will let them run on the same hardware level.
:
: Otherwise we don't compare chess-strength but hardware-power.
So, they've spent 10 years designing and building fast hardware, and they
get to throw all that work out the window to have a fair match? What does the
PC program have to throw out to make this equal? hmmm.... They get to throw
out all their code, so DB wins by default because the PC suddenly has no
program to make moves? Makes a lot of sense...
:
: I don't know if you can understand the point, because past has shown
: you don't understand it, but you cannot compare things that use
: different scales.
You can if you use an absolute scale: take one 64 square chessboard, with 16
pieces per side, arranged as per the rules of chess, and let two players that
follow these rules play a game, or a match of games. the winner is better, the
scale is not 'relative', and we can conclude which plays the stronger game of
chess. No discussion about the hardware, the software, etc. All of that is
unimportant when you look at the rules of chess. If you have 2x the neurons that
I do, I don't make you undergo a lobotomy before we play, we just play. *period*.
Peter W. Gillgasch
<mcl...@prima.ruhr.de> wrote:
[ snip ]
> One thing to the speed:
>
> Always and always it is wrongly said:
> Why should we reduce DB's speed???
>
> You should not reduce the speed of db to help the slow-pc-programs to win.
Certainly not. In this post I am trying pretty hard to explain my point
of view to you. This is a lengthy post, I suggest that you read it
carefully in full length and think about it. I have spent some time
writing it so *please* give it a serious try.
> But: in sport BOTH sides should have fair and the same conditions.
Of course. But this is no sport between athletes, this is competition
between *systems*.
Part of the system is the software, part of it is the hardware. A very
important part of the system is the human work invested into this
system. Let's look into this a bit deeper just to get an idea what I
mean with that.
> You cannot let Rebel8 run on PentiumPro200 and DB on it's fast hardware.
> Thats unfair.
> Both oponents should have the same hardware.
> I guess this is not unfair.
There is nothing like unfairness in this as far as I can tell.
Just to avoid excessive name dropping let's introduce three hypothetical
computer chess teams into the discussion:
The Y team - those use an off the shelf Pentium system
The X team - they use XYZ which is a 1,000 mhz Pentium system [ignore
the fact that this system doesn't exist in reality].
The *X team - they use *XYZ - a super computer with 1024 off the shelf
Pentiums, but those are interconnected. A kick ass
parallel machine.
I probably could understand your point in the following situation.
If (big if) a program X is running on an XYZ supercomputer which is
binary compatible to the Pentium but is still a serial machine (e.g. if
some arbitrary chess program is running on a GaAs implementation of the
Pentium clocked at 1 GHz) then I'd be the first to admit that comparing
it with another program running at a more usual system (the Y program)
doesn't tell us anything and is *probably* from the human point of view
(e.g. from the point of view of the programers) unfair since the
programmers of X simply draw on the speed of XYZ.
So in principle I can understand your argument. But of course I wouldn't
be the bad brute force boy Gillgasch if... So let's look at another
scenario:
Now if *XYZ is a set of 1024 normal Pentiums plus an interconnection
network things would be (for me that is) *entirely different*. The *X
team would need a *tremendous* amount of work to invest into utilizing
such a machine, they *did* go that hard road and if they then crush any
serial chess program....
Now if the Y team [or the loyal fanatic following of the Y team] shouts
"unfair" then the *X team is treated unfairly, because they invested
serious work that is *as good* as the work of the other teams [they have
concentrated their effort in a different area, but there really is no
difference].
The crux with FHH's toys is the following: A huge part of the work that
was invested into this was invested into a special purpose hardware
architecture. Their speed did *not* fall from the sky. They *did work
for it*. Trying to eliminate that part of their work is utterly unfair.
Can you accept that ? I hope so.
> Chris mentioned to reduce the nodes per second.
> I don't like this idea, it suggest chris would like to decrease deep-blues
> strength by force.
>
> If it is not possible to reduce deep-blues speed, or not possible to
> increase Rebel8s hardware, than we cannot compare.
Now here is the point where we probably will disagree upon.
The target of the development of all chess playing systems is to
maximize playing strength. Traditionally playing strength is measured by
playing games. So we can compare.
> But since we cannot compare, we cannot say: DB is stronger!
But since we can compare we will be able to say who is stronger.
> DB is strong BECAUSE OF ITS HARDWARE AND *Sophisticated software* !
Exactly.
> But a fair competition will let them run on the same hardware level.
Not neccessarily. The computer chess competitions are traditionally
split into two classes:
1 - The unlimited class
2 - The micro class
Every participant of (2) can also play in the (1) competition.
DarkThought did that last year. In the (1) competition you have to
accept that some of the participants have a *tremendous* advantage in
raw processing power [it you don't accept that you simply don't
participate].
> Otherwise we don't compare chess-strength but hardware-power.
Ah. NOW we come to the point. Chess strength and hardware-power are
related, so how do we compare ? Even worse, *if* we compare *what* do we
compare then ?
I agree with you 100% for the scenario that we run the Y program on the
Pentium and pair it versus itself running on the XYZ machine. In that
case we compare horsepower only.
If we pair X running on Y's machine versus Y running on the XYZ machine
then we compare two programs running in environments that are
*significantly* different to the environments they have been developed
for. We don't need to discuss unfairness in this case since the
contribution of the X team to the state of the art was basically flawed
by their decision to use the XYZ machine.
But things are not so easy. If we pair Y versus *X running on *XYZ.
What do we compare then ?
Food for thought.
> I don't know if you can understand the point, because past has shown
> you don't understand it, but you cannot compare things that use
> different scales.
I don't know if you can understand the point, because past has shown
you don't understand it, but you can compare things that are
build for achieving the same goal:
TO WIN AT CHESS
Peter W. Gillgasch
Tord Kallqvist Romstad <tor...@ifi.uio.no> wrote:
> Peter W. Gillgasch (gil...@ilk.de) wrote:
[ snipped some stuff about singular extensions ]
> : Maybe it is ok if you are doing a couple of millions nodes per second,
> : but on a micro ? Any (success) stories from someone ?
>
> Richard Lang introduced singular extensions in his program many years
> ago (in Mephisto Lyon, I believe). I recall having read somewhere that
> Ed Schroder also uses singular extensions.
I believe you are correct.
[ wades through a pile of ICCAJs looking for something. Ah. ]
ICCAJ 2/Vol. 14 [June 1991] contains two articles that are pretty
interesting:
T.S. Ananthamaran, Extension heuristics
[this is the guy who wrote the original ChipTest / Deep Thought search
software, __very__ interesting to read, maybe one of the best articles
ever published in the ICCAJ].
D.F. Beal, Report on the 11th World Micro Computer Chess Championship
This article contains a detailed description of the participating
programs. Quite interesting [pages 88 -90]. This event happened at a
time during which everybody was excited with the singularity idea.
I'll quote the stuff on extensions for the more notable programs and I
also add the nps rate to see if there is some correlation:
Gideon [Ed Schroeder]
"check extensions and singular extensions"
nps: 6,800
M-Chess [Marty Hirsch]
"check extensions and recapture extensions, but no singular extensions"
nps: 4,500
Mephisto [Richard Lang]
"check extensions and singular extensions"
nps: 10,000
The King [Johan de Koning]
"check extensions and chess-specific static rules for additional
extensions. Singular extensions are not used..."
nps: 4,000
Saitek-X [Dan & Kathe Spracklen]
"check extensions and singular extensions, but no other extension rules"
nps: 10,000
Hiarcs [Mark Uniacke]
"check extensions, singular extensions on a static analysis only, chess
specific 'part ply' extensions"
nps: 2,500
The other programs don't use singular extensions [they are all below
5,000 nps, most of them around 1,000 nps]. It is noteworthy that the
fastest programs used singular extensions at that time. Hiarcs was
rather slow and obviously avoided the cost of test searches by using the
evaluation for singularity decisions.
I never tried the later idea since I felt that the very idea of
singularity is to use the search to cover cases which I didn't consider
when writing the evaluation [the usual "the system is more powerful than
the sum of its components" belief of mine which is not understood in
some circles].
Simple question. Did the idea go away over time or is it getting more
attractive again as processing power increases ?
Obviously Ed could comment on that ;-)
-- Peter
Simon Read
OK, another comparison: we're trying to get the clock
speeds of the two platforms equal. So, let's reduce the
number of revolutions per minute of the car engine so
that it matches the number of beats-per-minute of the
horse's heart. Then we can see if they perform equally.
That would be a fair comparison.
Alternatively, let's make the diameter of the car's
wheels equal to the length of the horse's legs, and
make the car's wheels rotate at the same number of
revolutions per minute as the horse makes strides per
minute. That would be a fair comparison.
"The car is faster."
"But it's polluting!"
"The car is faster."
"But it makes a lot of noise!"
"The car is faster."
"But it's the wrong way to do it! It's unnatural."
"The car is faster."
"But horses have common sense. No horse would ever go off a
cliff-top, or crash into another horse, or ram a shop
window."
"The car is faster."
"Someday, we will learn how to build cars like horses, then
it will be done properly. Such a machine will outrun everything."
This argument will run and run... or maybe it will roll and roll...
Simon
Simon Read
OK, let me suggest something:
I want to find out if a car can run faster than a horse.
Why don't you try to port the car's software to the horse's
body, then both can be put on an equal-hardware footing,
so as to make a fair comparison?
Simon
Shawn K. Quinn
On 11 Sep 1996 04:08:13 GMT, Robert Hyatt <hy...@crafty.cis.uab.edu> wrote:
>If you have 2x the neurons that
>I do, I don't make you undergo a lobotomy before we play, we just play.
>*period*.
Well said. I could only imagine what the world would be like if one did have
to undergo a lobotomy before a game... :-)
Anyway, the questions I have are these:
1. Would it be possible to port DT II to a PC platform such as Linux?
2. If anyone knows, what language is the program behind DT II (and presumably
DB) written in?
SKQ
Robert Hyatt
Shawn K. Quinn (skq...@brokersys.com) wrote:
No. The problem is that while DT and DB are essentially "C" programs, they
have a built-in reliance on their special purpose hardware chips. These
chips do two things: (1) one maintains the state of the actual game board,
generates moves, makes/unmakes moves, etc; (2) once the proper depth is
reached, a chess processor searches this position using normal search rules
until depth=0, then executes a typical capture search, and returns the value
for that position.
These "chips" are so ingrained into the program, that removing them would
require a software make/unmake/move generator, and would so slow down the
program that it wouldn't really be an interesting machine to study. They've
invested a great deal of time and energy into developing these chips, which
is why I say they are "inseparable". Sort of like my taking Crafty, removing
the various search extensions and evaluation stuff I've done, the rotated
bitmaps, and anything else that has been something that I've developed over
time on my own, by testing, because then I'd have a plain vanilla chess
program. The parts that I have spent a lot of time on (the non-incremental
rotated bitmaps for example) are what makes Crafty unique. To remove them for
equality doesn't make much sense. In DT/DB's case, it would be horribly
crippling and it wouldn't be "deep blue on a PC" at all...
Bob
Vincent Diepeveen
>>Deep Blue and Bob Hyatt...
>>Bob is defending Deep Blue with heart and soul, no problem.
>I have had some experience with writing chess programs.
Add me to the list of guys who think DB is a looser.
>>But the truth is that Deep Blue *NEVER* proved that they have the
>>strongest chess engine of the world.
>Why not?
My guess: the DB people are
a) afraid to loose
b) want to prevent opponents from overcoming the computer shock.
So not to loose any information about its weaknesses. For example: suppose
DB cannot handle closed positions, and it seems very hard to work on this,
would you show it to the world?
c) financial problems+delay with hardware.
d) they have problems singleprocessor programs don't have
e) they are in no hurry to beat or loose anyone; they get paid to
do something, and not to do a lot (?)
f) they don't know much about chess, so they need more time to implement
chessknowledge.
g) they're amateurs
> Except for the tourney in Hong Kong, they have soundly
>beaten the opponents before....
That was another time in history, when the micro programs did have a lot
of troubles to even see a childrens trick. In that time outsearching
tactically was a great problem. Even amateurs at fast hardware could
at that time easily win from micros.
With the introduction of the PP200 this outsearch problem has been reduced.
>>So why don't they show it to us?
>>
>
>This is a really good question! May be they should put DB on ICC, so
>GMs can play some games with it....might serve as a great debugger -
>Remember that they have had all sorts of problems with Opening Books,
>some castling bug etc.
>
>>Bob's speculation is that DB will beat ALL the best PC programs.
>>Well let me speculate too!
>>
>>In a 10 game match at 40/2:00 Deep Blue will lose at least ONE match
>>against Genius, Fritz, Rebel or CM5000 ^^^
>>
>
>This is really no bet.... the dubious pruning/extensions that commercial
>programs do might actually *luck out* once in a while. I have seen
>CM5K do some real,real long PVs....
>
>Why don't you give better odds for the micros? I thought you said that
>DB really haven't proved that they are better - So methinks that you have
>some better micro program than DB .... thats fine, just give better odds
>for the match for the micros (since you think so highly of them).
>
>>How about a bet of $1000 Bob?
>>
>>40 games will show the (chess) world the state of art of Deep Blue.
>>Moreover we can draw conclusions about the future of computer chess.
>>
>>Speed or knowlegde?
In Diep i bet on knowledge, although we must not forget that most programs
lack knowledge because the implementing programmer lacks it. On the other
hand Diep is slow, causing it to search less ply than for example DB, which
is a tactical drawback; you must always win a game positionally. However,
Diep's style and games it did win look very human.
>>My opinion, the faster the hardware the more chess knowledge is needed.
>>I mean if you search 5 plies the 6th ply will bring 50 ELO or more.
>>But when you already search 16 ply, the 17th ply will bring almost NOTHING!
>>
>>- Ed Schroder -
>
>Yes you are right, but Bob has already pointed out that the extra speed is
>being used to do more selective extensions. According to Web Page, DB does
>12-15 ply searches. This is real bad if it is only doing a "brute force"
>search to 12-15 ply. The 12-15 ply includes a heck of extensions as well.
>
>I wonder what kind of PVS DB would throw out in end games? 30+ line PVS?
30+ ply is of course an advantage, but we must not forget that it is only
a tactical advantage. It will not tell DB what move to choose from, if
it has 2 legal moves, which both don't loose tactically.
Vincent Diepeveen
In <199609101607311842200@[194.121.104.142]> gil...@ilk.de (Peter W. Gillgasch) writes:
>> I am sure a PP-200 is another story.
>
>If the commercial state of the art programs have a *real* chance then
>this comes from superior selectivity, and man years spent on evaluation
>and book tuning. 3x speedup will not suffice, the gap is still too wide
>[altough it will help a little of course].
>
>Then Bob wrote:
>
>> >Right, but suppose you don't use the speed to get another ply? Suppose you
>> >use it to add additional selective extensions? Say, maybe instead of their
>> >singular extension idea, they try a "double extension" where they extend if
>> >there are only two good replies, rather than just one.
>
>Does anybody know if they are really into this ? I know that they
>speculated in a paper on variations of the singular extension scheme but
>other researchers and practioneers have tried singular extensions and
>after a first euphoric appreciation of the technique everybody has given
>up on that idea. Bob tried this a long time ago and gave it up, Hans
>tried it and found it to contribute no real improvement for Hitech, I
>think Jonathan Schaeffer tried it in Phoenix but wasn't too happy with
>it as well...
>
>I tried it myself for DarkThought [took a long time, really] and found
>it to be too costly. I was first impressed with it [ probably more due
>to the fact that I managed to get around most implementation problems, I
>did PV singular extensions only ] but we *never* used it in a tournament
>since the overall impact was unclear at best.
>Maybe it is ok if you are doing a couple of millions nodes per second,
>but on a micro ? Any (success) stories from someone ?
I use a modified version of singular extensions in Diep (1 of my modifications:
if you get a cutoff, then i give a cutoff, and don't search further for
singularity), and this works
fine, although at a high prize.
The reason i use singular extensions is NOT because my modified
(but still) singular extensions gives that great results, but because
it PREVENTS diep from going into for example a mate in 13.
A big problem in singular extensions if for example from opening
when searching for example:
1. e4,f6? 2. Qh5? g6 3. ...
Diep thinks 2...,g6 a singular move: it does not give a cutoff (because
f6 gives black a bad kingsafety), but it is way better than all other
moves.
Now this is a simple case, where you can even think out how to prevent it,
but there are numerous examples of positions you didn't even think of,
which it extends.
The problem of singular extension is that you extend so many positions
where a move is simply a forced move (which is the
idea of singular extensions).
The problem is that in Diep the move Qh5 gives a cutoff (stored in killermoves
etc), and that for example 2..,g6 does NOT give a cutoff, where i would
like to see a more human algorithm which gives also g6 a cutoff,
and then Qh5 a cutoff; there is no mechanism that gives 'killermoves' a cutoff,
if you have done bad moves in the variation already. If moves like g6 would
give a cutoff, then singular extensions would work really great.
>I cannot see how the idea of "double extensions" can be implemented
>without introducing a *lot* of additional test searches - I doubt that
>this is a win even for a machine as fast as Deep Blue.
I have very bad experiences with double extensions. Extensions are only
useful to overcome tactically the horizon. If you search already 15 ply,
then i don't see why one should extend ANYTHING!
I wouldn't even extend check in most cases!
>> >The fallacy of this argument is that you use more speed to search another
>> >ply, but there's nothing to force anyone to use the speed in that way. As
>> >I mentioned, Cray Blitz saw it's "hardware" go from roughly 80 MIPS on the
>> >first Cray 1, to 16,000 MIPS on the T90 (rough approximation of course.)
>That's pretty interesting. Can you quote some numbers how CB evolved
>over time in terms of speed / depth reached ? Specifically how much did
>you gain by tuning it to the architecture ?
>[ snip ]
>
>> *** A C C E P T E D ***
>
>I really hope that we can see this !
>
>> Still ONE minor (?) problem.
>> I want to see Deep Blue and not DT-II.
>
>Hm, you need a pretty big SP2 to run Deep Blue with the maximum number
>of chips. I doubt that IBM will give them machine time for that. OTOH DT
>2 is a simple RS/6000 with some micro channel cards that is probably
>bored in the moment. Your chance :-)
Whether you search 50,000,000 nodes a second or 1000,000,000. It will not
affect the outcome of the match. I think Rebel makes a great chance,
especially when Ed uses certain lines from the Rebel book... :)
>> I want to see the real beast in action, with 10 games we can draw some
>> careful conclusions and I am prepared to pay a $1000 for it.
>Interesting bet, I am impressed, but that will probably not pay for 10
>minutes of machine time on a big SP2 configuration...
>
> Peter
So why playing Kasparov with DB?
Too expensive, you loose another few hundreds of thousands!
--
...i think i can win a match from Kasparov... ...i have a big mouth,
i have bigger glasses, i have bigger chess pieces, i have a big
chessboard, i am more optimistic, i have more openingsbooks,
i have bigger shoes... ...and a sponsor :)
>--
>May God grant me the serenity to accept the things I cannot change,
>courage to choke the living shit out of those who piss me off,
>and wisdom to know where I should hide the bodies...
--
Peter Noehrenberg
Robert Hyatt wrote:
>
> Ed Schröder (10065...@CompuServe.COM) wrote:
> :
> : Deep Thought II - Rebel 8.0 than.
> :
> :
> : - Ed Schroder -
>
> Will Do...
An' how 'bout staging the match LIVE on one of the free servers
(ie FICS) so we can all watch and cheer our favorite on...
--
PCN
Q: What is the difference between
capitalism and socialism?
A: Under capitalism, man exploits man.
Under socialism, it's just the reverse.
--Anonymous
jack_clerk
6 games per year is a small sample size. But from these few games, can anyone
pinpoint a move that DB made which a PC wouldn't have - and which prevented the
loss of a game?
I think that most people would be hard pressed to find such a move.
In article <514vup$b...@juniper.cis.uab.edu>, hy...@crafty.cis.uab.edu says...
>
>mark glickman (glic...@fas.harvard.edu) wrote:
>: In article <511kiq$h82$1...@mhafc.production.compuserve.com>,
>: Ed Schröder <10065...@CompuServe.COM> wrote:
>: >In a 10 game match at 40/2:00 Deep Blue will lose at least ONE match
>: >against Genius, Fritz, Rebel or CM5000 ^^^
Robert Hyatt
Vincent Diepeveen (vdie...@cs.ruu.nl) wrote:
: >>Deep Blue and Bob Hyatt...
: >>Bob is defending Deep Blue with heart and soul, no problem.
: >Add me to the list of DB defenders.... Ok, I am no expert, but
: >I have had some experience with writing chess programs.
:
: Add me to the list of guys who think DB is a looser.
:
: >>But the truth is that Deep Blue *NEVER* proved that they have the
: >>strongest chess engine of the world.
:
: >Why not?
:
: My guess: the DB people are
: a) afraid to loose
Possibly. We can't read their minds, however, so...
: b) want to prevent opponents from overcoming the computer shock.
: So not to loose any information about its weaknesses. For example: suppose
: DB cannot handle closed positions, and it seems very hard to work on this,
: would you show it to the world?
very probable answer.
: c) financial problems+delay with hardware.
Not likely. hardware was working fine. doubt there's suddenly a problem
with it or with money since it's already working.
: d) they have problems singleprocessor programs don't have
true, but they have been solved pretty well based on the last match,
parallel processing was not a problem I don't think.
: e) they are in no hurry to beat or loose anyone; they get paid to
: do something, and not to do a lot (?)
that makes little sense. I doubt IBM will let this go on forever without
some sense of progress.
: f) they don't know much about chess, so they need more time to implement
: chessknowledge.
Hmm. GM Joel Benjamin is working with them, so I suppose this is not exactly
close to the mark...
: g) they're amateurs
And this is an outright insult. To call someone an amateur, you have to be
something more than that yourself. You were doing o.k. until here, but now
look like a kid throwing a tantrum again.
:
: > Except for the tourney in Hong Kong, they have soundly
: >beaten the opponents before....
:
: That was another time in history, when the micro programs did have a lot
: of troubles to even see a childrens trick. In that time outsearching
: tactically was a great problem. Even amateurs at fast hardware could
: at that time easily win from micros.
:
: With the introduction of the PP200 this outsearch problem has been reduced.
check your math. It's been reduced by a *miniscule* amount. Instead of being
3,000 times slower, now PCs are only 1,000 times slower. One hell of a big
difference.
:
: >>So why don't they show it to us?
: >>
: >
: >This is a really good question! May be they should put DB on ICC, so
: >GMs can play some games with it....might serve as a great debugger -
: >Remember that they have had all sorts of problems with Opening Books,
: >some castling bug etc.
: >
: >>Bob's speculation is that DB will beat ALL the best PC programs.
: >>Well let me speculate too!
: >>
: >>In a 10 game match at 40/2:00 Deep Blue will lose at least ONE match
: >>against Genius, Fritz, Rebel or CM5000 ^^^
: >>
: >
: >This is really no bet.... the dubious pruning/extensions that commercial
: >programs do might actually *luck out* once in a while. I have seen
: >CM5K do some real,real long PVs....
: >
: >Why don't you give better odds for the micros? I thought you said that
: >DB really haven't proved that they are better - So methinks that you have
: >some better micro program than DB .... thats fine, just give better odds
: >for the match for the micros (since you think so highly of them).
: >
: >>How about a bet of $1000 Bob?
: >>
: >>40 games will show the (chess) world the state of art of Deep Blue.
: >>Moreover we can draw conclusions about the future of computer chess.
: >>
: >>Speed or knowlegde?
:
: In Diep i bet on knowledge, although we must not forget that most programs
: lack knowledge because the implementing programmer lacks it. On the other
: hand Diep is slow, causing it to search less ply than for example DB, which
: is a tactical drawback; you must always win a game positionally. However,
: Diep's style and games it did win look very human.
:
: >>My opinion, the faster the hardware the more chess knowledge is needed.
: >>I mean if you search 5 plies the 6th ply will bring 50 ELO or more.
: >>But when you already search 16 ply, the 17th ply will bring almost NOTHING!
: >>
: >>- Ed Schroder -
: >
: >Yes you are right, but Bob has already pointed out that the extra speed is
: >being used to do more selective extensions. According to Web Page, DB does
: >12-15 ply searches. This is real bad if it is only doing a "brute force"
: >search to 12-15 ply. The 12-15 ply includes a heck of extensions as well.
: >
: >I wonder what kind of PVS DB would throw out in end games? 30+ line PVS?
:
: 30+ ply is of course an advantage, but we must not forget that it is only
: a tactical advantage. It will not tell DB what move to choose from, if
: it has 2 legal moves, which both don't loose tactically.
:
From where I've sat, it helps way more than you guess. It's lets them play
weak moves and recover, where their search avoids tactical mistakes that take
us many moves to reach.
Robert Hyatt
JackClerk wrote:
: 6 games per year is a small sample size. But from these few games, can anyone
: pinpoint a move that DB made which a PC wouldn't have - and which prevented the
: loss of a game?
:
: I think that most people would be hard pressed to find such a move.
I believe in the DB game against Kasparov, the PC's wanted one rook move
that later proved to be a loser, while DB went for another rook move (either
Rc6 or Rc7, which ever it played.) There were likely many others. If the
PC programs like 90 %, that 10% is more than enough to get 'em creamed.
:
: In article <514vup$b...@juniper.cis.uab.edu>, hy...@crafty.cis.uab.edu says...
: >
: >mark glickman (glic...@fas.harvard.edu) wrote:
: >: In article <511kiq$h82$1...@mhafc.production.compuserve.com>,
: >: Ed Schröder <10065...@CompuServe.COM> wrote:
: >: >In a 10 game match at 40/2:00 Deep Blue will lose at least ONE match
: >: >against Genius, Fritz, Rebel or CM5000 ^^^
: >: >
: >: >How about a bet of $1000 Bob?
: >:
: >
Robert Hyatt
Robert Hyatt
James Garner (da...@laraby.tiac.net) wrote:
: Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
:
: : These "chips" are so ingrained into the program, that removing them would
: : require a software make/unmake/move generator, and would so slow down the
: : program that it wouldn't really be an interesting machine to study. They've
: : invested a great deal of time and energy into developing these chips, which
: : is why I say they are "inseparable".
: So, they have spent much time letting it be fast, and less time
: making it be smart.
:
: Ok, that's fair. There's our answer. It's not as smart, but it's
: much faster, and therefore wins more often than the smarter but much
: slower PC programs. When the day comes that we have 1,000,000 Mhz PC's
: running DOS, then we can play Mchess vs. DB, and Mchess will probably win.
Unlikely, because DB IV will still be 1,000 times faster, most likely, if
the project is still alive. As always. special-purpose hardware is going
to stay way ahead in the speed curve, but does lack general-purpose features
that make the cost difficult to bear. However, the question will be then,
can the PC's of tomorrow beat the world champion. I hope so. If there
are a bunch of machines fighting for the title, I won't mind losing to a
special-purpose chess engine, particularly if I have the best PC based
program. However, I think the "war" will wage for a long time, because
we haven't yet reached that point by a long shot. It's not even clear that
DB belongs on the same board with Kasparov at present, although it is pretty
obvious that none of the other programs are near as good. This seems to be
a "financial" thing rather than a "chess" thing if you ask me. Otherwise,
the match wold be played somewhat differently.
hmmm is about all I can come up with at present, although I think the match
will be entertaining again... from the human perspective...
Bob
Robert Hyatt
: : : slower PC programs. When the day comes that we have 1,000,000 Mhz PC's
: : : running DOS, then we can play Mchess vs. DB, and Mchess will probably win.
: :
: : Unlikely, because DB IV will still be 1,000 times faster, most likely, if
: : the project is still alive.
: But you understand my point. If the issue is which is "smarter",
: more "chess-aware", then the answer is Mchess Pro or one of the other PC
: programs. If the question is which has been designed to integrate well
: and run with blazingly fast chips, then it's DB.
:
: I think there is no more point to arguing, unless the issue now
: is which is better to program for.
I've never disagreed with the above. Just that the DB hardware more than
offsets any currently available knowledge in the commercial programs, and
by a really significant amount. I define "best" as winning the most games,
just like I do in human vs human games. Too many want to change the
definition when the magic word "Deep Blue" is mentioned. Now it's "if they
weren't so fast, they would be no better and likely worse than a micro."
While it's true, it's also "moot"...
Robert Hyatt
: : offsets any currently available knowledge in the commercial programs, and
: : by a really significant amount. I define "best" as winning the most games,
: : just like I do in human vs human games.
: But what is being argued is not "best" as in "winning the most
: games" but "best" as in "the best program". DB is not the best program,
: conclusively. DB is the STRONGEST program, much like a crane is stronger
: than a human being. But they are not in the same league, even if they are
: both trying to do the exact same thing, e.g. lift a 200 pound bucket of
: sand.
:
In that case, let's say they are the "strongest". At least we know how
to test that. However, for the "best" program, how do we decide? That's
an abstract thing without a quantitative way (such as a match) to compare
them...
Valavan Manohararajah
In article <51bcm8$n...@news-central.tiac.net>,
>: : much faster, and therefore wins more often than the smarter but much
>: : slower PC programs. When the day comes that we have 1,000,000 Mhz PC's
>: : running DOS, then we can play Mchess vs. DB, and Mchess will probably win.
>:
>: Unlikely, because DB IV will still be 1,000 times faster, most likely, if
>: the project is still alive.
>
> But you understand my point. If the issue is which is "smarter",
>more "chess-aware", then the answer is Mchess Pro or one of the other PC
>programs. If the question is which has been designed to integrate well
>and run with blazingly fast chips, then it's DB.
>
> I think there is no more point to arguing, unless the issue now
>is which is better to program for.
I really don't know if you can say that DB is not as "chess-aware" as the PCs.
Isn't Murray Campbell a strong player (expert perhaps?). Besides they have
had a few GMs help them...I believe Joel Benjamin works with them.
It is also becoming real, real hard to say what exactly is meant by
"chess-awareness". Sure eval functions are one thing, but somethings can be
discovered by search and thus don't need to be coded anyway. Look at Fritz..
It runs as fast as hell and trys to go real deep... I doubt it has very much
knowledge in its eval functions... but the deep search is in place of the
reduced knowledge. Fritz is also one of the strongest programs around...so
"chess-awareness" doesn't really mean much when singled out. What you should
really look at is the combination of search and eval functions.
So comparing chess programs/systems eval functions or knowledge content
doesn't really make any sense.
--
-------------------------------------------------------------------------------
man...@ecf.utoronto.ca | 3rd year Comp Eng., University of Toronto
Valavan Manohararajah |
-------------------------------------------------------------------------------
Chris Whittington
Here we go again .....
Fatal flaw in reasoning: 'all knowledge can arise from search'.
No it can't. Not in the available space.
An example:
You know that white pawn on a2, king on b3, bishop on b8 (eg wrong
colour bishop), black king c5 is a draw.
You can code this knowledge into an evaluation function, so the
program knows its a draw at ply 1.
If you don't code this knowledge, it will take a program 6 x 50
move rule = 300 moves to conclude its drawn. The program will
shuffle the bishop all over the board, and delay pawn pushing
until the 50-move rule demands it.
So, this is an extreme case, knowledge = 300 plies.
Other knowledge is worth 1 ply, or 2 plies or 20 plies or whatever.
In some situations (often in my view) you can watch programs
play each other and they both have no idea about what is going on.
Their search doesn't discover the essential elements of the position,
and often, these elements are obvious to a human.
Search can find *combinational* knowledge, eg. a tactical resolution;
but it can't find *positional* knowledge unless the positional
is resolvable within the search space.
Think of DB game 6. DB didn't understand the game. It did not have
the necesary knowledge coded in its evaluation, and its search was
unable to discover the problem (until too late).
Chris Whittington
Robert Hyatt
Chris Whittington (chr...@cpsoft.demon.co.uk) wrote:
I didn't read "all knowledge comes from search" in the above. What I read
was that some knowledge, amplified by deeper search, can produce results
that "some knowledge" by itself can't, and that, within reason, "some
knowledge" continues to improve proportionally with depth, perhaps in a
diminishing way. However, his point about Fritz is quite accurate. It does
play well, and is right near the top of the SSDF (within 50 points or so I
believe) which means a tiny eval, augmented by a deep search, can produce
good results. The question is, what happens as they add knowledge? Does
it get better? Or worse? because it will obviously lose depth...
Mark Rawlings
Chris Whittington <chr...@cpsoft.demon.co.uk> wrote:
<snip>
>Think of DB game 6. DB didn't understand the game. It did not have
>the necesary knowledge coded in its evaluation, and its search was
>unable to discover the problem (until too late).
>Chris Whittington
How do the top PC programs do when evaluating game 6? Do they
understand the game better? I'll try it with Genius 4. ( Although I'm
not sure *I'll* be able to understand whether Genius understands...)
Mark
mcl...@prima.ruhr.de
In article <84269879...@cpsoft.demon.co.uk>, Chris Whittington
<chr...@cpsoft.demon.co.uk> writes:
>Think of DB game 6. DB didn't understand the game. It did not have
>the necesary knowledge coded in its evaluation, and its search was
>unable to discover the problem (until too late).
>
>Chris Whittington
The following is ironicle:
Maybe we should really switch into another newsgroup!
In my past I read Pirsig: Zen and the Art of Motorcycle Maintenance
and Pirsigs: Lila, an Inquiry into Morals.
I always thought that Americans could be proud to have persons
like Pirsig and that they CAN UNDERSTAND the difference
between quality and quantity.
Now it looks for me, that philosophy, moral and values are not much
known in US.
------------------------
The following stuff is purely redundant:
Chris, they don't want to understand. They really think that
search-depth into a range < infinite is the same like knowledge.
They have not understood the story of adam and eve and the tree of
knowledge.
They know that chess can never be outcomputed, but they still
believe in their machines.
I will never buy an american car! It's produced by people who believe
that machines have spirit or will have spirit if they search deep
enough or eat enough fuel or have enough PS like a horse. :-)
We should really give it up ! Also we should give up to
convince them by showing them games.
They don't replay them! How can you convince a chess player about
chess games, if they don't replay them or comment on the games.
They've asked for games. We show them games. Now they do not say
anything.
I give up.
I do not believe that stupid ideas or programs will lead the world.
I don't want to. And I don't see them lead the swedish-rating-list.
When was the last time, a fast/stupid program was the leader of the
swedish-rating-list, can the swedish-guys find out when this was:
before Jesus Christ or after him ?
You are right.
Chris and me are dreamers. We have an ideal, and we believe in it.
And we do not call us realists, but dreamers.
You call yourself realists. But you are dreamers too.
Marty, Ed, Richard, Mark are dreamers too ! When will they give
up their ideas in knowledge and develop into materialists ?
Construct the world as materialistic as you want.
Make communism be the bad klingons.
A nation that assassinated most of their presidents is a strange
nation. A nation that elects actors as presidents is a little
strange.
How can Kirk beat Spock in chess if Spock plays more logic ?
How can Deanna Troi beats Data in chess, if Data is a computer.
How can Lasker beat Capablanca, the "chess-machine".
When will Fischer beat Kasparov (in normal or fischer-randomn-chess)?
I want to excuse for my above words, but they express my feelings.
Robert Hyatt
mcl...@prima.ruhr.de wrote:
: In article <84269879...@cpsoft.demon.co.uk>, Chris Whittington
However we immoral Americans can certainly recognize ignorance...
And we can recognize what belongs in this newsgroup (computer chess)
and what doesn't (political ranting...)
:
: ------------------------
:
: The following stuff is purely redundant:
:
: Chris, they don't want to understand. They really think that
: search-depth into a range < infinite is the same like knowledge.
Don't put words in my mouth. However, it's provable that search and
knowledge work together, and that an increase in one can offset a
decrease in the other. Check the literature... If you bother to read
any...
:
: They have not understood the story of adam and eve and the tree of
: knowledge.
:
: They know that chess can never be outcomputed, but they still
: believe in their machines.
It seems to be working so far. At least DB is the strongest machine
in existance.
:
: I will never buy an american car! It's produced by people who believe
: that machines have spirit or will have spirit if they search deep
: enough or eat enough fuel or have enough PS like a horse. :-)
Hmmm... tell that to Ferarri... They seem to be pretty fanatical about
building "machines" that go fast, even if they aren't particularly quiet,
or fuel-efficient, or comfortable, or practical, or anything but *fast*.
Last time I was in Europe, there were plenty of 'em running on the roads
there...
:
: We should really give it up ! Also we should give up to
: convince them by showing them games.
:
: They don't replay them! How can you convince a chess player about
: chess games, if they don't replay them or comment on the games.
:
: They've asked for games. We show them games. Now they do not say
: anything.
What on earth are you ranting about here? What games? I've said it
once, and I'll say it again. The "best" program is the one that wins
the *most* games. You keep wanting to change the definition until you
find one that no one comments on. However, best is still "the one
that wins the most games" in the world of chess. Otherwise we'll be
forever arguing whether Tal was better than <your choice> here, because
even if Tal beat <your choice> he played wild, tactical chess that was
not as "good" as <your choice>... OTB is a common measure of "best"...
Works for me.
If you are talking about the 3 games you posted, I looked at 'em, quickly
decided they were worthless because of the openings played, and went on with
business as usual. Chris and I will have a chance to see how CSTal does
against a "real" crafty soon. Until then, the games you posted are not
interesting. No information about hash table size, about how you set the
time control, using a book that is next to useless (as is mentioned in the
read.me, BTW, because it has no games, just singular lines from lots of
opening texts with no win/lose/draw results to help crafty pick lines, etc.)
If you are ranting about that, keep right on ranting. It's not likely I'd
pay much attention to games posted by a "disinterested party" anyway. If
you "don't think that CSTal has too many problems with this version of
Crafty", that's your right to think what you want, Whether or not that's
reality is another thing altogether. Crafty's pretty well proven itself to
be a reasonably strong program. Ask Lonnie Cook. He's beat on it with
Genius, Fritz, Wchess and now CM5000. Losing 3 in a row to any of those
is not something I'd bet on, although statistically I see it happen. I
can recall a 4 game match with Ferret where crafty won all 4 and Bruce
was very "interested." The next match Ferret won all 4, so your 3 games
don't mean much in the great scheme of things, particularly when you don't
explain the circumstances of how you set things up. Send me log files
from Crafty and I can tell exactly what went on. With just moves, I can't
tell anything.
Of course, this has *nothing* to do with Deep Blue...
:
: I give up.
: I do not believe that stupid ideas or programs will lead the world.
: I don't want to. And I don't see them lead the swedish-rating-list.
: When was the last time, a fast/stupid program was the leader of the
: swedish-rating-list, can the swedish-guys find out when this was:
: before Jesus Christ or after him ?
Hmmm... you must really make your parents proud...
: You are right.
: Chris and me are dreamers. We have an ideal, and we believe in it.
: And we do not call us realists, but dreamers.
: You call yourself realists. But you are dreamers too.
: Marty, Ed, Richard, Mark are dreamers too ! When will they give
: up their ideas in knowledge and develop into materialists ?
:
: Construct the world as materialistic as you want.
: Make communism be the bad klingons.
: A nation that assassinated most of their presidents is a strange
: nation. A nation that elects actors as presidents is a little
: strange.
As opposed to a nation that murders or shackles its citizens, eh?
Or one that tried to conquer Europe eh? Stupid comments to post in
this newsgroup. Again, no facts, just rantings (or droppings, depending
on your perspective...)
:
: How can Kirk beat Spock in chess if Spock plays more logic ?
Chris Whittington
raw...@erols.com (Mark Rawlings) wrote:
>
> Chris Whittington <chr...@cpsoft.demon.co.uk> wrote:
>
>
> <snip>
>
> >the necesary knowledge coded in its evaluation, and its search was
> >unable to discover the problem (until too late).
>
> >Chris Whittington
>
> understand the game better? I'll try it with Genius 4. ( Although I'm
> not sure *I'll* be able to understand whether Genius understands...)
>
> Mark
>
I doubt their understanding is any better.
Chris Whittington
Mark Rawlings
Chris Whittington <chr...@cpsoft.demon.co.uk> wrote:
>Chris Whittington
I took a look at the Rebel 7 analysis of game 6 (from Ed's web page)
and from the eval scores it looks like the program doesn't understand
that it is getting crushed by Kasparov. I'm sure you are right Chris.
Mark
Chris Whittington
kirk...@ee.utah.edu (Dan Kirkland) wrote:
>
> In article <84269879...@cpsoft.demon.co.uk>,
> Chris Whittington <chr...@cpsoft.demon.co.uk> writes:
>
>
> >Here we go again .....
>
> where you stand. Enough already!
>
> >Fatal flaw in reasoning: 'all knowledge can arise from search'.
>
> This is only your skewed view of what others think.
>
> >No it can't. Not in the available space.
>
> even thinks this is true.
>
> >An example:
> >
> >You know that white pawn on a2, king on b3, bishop on b8 (eg wrong
> >colour bishop), black king c5 is a draw.
> >
> >You can code this knowledge into an evaluation function, so the
> >program knows its a draw at ply 1.
> >
> >If you don't code this knowledge, it will take a program 6 x 50
> >move rule = 300 moves to conclude its drawn. The program will
> >shuffle the bishop all over the board, and delay pawn pushing
> >until the 50-move rule demands it.
>
> Well, a draw is a draw...
>
> Yes, I know you could come up with a better example,
> just as most of the rest of us can. But why bother!
> We have ALL seen it before.
>
> >So, this is an extreme case, knowledge = 300 plies.
> >Other knowledge is worth 1 ply, or 2 plies or 20 plies or whatever.
> >
> >In some situations (often in my view) you can watch programs
> >play each other and they both have no idea about what is going on.
> >
> >Their search doesn't discover the essential elements of the position,
> >and often, these elements are obvious to a human.
> >
> >Search can find *combinational* knowledge, eg. a tactical resolution;
> >but it can't find *positional* knowledge unless the positional
> >is resolvable within the search space.
> >
> >the necesary knowledge coded in its evaluation, and its search was
> >unable to discover the problem (until too late).
>
> ever written will make mistakes now and then. Something you
> don't seem to understand is that with computer vs computer,
> if one searches very much deaper than the other (either full-
> width OR selective), the deeper searching one will win most
> of the time. But then, if your program is only searching
> 4,000 nps, then I'm sure you will find this out...
>
> >Chris Whittington
>
> blah blah blah
>
> First, quite trying to imply that others think that the
> search is all knowing, or that speed is all important.
> Nobody thinks either. (Although I'm sure you will try
> and twist something I said in a feeble attemp to make
> it sound like I said such. Silly!)
>
> Second, you don't need to tell us what you think some
> 5-10 times a day. I think we know what you think. If
> not, then we never will. Either way, there is no need
> to do it over and over and over and ...
>
> Third, quit attacking search speed...
> Speed is one of the main advantages of a computer.
> Not using this speed is senseless. A computer is NOT
> human. A computer can NOT play chess like a human.
> A computer will NEVER play chess like a human. If you
> think otherwise, it is your loss...
>
> Because human chess players do something a certain way
> does NOT mean it is the best way for computers. Most
> of us realize that the best computer chess will be the
> RIGHT combination of speed and knowlege (whatever that
> combination actually is...???).
>
> While you may want to slow it down and make it play like
> a human, most of us realize this is pointless. But hey,
> while you are at it, here are some things you need to
> look at...
>
> Most top level players spend more time before a match or
> game preparing than they do playing the match or game.
>
> Most top level players know pretty much what the other
> player will play in many situations (sometimes for very
> many moves).
>
> Most top level players try to get the other player in
> unfamiliar positions.
>
> (Better add this to your turblence stuff...)
>
> Finally, IF you EVER get your program so that it beats
> all the rest (ya right!), I will take the time to sit
> down and actually write (or help write) a chess program
> that will blow yours away! (Please save this for future
> reference...)
>
>
> That was fun! :)
> Now I just can't wait for another day on RGCC with
> Chris and Don repeating and re-repeating and
> re-re-repeating and ... the same ol stuff.
>
> dan
Well, I thought I'ld take a look at all your posts that are
coming up on my browser.
I found six. Three were direct replies to me, and the other three
were one or two removed, but following one of my posts.
Obviously you like reading them. Obviously they annoy you.
Normally your replies are agressive, rude, macho-challenging
which is why I ignore them. Probably this annoys you more.
If you really can't stop reading my posts (they're marked at the
top 'Chris Whittington' - quite easy to avoid), and then
getting angry; I'll be forced to assume that you love me really.
In so far as you want answers to your questions (in contradiction
to your suggestions that I shut-up), you'll find them all
in various postings dotted around.
Please don't think I want a flame war with you. I know you're
trying to provoke me, but I have other, more rewarding areas to
reserve myself for.
Happy hunting.
Chris Whittington
Dan Kirkland
Jouni Uski
One example: when DB give up in gabe 6, Fritz 4 evaluates position still about equal!
Jouni
Robert Hyatt
Jouni Uski (Jouni...@nce.nokia.com) wrote:
: One example: when DB give up in gabe 6, Fritz 4 evaluates position still about equal!
:
: Jouni
Right. I forgot what Murray said DB was seeing, but it was at least -9 and
might have been a mate. In any case, it *knew* it was lost. Crafty thought
black was terrible for the whole game, but it was nowhere near bad enough to
resign, it just didn't like the position with all the white pawns, mobility,
and so forth, while black was squeezed into a couple of ranks basically.
DB's impressive... even if it plays bad chess at times... :) It discovers
just how bad it is way before anybody else. Now, if they quit playing bad
positional chess, *ouch*...