Deep Blue vs Micros
Robert Hyatt
Since we can't get past opinion and prejudice here very easily, I
thought I would write a simple to follow "paper" something like you
might produce for a homework assignment in a basic statistical
class.
Hypothesis: Deep Thought/Deep Blue is at least 200 Elo rating
points better than any microcomputer playing. It has also been
at least 200 Elo rating points better than any microcomputer at
any point in time.
Methodology to prove this: A 200 point rating difference should
produce 3 wins out of every 4 games if memory serves me correctly.
If this is wrong, I'm sure it will be corrected, but I don't think
that it's going to affect the outcome, because I believe that winning
3 out of 4 games is a lower bound for DT/DB.
To come up with the wins and losses, I'm going back through my old
ACM and WMCCC event file (paper) and am going to extract, year by
year, results of DT against the best micros of that year. I'll
cull wins over unknown/weak opponents to avoid biasing the results,
but I'll include Cray Blitz, HiTech, *Socrates and other programs
that are non-micro-based, but which are also provable as being better
than micros of the era by this same methodology.
1987, ACM18, Dallas Texas. Chiptest (predecessor to Deep Thought and
Deep Blue) won this event with 4 wins and 0 losses. Chiptest beat Lachex,
a strong program running on a Cray, Sun Phoenix, a strong program
running on a cluster of Sun's, and Cray Blitz, whom we already know
about. One win over Levy's program I'm not counting, although Cyrus
finished in 5th place. Novag finished in 7th, Belle in 8th, Waycool
(parallel hypercube program) finished in 9th, gnuchess in 10th, etc.)
3 wins 0 losses, against strong competition.
1988, ACM19, In this event, Deep Thought finished in first place.
It drew with Fidelity, beat Waycool, beat HiTech and beat Mephisto.
3 wins, 0 losses, 1 draw.
1989, ACM20, In this event, HiTech finished in first place, Deep
Thought in second. Deep thought beat phoenix, BP, HiTech, Rebel,
but lost to Mephisto in round 5.
4 wins 1 loss
1990, ACM21, in this event, Mephisto won, with Deep Thought (first
version) in second place. DT beat BeBe, a strong special purpose
machine, beat Belle, Beat Mephisto, lost to Hitech, and beat
Zarkov.
4 wins 1 loss.
1991, ACM22, In this event, Deep Thought once again won. It beat
Zarkov, MchessPro, HiTech, Cray Blitz, and ChessMachine Shroeder.
5 wins, 0 losses.
1993, ACM23, In this event, DT did not participate. It was won by
*Socrates, with Cray Blitz in 2nd, HiTech in 3rd, and *Tech in 4th
(Hitech on a CM5 machine). Zarkov was 5th, Chess Machine Shroeder
was 7th, Mchess Pro was 10, with 12 participants total. No results
here, other than to notice that four "supercomputer" programs
finished in the top 4 places, with the micros left in the remaining
8 places.
1994, ACM24, In this event, DT finished in first, despite a first
round forfeit on time. A severe thunderstorm knocked out power to
the IBM Watson center, and it was not restored in a reasonable
time. Since the first day had two rounds scheduled, delaying beyond
a point was impractical, so TD Mike Valvo forfeited DT in round 1
against (I believe) Mchess Pro. I won't count this because it was
not an OTB loss, and was clearly beyond their control. However,
the interesting thing is they won the tournament convincingly. By
winning the next 4 rounds easily. I can't find my results for this
event, and will simply record this as 4 wins and no losses. I do
recall that Wchess was there, Mchess Pro, Cray Blitz, HiTech,
*Socrates, and others, so the field was very strong as usual.
4 wins, 0 losses.
Now for the grand total:
23 wins 2 losses 2 draws.
that's 23 wins out of 27 games, which I claim is greater than
a 75% win ratio (it's actually 85% for those interested.) The
*only* conclusion one can draw from this is that DT was at
*least* 200 rating points better than the competition in those
years, as proved by Elo's calculations. That held through 1994
with no variance, because you will notice their worst finish ever
in a tournament was only one loss, while it was typically 4-0 or
5-0 and out for them...
Summary: After reviewing the above results, the conclusion is
quite easy to reach, because Elo's formula precisely defines a
rating advantage for a specific win/loss ratio. I don't have
the Hong Kong data to see who they played, but we know of course
that they won 3, lost 1 (Fritz) and drew once (I believe) for
3.5 out of 5. If you factor that in, it does not change a
thing, and simply shows that they have dominated like no other
program since they started playing.
Now, to 1997. The fastest version I recall of this machine was
searching around 10M nodes per second in 1994, if memory serves
me correctly... They had (again, from memory) 12 processors,
each searching around 700K positions per second (Hsu, if this
is incorrect, refresh my memory, I'm afraid my DRAM refresh
circuitry is slowing down and I occasionally turn 1's into
0's now and then I'm afraid, although I'm having this worked
on.. :) )
Let's give 'em 10M in 1994, and according to current reports,
rumors, and speculation, as well as some hints from the IBM
web page, let's give 'em 1000M now, which seems reasonable,
and is consistent with their goal from day one at IBM. Since
1994 then, 100x faster, with also some hardware design improvements
that help the eval, help with determining checkmate/draws deep in the
tree, and so forth.
In 1994, what was the fastest machine we had, PC-wise? I'll go
out on a limb and say a P5/90 sometime that year, although I'm
not sure. There were a gaggle of 486/100's around then, so
maybe we should use the 486/100 as the PC speed in 1994. On
to 1997 and the p6/200, which is 2.5x faster than a P5/133, which is
about 2x faster than a P4/100 (roughly, someone might have a better
number to plug in here). So, here in 1997, the Micros are 5x faster
than 3 years ago, DB is 100X faster than 3 years ago, yet it was
crushing everyone back then.
To a reasoned and scientific mind, what is *your* conclusion, based
on *facts*? (note that none of the above is idle speculation or
rumor, it's all available as factual data. The source of the tournament
results is my huge file on chess events, but all of that was published
in the ICCA and other places yearly as well.
*my* conclusion is that in 1994 DT was *at least* 200 rating points
better than the next closest program, which was *not* a PC based
program at all, but was a member of the set {Cray Blitz, HiTech,
*Socrates, LaChex, Phoenix, to name a few}. This is based on the
finishing order at these events, where the above programs were
always at the top, with the micros falling somewhere below them.
*my* conjecture (because there are no games to support this, only
the raw performance numbers above comparing DT/94 and DB/97 to the
PC/94 and PC/97 platform) is that the distance has *not* closed, it
has widened.
Now for the interesting discussion. Feel free to challenge anything
above you like, regarding my methodology. Someone might also cross-check
my reported results year-by-year, because some of this was stuff handed
out at the tournament, hot off the Xerox, and could well have had errors,
although I believe they jive with my memory perfectly.
the above "conjecture" is just that, but given the evidence that their
speed made them far superior to everyone else in 1994 and prior years,
and that their speed advantage has increased by 20X since then (PC's
gained 5X, they got 100X), and since there's been no remarkable break-
through in the microcomputer chess programs to close this gap, then
my conjecture is nearly a "proof".
Comments???
Dan Schmidt
hy...@crafty.cis.uab.edu (Robert Hyatt) writes:
| Since we can't get past opinion and prejudice here very easily, I
| thought I would write a simple to follow "paper" something like you
| might produce for a homework assignment in a basic statistical
| class.
|
| Hypothesis: Deep Thought/Deep Blue is at least 200 Elo rating
| points better than any microcomputer playing. It has also been
| at least 200 Elo rating points better than any microcomputer at
| any point in time.
|
| [...]
|
| Now for the grand total:
|
| 23 wins 2 losses 2 draws.
Or with the other +3-1=1 performance added in, +26-3=3, which is an
86% score.
| Now for the interesting discussion. Feel free to challenge anything
| above you like, regarding my methodology.
Well, my only problem is that you have shown that Deep Blue is
approximately X points better than its average micro competitor, not
that it's X points better than the next best micro, which was your
assertion.
It would be nice to know what value of X corresponds to an 86%
score. I can't find the Elo formula online, though.
Robert Hyatt
Dan Schmidt (df...@harmonixmusic.com) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) writes:
: | Since we can't get past opinion and prejudice here very easily, I
: | thought I would write a simple to follow "paper" something like you
: | might produce for a homework assignment in a basic statistical
: | class.
: |
: | Hypothesis: Deep Thought/Deep Blue is at least 200 Elo rating
: | points better than any microcomputer playing. It has also been
: | at least 200 Elo rating points better than any microcomputer at
: | any point in time.
: |
: | [...]
: |
: | Now for the grand total:
: |
: | 23 wins 2 losses 2 draws.
: Or with the other +3-1=1 performance added in, +26-3=3, which is an
: 86% score.
: | Now for the interesting discussion. Feel free to challenge anything
: | above you like, regarding my methodology.
: Well, my only problem is that you have shown that Deep Blue is
: approximately X points better than its average micro competitor, not
: that it's X points better than the next best micro, which was your
: assertion.
Note that the above is not the results I gave. I didn't include any of
the "weak" programs (there were a few wins over those that I left out)
although DT didn't play many weak ones since it was always seeded #1.
However, we do need to add a 5-0 for the 1989 WCCC in
Alberta, which they swept cleanly, beating Rebel, Fidelity, Cray
Blitz and HiTech in the process, *not* a favorable set of pairings...
So we have +31 -3 =3 which is even better. (or worse, depending on
your perspective... :) )
In fact, that is *almost* 9 wins for every loss, ignoring the draws, which
is *pretty close* to what I've said all along... about how they would do
against Micros. Note that at least 3 of those wins are against Cray Blitz,
which is *not* a micro by any stretch of the imagination... it's better...
: It would be nice to know what value of X corresponds to an 86%
brucemo
Dan Schmidt wrote:
> It would be nice to know what value of X corresponds to an 86%
> score. I can't find the Elo formula online, though.
86% is 315 Elo points.
elo = 400 * LOG((1/S)-1)
Where S is the score, like 0.86 in this case, and LOG is log base
10.
S = 1/(1+(10^(elo/400)))
is how you flip the formula around, in case you want to do this,
and would like to avoid having to do the work yourself.
"^" is "raised to the power of" in this case (not "XOR" :-)).
bruce
Martin Borriss
In article <330A92...@nwlink.com>,
brucemo <bru...@nwlink.com> writes:
>Dan Schmidt wrote:
>
>> It would be nice to know what value of X corresponds to an 86%
>> score. I can't find the Elo formula online, though.
>
>86% is 315 Elo points.
>
>elo = 400 * LOG((1/S)-1)
>
Is this USCF rating?
FIDE Elo would be (86-50) * 7 which it makes it 252 points difference.
Martin
A.Mader
Due to the formula from Prof. A. Elo 86% is a difference of 315 points.
It doesn't matter if its USCF or FIDE. ELO ratings are not computed
using a liner formula (like you do).
Best wishes
Andreas Mader
A.Mader
A.Mader wrote:
> Due to the formula from Prof. A. Elo 86% is a difference of 315 points.
> It doesn't matter if its USCF or FIDE. ELO ratings are not computed
> using a liner formula (like you do).
>
> Best wishes
> Andreas Mader
Of course this should be "LINEAR" formula.....
Sorry for that
Andreas Mader
Martin Borriss
In article <330B64...@siemens.at>,
"A.Mader" <andrea...@siemens.at> writes:
>Martin Borriss wrote:
>>
>> In article <330A92...@nwlink.com>,
>> brucemo <bru...@nwlink.com> writes:
>> >Dan Schmidt wrote:
>> >
>> >> It would be nice to know what value of X corresponds to an 86%
>> >> score. I can't find the Elo formula online, though.
>> >
>> >86% is 315 Elo points.
>> >
>> >elo = 400 * LOG((1/S)-1)
>> >
>>
>> Is this USCF rating?
>>
>> FIDE Elo would be (86-50) * 7 which it makes it 252 points difference.
>>
>
>It doesn't matter if its USCF or FIDE. ELO ratings are not computed
>using a liner formula (like you do).
>
To the best of my knowledge they are computed the way I said. It used to be
non-linear some years ago, now it is linear.
I guessed at USCF because of this "400" constant in the formula. In FIDE Elo
I have to score 100% against players at least 350 points weaker. In
practice, the formula I used always gives me the correct (same as FIDE)
result.
No big deal, these big difference are always fuzzy.
Martin
Joe McCaughan
Hmmmm....
Things seem to be getting a little quiet out there...
Ed? Anything to say?
...
Ed Schroder
From: shi...@nntp.best.com (Joe McCaughan)
: Hmmmm....
: Things seem to be getting a little quiet out there...
: Ed? Anything to say?
Hi,
In principal I have said everything what is on my mind and I don't like
repetitions, they as you know only end in a draw... :)
But what I really have missed is that it was a discussion sofar between
me and Bob only. Where is the rest of you?
Bob and I are just 2 persons with only 2 brains and 2 opinions and
thoughts. There are hundreds of people here all with opinions about
this subject.
So write them, I like to hear them...
I like to know for example your opinions about breakthroughs. I am
quite happy with a 25 ELO gain in one year on the same hardware and I
call 50 ELO due to software improvements only simply a breakthrough.
Going from 1600 to 1650 is nothing compared to 2410 to 2460.
Going from 1600 to 1650 is no breakthrough.
Going from 2410 to 2460 is a breakthrough.
Just ask any IM, if he (she) makes such an ELO jump the IM in question
is the happiest man (woman) in the world!
I have read Bob's comments on that and besides the fact I find his
comments very unfounded I like to hear yours.
Are Bob and I the only ones here in RGCC?
I hope not...
- Ed Schroder -
Ed Schroder
From: shi...@nntp.best.com (Joe McCaughan)
: Hmmmm....
: Things seem to be getting a little quiet out there...
: Ed? Anything to say?
Hi,
In principal I have said everything what is on my mind and I don't like
repetitions, they as you know only end in a draw... :)
But what I really have missed is that sofar it was a discussion between
Bob and me only. Where is the rest of you?
Ed Schroder
From: shi...@nntp.best.com (Joe McCaughan)
: Hmmmm....
: Things seem to be getting a little quiet out there...
: Ed? Anything to say?
Hi,
In principal I have said everything what is on my mind and I don't like
repetitions, they as you know only end in a draw... :)
But what I really have missed is that it was a discussion sofar between
me and Bob only. Where is the rest of you?
Torstein Hall
I find this rating talk to be a little funny. If you put Rebel up against a
IM with 2465 in rating (Rebel 8 rating on SSDF) I would be willing to bet
quite a lot, that it would lose this match with time controls set 2h 40
moves.
This puts the talk about rating improvments into its right position. It
only counts for play agains computers as its only measured agains
computers.
Would you put Rebel 8 up against an IM in such a match and calculate its
rating next. I guess it would end up around in the area of 2200-2300.
Anyway it would have been very interesting to see such a match.
--
Torstein Hall
tors...@eunet.no
http://login.eunet.no/torshall/sjakk.html
Ed Schroder <rebc...@xs4all.nl> skrev i artikkelen
<5emti5$2hc$1...@news1.xs4all.nl>...
> From: shi...@nntp.best.com (Joe McCaughan)
>
> : Hmmmm....
>
> : Things seem to be getting a little quiet out there...
>
> : Ed? Anything to say?
>
>
> Hi,
>
> In principal I have said everything what is on my mind and I don't like
> repetitions, they as you know only end in a draw... :)
>
> But what I really have missed is that sofar it was a discussion between
> Bob and me only. Where is the rest of you?
Ed Schroder
From: "Torstein Hall" <tors...@eunet.no>
: I find this rating talk to be a little funny. If you put Rebel up
: against a IM with 2465 in rating (Rebel 8 rating on SSDF) I would be
: willing to bet quite a lot, that it would lose this match with time
: controls set 2h 40 moves.
: This puts the talk about rating improvments into its right position.
: It only counts for play agains computers as its only measured agains
: computers.
: Would you put Rebel 8 up against an IM in such a match and calculate
: its rating next. I guess it would end up around in the area of
: 2200-2300. Anyway it would have been very interesting to see such a
: match.
Perhaps you underestimate the micro's.
Look at AEGON (man versus machine) Rebel and other micro's played very
well against strong IM's and GM's.
ELO TPR for Rebel the last 3 years...
AEGON 1996 : 2530
AEGON 1995 : 2460
AEGON 1994 : 2470
That's not in the 2200-2300 range but clearly above.
It's also not a lucky shot since it's from the last 3 years and quite
steady... The same counts for other micro's especially Fritz (1996)
and Hiarcs (1995).
AEGON 1996 (6 rounds)
Rebel - GM Christianssen 1-0
Rebel - GM Vaganian 1/2 - 1/2
Rebel - IM Hoeksema 1/2 - 1/2
Playing level 40 in 1:30:00
After that 15:00 for the rest of the game
Bronstein method: add 0:20 for every move
This comes pretty close to 40 in 2:00
Average is about 2:15 per move.
- Ed Schroder -
: Torstein Hall
: tors...@eunet.no
: http://login.eunet.no/torshall/sjakk.html
Robert Hyatt
Ed Schroder (rebc...@xs4all.nl) wrote:
: From: shi...@nntp.best.com (Joe McCaughan)
: : Hmmmm....
: : Things seem to be getting a little quiet out there...
: : Ed? Anything to say?
: Hi,
: In principal I have said everything what is on my mind and I don't like
: repetitions, they as you know only end in a draw... :)
: But what I really have missed is that sofar it was a discussion between
: Bob and me only. Where is the rest of you?
: Bob and I are just 2 persons with only 2 brains and 2 opinions and
: thoughts. There are hundreds of people here all with opinions about
: this subject.
: So write them, I like to hear them...
: I like to know for example your opinions about breakthroughs. I am
: quite happy with a 25 ELO gain in one year on the same hardware and I
: call 50 ELO due to software improvements only simply a breakthrough.
: Going from 1600 to 1650 is nothing compared to 2410 to 2460.
There you are wrong. The Elo formula computes a spread of 50 between two
1500 players exactly the same as a spread of 50 between to 2600 players. It
simply gives the expected # of wins for each in a match.
: Going from 1600 to 1650 is no breakthrough.
: Going from 2410 to 2460 is a breakthrough.
If you were on the very upper end of the *real* Elo scale, like, say a Kasparov,
you are correct, because a 50 point gain is difficult. But only because you don't
get anything for beating players 400 points below you. And if you are on the top,
you are only playing players below you. And if you are Kasparov, most are 2-300
points below you. Even one loss costs you 32 points. it might take 32 wins to
pick that back up and break even.
*But*, on the SSDF, this isn't true. Look how the programs are bunched up on
the top, with over a dozen within 100 points of each other. No one is way out
front, and, as a result, they don't suffer from the "compression" problem as
above.
I see this all the time on ICC. Crafty stays in the top 10 or so at blitz all
the time. It's had a peak ovf 2850, but it's nearly impossible to stay there,
because it is playing opponents rated at 2200 also. And one loss = -32 points,
while one win = 0. Against a real master, it would have to win *every* game
to stick up at that lofty pinnacle. It won't happen. On occasion, a GM will
go crazy and play it 100 games, and lose 3/4 of them, and drive it up to such
a rating, but when he goes away, the normal rating pool simply drags it right
back down because there's not many rated over 2600, that play actively on ICC.
There's far move 2300-2400 IM's, but if a 2800 player beats a 2300 player he
gets nothing. He loses a lot when he draws or loses however.
The bottom line is as you distance yourself from the pack, rating changes are
much more hard to produce as you suggest. But on the SSDF, no one is distanced
+200 from the rest. or even +100 from the rest. The top of the rating pool is
quite compressed, making the "spread" calculation very accurate, but also +50
at the top is just as significant and difficult to obtain as +50 in the middle.
: Just ask any IM, if he (she) makes such an ELO jump the IM in question
: is the happiest man (woman) in the world!
But not particularly more happy than a regular master that jumps +50, or
an expert that jumps +50... I agree that +50 represents some work on the
IM's part, to be sure. But notice also that he's playing in a big pool of
players with reliable ratings based on the Elo formula. The SSDF rating pool
is much smaller, with rating that are distorted by incest. The difference
between two programs gives an indication of the expected win/loss ratio, but
the actual rating doesn't mean a lot, other than when compared to another program
for the spread. Elo's main "claim to fame" is the "spread" between two players,
not the "rating" of a single player. You could add +300 to everyone's rating,
and the rating pool would still be stable at that new rating level for everyone.
: I have read Bob's comments on that and besides the fact I find his
Kim Hvarre
Torstein Hall <tors...@eunet.no> wrote in article
<01bc20cd$2e5c0ba0$57fb47c1@Ptorshall>...
> I find this rating talk to be a little funny. If you put Rebel up against
a
> IM with 2465 in rating (Rebel 8 rating on SSDF) I would be willing to bet
> quite a lot, that it would lose this match with time controls set 2h 40
> moves.
>
> This puts the talk about rating improvments into its right position. It
> only counts for play agains computers as its only measured agains
> computers.
>
> Would you put Rebel 8 up against an IM in such a match and calculate its
> rating next. I guess it would end up around in the area of 2200-2300.
> Anyway it would have been very interesting to see such a match.
>
>
> >
> > : Hmmmm....
> >
> > : Things seem to be getting a little quiet out there...
> >
> > : Ed? Anything to say?
> >
> >
> > Hi,
> >
> > In principal I have said everything what is on my mind and I don't like
> > repetitions, they as you know only end in a draw... :)
> >
> > But what I really have missed is that sofar it was a discussion between
> > Bob and me only. Where is the rest of you?
> >
> > Bob and I are just 2 persons with only 2 brains and 2 opinions and
> > thoughts. There are hundreds of people here all with opinions about
> > this subject.
> >
> > So write them, I like to hear them...
> >
> > I like to know for example your opinions about breakthroughs. I am
> > quite happy with a 25 ELO gain in one year on the same hardware and I
> > call 50 ELO due to software improvements only simply a breakthrough.
> >
> > Going from 1600 to 1650 is nothing compared to 2410 to 2460.
> > Going from 1600 to 1650 is no breakthrough.
> > Going from 2410 to 2460 is a breakthrough.
> > Just ask any IM, if he (she) makes such an ELO jump the IM in question
> > is the happiest man (woman) in the world!
> >
> > I have read Bob's comments on that and besides the fact I find his
> > comments very unfounded I like to hear yours.
> >
> > Are Bob and I the only ones here in RGCC?
> >
> > I hope not...
> >
> > - Ed Schroder -
> >
> >
> >
>
Funny thing to observe - a year or 2 ago I had a debate with Hyatt in the
same neighborhood, i.e. whether the ratingpool of SSDF in any way was
representative in terms of relative strength, programs vs man. He then
claimed and still does that the very and only thing the SSDF-list "tells"
is how computers performs paired to each other (and therefore is of little
interest relative humans).
In statistical terms he's of cause right, but he totally ignored the fact,
that the absolute level of the SSDF-list from time to time is "adjusted"
up/down against the participants performance against humans, e.g. AEGON and
others and therefore gives a pretty good picture of the potential strength
relative humans (for some years the list has been even a little
"pessimistic" compared to actual results against humans!
The point is obvious. Rebel and other don't have the knowledge of an IM. Do
not play "like" an IM – but perform like an IM against IMs as well! And
this only proves that chess is a wonderful game with very few "that how it
is…"
No coin at all.
kim
--
I wonder why!?
Kim Hvarre
ph...@doubt.com
clea...@cybernet.dk
Joe Stella
Ed Schroder <rebc...@xs4all.nl> wrote:
>But what I really have missed is that it was a discussion sofar between
>me and Bob only. Where is the rest of you?
>[...]
>- Ed Schroder -
I think part of the problem here is that the discussion isn't really
about anything tangible. You guys are arguing about definitions of
words, and also state opinions that can't be verified at this time.
For example: whether you call 50 rating points a "breakthrough" is
really up to you. You can make your choice one way or the other,
and give your reasons, but no one can say you are wrong. It's
just your opinion.
As for the "DB vs. Micros" issue, this seems to me to be more of
the same. Bob believes that DB will trounce any micro in a match,
and has stated why he believes this. Ed states why he thinks this
is wrong. But really, who knows? Until DB starts playing more
games publicly, we have no way of knowing who is right here.
So rather than join in on this debate, I am just waiting to see
how DB plays against K. in another few months. Maybe the DB
team will also play more public games after the match, and the
whole issue will be solved. I suspect that lots of other people
are doing the same, just waiting and watching and not bothering
to argue/discuss...
Joe Stella
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
In my previous posting I said the following:
I like to know for example your opinions about breakthroughs. I am
quite happy with a 25 ELO gain in one year on the same hardware and I
call 50 ELO due to software improvements only simply a breakthrough.
Going from 1600 to 1650 is nothing compared to 2410 to 2460.
Going from 1600 to 1650 is no breakthrough.
Going from 2410 to 2460 is a breakthrough.
Just ask any IM, if he (she) makes such an ELO jump the IM in question
is the happiest man (woman) in the world!
Bob answered I am wrong...
Maybe he is right, but to my best knowledge it's much more difficult
to make an ELO jump from 2410 to 2460 than from 1600 to 1650.
I know it is in computer chess, but to me this also counts for humans.
Am I wrong?
- Ed Schroder -
Robert Hyatt
Kim Hvarre (clea...@cybernet.dk) wrote:
: >
: Funny thing to observe - a year or 2 ago I had a debate with Hyatt in the
: same neighborhood, i.e. whether the ratingpool of SSDF in any way was
: representative in terms of relative strength, programs vs man. He then
: claimed and still does that the very and only thing the SSDF-list "tells"
: is how computers performs paired to each other (and therefore is of little
: interest relative humans).
: In statistical terms he's of cause right, but he totally ignored the fact,
: that the absolute level of the SSDF-list from time to time is "adjusted"
: up/down against the participants performance against humans, e.g. AEGON and
: others and therefore gives a pretty good picture of the potential strength
: relative humans (for some years the list has been even a little
: "pessimistic" compared to actual results against humans!
A statistician would have to ask however, how to you "adjust" the whole pool
based on what happens in a match like Aegon? Most FIDE ratings are based on
40/2 or 45/2.5, and the like. Aegon is faster. Most FIDE events have a
substantial incentive to win, like Linares, or whatever. I think the Aegon
results are always interesting, but I don't think they necessarily show how
the GM's vs programs will do. Check out Bronstein in the last Aegon. He made
the machines look like sub-2000 players when he got hold of them. Kasparov
made DB look like a sub-2000 player in a couple of games. We know the programs
aren't that weak. But I'm a good enough chess player to see all of the holes
in the current programs, and these GM's and even good IM's can drive a truck
through that hole.
: The point is obvious. Rebel and other don't have the knowledge of an IM. Do
: not play "like" an IM – but perform like an IM against IMs as well! And
: this only proves that chess is a wonderful game with very few "that how it
: is…"
Yes... but when you analyze carefully, the programs are using search to make up
for an acute lack of knowledge. IM's *and* Gm's can and do figure this out, and
once they see what you don't have, it's over. What I think that Ed, or anyone
ought to do is automatic play on a chess server. That is an immensely hostile
environment, because then everyone has a chance to play you over and over, and
find your weaknesses, and if you wait for a year between release versions,
you'll end up rated at 1700. The next year you might work your way back up
to 2500 for a bit after fixing those holes, but as they learn new ones, you
go right back down. The humans learn and adapt quite quickly. If the machines
have a glaring weakness, that is one missing thing...
: No coin at all.
Robert Hyatt
Ed Schroder (rebc...@xs4all.nl) wrote:
: From: "Torstein Hall" <tors...@eunet.no>
: : I find this rating talk to be a little funny. If you put Rebel up
: : against a IM with 2465 in rating (Rebel 8 rating on SSDF) I would be
: : willing to bet quite a lot, that it would lose this match with time
: : controls set 2h 40 moves.
: : This puts the talk about rating improvments into its right position.
: : It only counts for play agains computers as its only measured agains
: : computers.
: : Would you put Rebel 8 up against an IM in such a match and calculate
: : its rating next. I guess it would end up around in the area of
: : 2200-2300. Anyway it would have been very interesting to see such a
: : match.
: Perhaps you underestimate the micro's.
: Look at AEGON (man versus machine) Rebel and other micro's played very
: well against strong IM's and GM's.
: ELO TPR for Rebel the last 3 years...
: AEGON 1996 : 2530
: AEGON 1995 : 2460
: AEGON 1994 : 2470
: That's not in the 2200-2300 range but clearly above.
: It's also not a lucky shot since it's from the last 3 years and quite
: steady... The same counts for other micro's especially Fritz (1996)
: and Hiarcs (1995).
: AEGON 1996 (6 rounds)
: Rebel - GM Christianssen 1-0
: Rebel - GM Vaganian 1/2 - 1/2
: Rebel - IM Hoeksema 1/2 - 1/2
I agree that all of us can do well against IM's. And GM's... if you look
at the computer vs GM event on ICC, 5 programs (Ferret, Rebel, Crafty,
Genius and I think Fritz but I'm not sure) played 5 GM's in game/30. The
worse computer had a better result than the best GM. Do I believe that any
of those programs are really better than the GM's? No.
There are two issues to consider. The first is that most GM's make a living
at playing chess. This means that they expend a lot of mental energy and
time to prepare for an event, because there's money on the line. In this
GM/computer event, there's no incentive for them to show off their latest pet
opening innovation, and *every* one of them has a bag full of these as they
will readily tell you if you are bold enough to ask. They won't show you, but
they will admit they have some.
Therefore, I'd consider it interesting that such results happen at Aegon,
or in other matches like the chess.net match, but that's not quite the same
thing as getting a real IM to sit down across the table and play real chess,
at the level he's capable of. If I believed that IM's play their best chess
all the time, I'd claim a 2500+ Elo for Crafty, because I have a couple of
thousand games against a particularly strong IM, at long time controls, like
15 30, which is a pretty long game. However, I know they don't play their very
best, any more than I'd be able to play my very best at a local chess club
Wednesday nite tourney.
The DB/Kasparov match is a horse of a different color, however, because it is
a real game, at a real GM-type tourney time control, and with enough money on
the line that I'd certainly go into hibernation and prepare were I Kasparov,
simply because I've never had $700,000 offered to me at any time in my life.
*That* is one heck of an incentive...
I still believe, that if the best micros could play in the FIDE circuit, we'd
be seeing ratings of 2300-2350... maybe a *very* rare 2400, but I doubt it at
present. And that's not slamming anybody, because I'm one of the micro guys
at present. And I do believe that will slowly inch up. It's just that we see
so much of the SSDF 2450, 2500 numbers that they begin to appear to be absolute
numbers, when they aren't...
: Playing level 40 in 1:30:00
: After that 15:00 for the rest of the game
: Bronstein method: add 0:20 for every move
: This comes pretty close to 40 in 2:00
: Average is about 2:15 per move.
: - Ed Schroder -
: : Torstein Hall
: : tors...@eunet.no
: : http://login.eunet.no/torshall/sjakk.html
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: Perhaps you underestimate the micro's.
: Look at AEGON (man versus machine) Rebel and other micro's played very
: well against strong IM's and GM's.
: ELO TPR for Rebel the last 3 years...
: AEGON 1996 : 2530
: AEGON 1995 : 2460
: AEGON 1994 : 2470
: That's not in the 2200-2300 range but clearly above.
: It's also not a lucky shot since it's from the last 3 years and quite
: steady... The same counts for other micro's especially Fritz (1996)
: and Hiarcs (1995).
: AEGON 1996 (6 rounds)
: Rebel - GM Christianssen 1-0
: Rebel - GM Vaganian 1/2 - 1/2
: Rebel - IM Hoeksema 1/2 - 1/2
> I agree that all of us can do well against IM's. And GM's... if you
> look at the computer vs GM event on ICC, 5 programs (Ferret, Rebel,
> Crafty, Genius and I think Fritz but I'm not sure) played 5 GM's in
> game/30. The worse computer had a better result than the best GM.
> Do I believe that any of those programs are really better than the
> GM's? No.
Agreed...
> There are two issues to consider. The first is that most GM's make a
> living at playing chess. This means that they expend a lot of mental
> energy and time to prepare for an event, because there's money on the
> line. In this GM/computer event, there's no incentive for them to
> show off their latest pet opening innovation, and *every* one of them
> has a bag full of these as they will readily tell you if you are bold
> enough to ask. They won't show you, but they will admit they have some.
You are misinformed, at Aegon money is at steak. Also no GM likes to
lose from a computer since this is bad for his reputation. The next days
it's in all dutch papers, then the loss is in all the chess magazines
worldwide. That's not funny for a GM. Computers are suppose to lose
from a GM... :)
Also I have seen the GM's sweat while playing and their face wasn't
exactly happy when they lost. I think you are wrong, you should come
this year to Aegon and taste it yourself. Aegon is dead serious chess.
> I still believe, that if the best micros could play in the FIDE
> circuit, we'd be seeing ratings of 2300-2350... maybe a *very* rare
> 2400, but I doubt it at present. And that's not slamming anybody,
> because I'm one of the micro guys at present. And I do believe that
> will slowly inch up. It's just that we see so much of the SSDF 2450,
> 2500 numbers that they begin to appear to be absolute numbers, when
> they aren't...
You keep on saying this but the facts are against you. See the
above results of Rebel: 2530, 2460 and 2470 steady in 3 years.
Here are the amazing AEGON 96 results of Quest (Fritz)...
Quest - Loewenthal (about 2000) 1-0
Peng Zhaoqin (2410) - Quest 0-1
Quest - David Bronstein (GM) 1-0
Quest - Jonathan Speelman (GM) 1/2 - 1/2
Roberto Cifuentes (GM) - Quest 1/1 - 1/2
Quest - John van der Wiel (GM) 1/2-1/2
TPR (FIDE) ELO rating Quest 2610!
Impressive isn't it?
Four GM's and no loss, to me a highlight in computer chess...
Also the ratings of Quest, Hiarcs and Mchess were VERY GOOD the last
three years. Steady results no lucky shots!
Bob you are totally wrong here.
See also the posting of Enrique one of the best postings I ever read in
RGCC concerning this subject because it's very informative about the
results and ratings of the micro's against strong human players (mainly
GM's) I am curious on your reply of this posting, so many facts.
Till now the victory list of the micro's is much more impressive than
the list from DB. Wish they would play more...
- Ed Schroder -
Ed Schroder
Len
Ed Schroder wrote:
[snip]
> Maybe he is right, but to my best knowledge it's much more difficult
> to make an ELO jump from 2410 to 2460 than from 1600 to 1650.
>
> I know it is in computer chess, but to me this also counts for humans.
>
> Am I wrong?
>
> - Ed Schroder -
In the USA, I believe you only get 8 points for winning a game if your
rating is over 2100. Under 2100 you get 16 points (assuming you play an
opponent rated the same as you). This means you need to win twice as
many games to go from 2410 to 2460 than from 1600 to 1650.
Len
Kim Hvarre
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
<5enjk0$f...@juniper.cis.uab.edu>...
> Kim Hvarre (clea...@cybernet.dk) wrote:
> : >
SNIP
Think that most statisticians will settle with "pretty good" (se above) +/-
;-)
…and in regards of how certain GMs makes app.s look like (rarely), well the
same thing happens between GMs all the time – then we call it a bad day,
fatigue, etc.
Again I think that Ed has answered this question elsewhere. If it really
was that "easy" to outlook "computerplay", then the "accumulated score" up
till now GM/IM vs. computers ought to look quite different, i.e. in clear
favor to the humans, but…
And agree – every programmer should consider to benefit of the experiences
gained from the chessservers. Beating humans is the goal!
Torstein Hall
Robert Hyatt <hy...@crafty.cis.uab.edu> skrev i artikkelen
<5enj2m$f...@juniper.cis.uab.edu>...
> Ed Schroder (rebc...@xs4all.nl) wrote:
> : From: "Torstein Hall" <tors...@eunet.no>
>
> : : I find this rating talk to be a little funny. If you put Rebel up
> : : against a IM with 2465 in rating (Rebel 8 rating on SSDF) I would be
> : : willing to bet quite a lot, that it would lose this match with time
> : : controls set 2h 40 moves.
>
> : : This puts the talk about rating improvments into its right position.
> : : It only counts for play agains computers as its only measured agains
> : : computers.
>
> : : Would you put Rebel 8 up against an IM in such a match and calculate
> : : its rating next. I guess it would end up around in the area of
> : : 2200-2300. Anyway it would have been very interesting to see such a
> : : match.
>
>
> : Perhaps you underestimate the micro's.
>
> : Look at AEGON (man versus machine) Rebel and other micro's played very
> : well against strong IM's and GM's.
>
> : ELO TPR for Rebel the last 3 years...
> : AEGON 1996 : 2530
> : AEGON 1995 : 2460
> : AEGON 1994 : 2470
>
> : That's not in the 2200-2300 range but clearly above.
Perhaps I underestimated your programs a little bit, but at what
timecontrol was Aegon? Not 2 hours 40 moves?
> : It's also not a lucky shot since it's from the last 3 years and quite
> : steady... The same counts for other micro's especially Fritz (1996)
> : and Hiarcs (1995).
>
> : AEGON 1996 (6 rounds)
> : Rebel - GM Christianssen 1-0
> : Rebel - GM Vaganian 1/2 - 1/2
> : Rebel - IM Hoeksema 1/2 - 1/2
I cant find the Aegon pages anymore, but was it not an all out humans at
top last year?
> I agree that all of us can do well against IM's. And GM's... if you
look
> at the computer vs GM event on ICC, 5 programs (Ferret, Rebel, Crafty,
> Genius and I think Fritz but I'm not sure) played 5 GM's in game/30. The
> worse computer had a better result than the best GM. Do I believe that
any
> of those programs are really better than the GM's? No.
>
> There are two issues to consider. The first is that most GM's make a
living
> at playing chess. This means that they expend a lot of mental energy and
> time to prepare for an event, because there's money on the line. In this
> GM/computer event, there's no incentive for them to show off their latest
pet
> opening innovation, and *every* one of them has a bag full of these as
they
> will readily tell you if you are bold enough to ask. They won't show
you, but
> they will admit they have some.
>
> Therefore, I'd consider it interesting that such results happen at Aegon,
> or in other matches like the chess.net match, but that's not quite the
same
> thing as getting a real IM to sit down across the table and play real
chess,
> at the level he's capable of. If I believed that IM's play their best
chess
> all the time, I'd claim a 2500+ Elo for Crafty, because I have a couple
of
> thousand games against a particularly strong IM, at long time controls,
like
> 15 30, which is a pretty long game. However, I know they don't play
their very
> best, any more than I'd be able to play my very best at a local chess
club
> Wednesday nite tourney.
Play on ICC and over a chess table in a competition is very different! Its
quite a lot of disturbance at home, like TV, Kids and by the way I think
chess is played on a wooden board for a human.
> The DB/Kasparov match is a horse of a different color, however, because
it is
> a real game, at a real GM-type tourney time control, and with enough
money on
> the line that I'd certainly go into hibernation and prepare were I
Kasparov,
> simply because I've never had $700,000 offered to me at any time in my
life.
> *That* is one heck of an incentive...
>
> I still believe, that if the best micros could play in the FIDE circuit,
we'd
> be seeing ratings of 2300-2350... maybe a *very* rare 2400, but I doubt
it at
> present. And that's not slamming anybody, because I'm one of the micro
guys
> at present. And I do believe that will slowly inch up. It's just that
we see
> so much of the SSDF 2450, 2500 numbers that they begin to appear to be
absolute
> numbers, when they aren't...
I agree very completle to that las paragraph!
vince
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
<5enjk0$f...@juniper.cis.uab.edu>...
....The humans learn and adapt quite quickly. If the machines
> have a glaring weakness, that is one missing thing...
>
what about the learning functions? wouldnt it be fair for a computer to be
able to build learning functions for each particular opponent? humans can,
so computers should be able to also. the next question is whether the
functions are valid enough and quick enough to keep up.
vince
vince
Ed Schroder <rebc...@xs4all.nl> wrote in article
<5enfe4$b75$1...@news1.xs4all.nl>...
>
...but to my best knowledge it's much more difficult
> to make an ELO jump from 2410 to 2460 than from 1600 to 1650.
>
> I know it is in computer chess, but to me this also counts for humans.
>
if that is true, that may explain much of the argument about diminishing
returns from increasing speed. if so, then adding knowledge to strong
programs should suffer similar diminishing returns.
vince
Robert Hyatt
vince (v...@ziplink.net) wrote:
: Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
: <5enjk0$f...@juniper.cis.uab.edu>...
: vince
No... computer learning is *so* primitive. Crafty learns that book lines
are good or bad, and it learns that a particular position is good or
bad... but it will take forever to learn how to beat kasparov. The chess
tree is *so* large...
Humans learn different things... for example, what weakness a program has,
and then they play to this weakness over and over... Not the same identical
position, or opening line, but the same "theme." That's a very difficult
thing to even think about how to do...
Robert Hyatt
vince (v...@ziplink.net) wrote:
: Ed Schroder <rebc...@xs4all.nl> wrote in article
: vince
:
What Ed was talking about, in simple terms, is simple knowledge steps will
improve you from 1600-1650... but once you get to 2300 or so, all of the
simple and easy-to-implement knowledge is in... and what's left is much
more difficult to program.
What I was talking about is that it doesn't take much, when playing two
programs against each other, to upset the applecart... A minor change can
produce a big match improvement in versions of the same program that only
differ in some small knowledge component.
Quite often you can add something, and if you auto-play 100 games, and
run it thru Elo's formula, you get something quite good... but when you
play it against a group of humans, there's little or no improvement. This
has been reported many times... and is aggravating to be sure...
mclane
Ed Schroder <rebc...@xs4all.nl> wrote:
>From: shi...@nntp.best.com (Joe McCaughan)
>: Hmmmm....
>: Things seem to be getting a little quiet out there...
>: Ed? Anything to say?
>Hi,
>In principal I have said everything what is on my mind and I don't like
>repetitions, they as you know only end in a draw... :)
>But what I really have missed is that sofar it was a discussion between
>Bob and me only. Where is the rest of you?
What shall * I * say ?? Whenever I say something I am attacked
BECAUSE I write.
In your thread many of your sentences could have been out of MY mouth.
So I read anything you wrote with great amusement and thought:
Thay always try to damage Chris reputation when they have no further
arguments/points against him.
I think: they always could say: Oh - this mclane is a nobody, he
counts nothing. He is not a programmer. He is maybe crazy.
If Rolf posts anything they say he is crazy.
I then think: they cannot do the same GAME with you. It will not work,
because you are a historical person, you normally do not overreact,
you have success and you are a world-champion
the only pity for them is:
that you are often Chris' and my opinon.
E.g. in this case.
So - these are my thoughts.
I have tried to fight many threads again, my
>Bob and I are just 2 persons with only 2 brains and 2 opinions and
>thoughts. There are hundreds of people here all with opinions about
>this subject.
Right, we should ask intelligent guys like Moritz Berger, Alexander
Fuchs, Dirk Frickenschmidt, Stefan-Meyer-Kahlen, Enrique Irazoqui,
Peter Schreiner, Vincent, Peter Gillgasch, and and and the rest,
>So write them, I like to hear them...
>I like to know for example your opinions about breakthroughs. I am
>quite happy with a 25 ELO gain in one year on the same hardware and I
>call 50 ELO due to software improvements only simply a breakthrough.
>Going from 1600 to 1650 is nothing compared to 2410 to 2460.
Of course true.
>Going from 1600 to 1650 is no breakthrough.
>Going from 2410 to 2460 is a breakthrough.
Right.
>Just ask any IM, if he (she) makes such an ELO jump the IM in question
>is the happiest man (woman) in the world!
Right.
>I have read Bob's comments on that and besides the fact I find his
>comments very unfounded I like to hear yours.
>Are Bob and I the only ones here in RGCC?
Certainly not.
>I hope not...
>- Ed Schroder -
But Ed - what do you want ?
In my opinon you want a fight against Deep Blue or another big iron.
I am not sure if they will give you the chance, because they could be
afraid a little. Your program is at the peak in the moment.
So they will say: we consider about, have to organize and other stuff
like that.
In the end, it is too late for Rebel8 to show it's strength.
I am looking forward for a fight. I think Rebel is still the strongest
program. Than Mchess and than maybe Hiarcs.
You have made a big jump with your rebel.
It was easy to find out. I did not need much time to find out.
So much progress was it.
But know you are number one and anybody is haunting you and rebel.
Thats the nice situation Richard always had. How do you feel now?
mclane
Ed Schroder <rebc...@xs4all.nl> wrote:
>From: shi...@nntp.best.com (Joe McCaughan)
>: Hmmmm....
>: Things seem to be getting a little quiet out there...
>: Ed? Anything to say?
>Hi,
>In principal I have said everything what is on my mind and I don't like
>repetitions, they as you know only end in a draw... :)
In principal Ed does not like repetitions. Therefore he sents anything
thrice.
Ed - there is a nice Science-Fictions classics called
"Thrice upon a time" by James P. Hogan.
Ever read ???
Thoralf Karlsson
In article <5enjk0$f...@juniper.cis.uab.edu>, hy...@crafty.cis.uab.edu
says...
>
>Kim Hvarre (clea...@cybernet.dk) wrote:
>: >
>: In statistical terms he's of cause right, but he totally ignored the
fact,
>: that the absolute level of the SSDF-list from time to time is
"adjusted"
>: up/down against the participants performance against humans, e.g.
AEGON and
>: others and therefore gives a pretty good picture of the potential
strength
>: relative humans (for some years the list has been even a little
>: "pessimistic" compared to actual results against humans!
>
>A statistician would have to ask however, how to you "adjust" the whole
pool
>based on what happens in a match like Aegon?
The level of the SSDF rating list has not been changed since 1991. It is
based on 337 tournament games against humans played in Sweden between
1987 and 1991. Since then we have played very few games against humans at
40/2. We have never used games played against humans in other countries
for adjusting the level of our list.
Thoralf Karlsson
SSDF
Ed Schroder
From: mcl...@prima.ruhr.de (mclane)
> But what I really have missed is that sofar it was a discussion between
> Bob and me only. Where is the rest of you?
: What shall * I * say ?? Whenever I say something I am attacked
: BECAUSE I write.
Well I do hope that it will not stop you from posting here. I agree on
many chess related views you have I only dislike words as Hitler and
other politic issues since there is absolutely no relationship to chess.
Jakarta, sure the WC was there.
: In your thread many of your sentences could have been out of MY mouth.
: So I read anything you wrote with great amusement and thought:
: Thay always try to damage Chris reputation when they have no further
: arguments/points against him.
: I think: they always could say: Oh - this mclane is a nobody, he
: counts nothing. He is not a programmer. He is maybe crazy.
I do not share your opinion, you are somebody from the first hour with
lots of experience and therefore a lot to share.
: I then think: they cannot do the same GAME with you. It will not work,
: because you are a historical person, you normally do not overreact,
: you have success and you are a world-champion
: the only pity for them is:
: that you are often Chris' and my opinon.
True Chris and I have much in common about "the way to go" in computer
chess concerning playing strength, but there are other opinions too and
I like to listen to them and learn from them.
: E.g. in this case.
: So - these are my thoughts.
: I have tried to fight many threads again, my
> Bob and I are just 2 persons with only 2 brains and 2 opinions and
> thoughts. There are hundreds of people here all with opinions about
> this subject.
: Right, we should ask intelligent guys like Moritz Berger, Alexander
: Fuchs, Dirk Frickenschmidt, Stefan-Meyer-Kahlen, Enrique Irazoqui,
: Peter Schreiner, Vincent, Peter Gillgasch, and and and the rest,
Yes.
[ snip ]
: But Ed - what do you want ?
: In my opinon you want a fight against Deep Blue or another big iron.
: I am not sure if they will give you the chance, because they could be
: afraid a little. Your program is at the peak in the moment.
Oh no!
Playing against "the beast" would be nice but it is (was) never my
intention to challenge them. This is silly, they can't afford the risk!
Suppose, just suppose they lose... IBM is right, a Deep Blue versus PC
match will never happen.
I can imagine that my discussion with Bob sounded that I was looking for
such a match but I was just jumping in the discussion because in my
opinion Bob overreacted about the strength of Deep Blue. So I chose
sides for the poor Pc guys in the underdog position fighting against
the NPS dream... :)
To me with a lot of good statements from both sides.
: So they will say: we consider about, have to organize and other stuff
: like that.
: In the end, it is too late for Rebel8 to show it's strength.
: I am looking forward for a fight. I think Rebel is still the strongest
: program. Than Mchess and than maybe Hiarcs.
: You have made a big jump with your rebel.
: It was easy to find out. I did not need much time to find out.
: So much progress was it.
: But know you are number one and anybody is haunting you and rebel.
: Thats the nice situation Richard always had. How do you feel now?
Hunted... :)
I feel the breath of Mark... :)
- Ed Schroder -
Chris Whittington
--
http://www.demon.co.uk/oxford-soft
Ed Schroder <rebc...@xs4all.nl> wrote in article
<5enq1g$gcu$1...@news1.xs4all.nl>...
> From: hy...@crafty.cis.uab.edu (Robert Hyatt)
>
> : Perhaps you underestimate the micro's.
>
> : Look at AEGON (man versus machine) Rebel and other micro's played very
> : well against strong IM's and GM's.
>
> : ELO TPR for Rebel the last 3 years...
> : AEGON 1996 : 2530
> : AEGON 1995 : 2460
> : AEGON 1994 : 2470
>
> : That's not in the 2200-2300 range but clearly above.
>
> : It's also not a lucky shot since it's from the last 3 years and quite
> : steady... The same counts for other micro's especially Fritz (1996)
> : and Hiarcs (1995).
>
> : AEGON 1996 (6 rounds)
> : Rebel - GM Christianssen 1-0
> : Rebel - GM Vaganian 1/2 - 1/2
> : Rebel - IM Hoeksema 1/2 - 1/2
>
> > I agree that all of us can do well against IM's. And GM's... if you
> > look at the computer vs GM event on ICC, 5 programs (Ferret, Rebel,
> > Crafty, Genius and I think Fritz but I'm not sure) played 5 GM's in
> > game/30. The worse computer had a better result than the best GM.
> > Do I believe that any of those programs are really better than the
> > GM's? No.
>
> Agreed...
>
>
> > There are two issues to consider. The first is that most GM's make a
> > living at playing chess. This means that they expend a lot of mental
> > energy and time to prepare for an event, because there's money on the
> > line. In this GM/computer event, there's no incentive for them to
> > show off their latest pet opening innovation, and *every* one of them
> > has a bag full of these as they will readily tell you if you are bold
> > enough to ask. They won't show you, but they will admit they have
some.
>
> You are misinformed, at Aegon money is at steak. Also no GM likes to
> lose from a computer since this is bad for his reputation. The next days
> it's in all dutch papers, then the loss is in all the chess magazines
> worldwide. That's not funny for a GM. Computers are suppose to lose
> from a GM... :)
>
> Also I have seen the GM's sweat while playing and their face wasn't
> exactly happy when they lost. I think you are wrong, you should come
> this year to Aegon and taste it yourself. Aegon is dead serious chess.
>
>
> > I still believe, that if the best micros could play in the FIDE
> > circuit, we'd be seeing ratings of 2300-2350... maybe a *very* rare
> > 2400, but I doubt it at present. And that's not slamming anybody,
> > because I'm one of the micro guys at present. And I do believe that
> > will slowly inch up. It's just that we see so much of the SSDF 2450,
> > 2500 numbers that they begin to appear to be absolute numbers, when
> > they aren't...
>
> You keep on saying this but the facts are against you. See the
> above results of Rebel: 2530, 2460 and 2470 steady in 3 years.
>
> Here are the amazing AEGON 96 results of Quest (Fritz)...
>
> Quest - Loewenthal (about 2000) 1-0
> Peng Zhaoqin (2410) - Quest 0-1
> Quest - David Bronstein (GM) 1-0
> Quest - Jonathan Speelman (GM) 1/2 - 1/2
> Roberto Cifuentes (GM) - Quest 1/1 - 1/2
> Quest - John van der Wiel (GM) 1/2-1/2
>
> TPR (FIDE) ELO rating Quest 2610!
>
> Impressive isn't it?
>
> Four GM's and no loss, to me a highlight in computer chess...
>
> Also the ratings of Quest, Hiarcs and Mchess were VERY GOOD the last
> three years. Steady results no lucky shots!
>
> Bob you are totally wrong here.
>
> See also the posting of Enrique one of the best postings I ever read in
> RGCC concerning this subject because it's very informative about the
> results and ratings of the micro's against strong human players (mainly
> GM's) I am curious on your reply of this posting, so many facts.
>
> Till now the victory list of the micro's is much more impressive than
> the list from DB. Wish they would play more...
Ed, there's always a danger with so many programs playing that you can just
cherry pick the best results, or best results series.
Suppose you were to cooect up all the results from say the last three years
of all micros and dedicated machines that were within, say, 150 elo's of
the top of the ssdf, and then generate a 'combined' cptr-human elo ......
What then ?
Chris Whittington
>
> - Ed Schroder -
>
>
>
Ed Schroder
From: "Chris Whittington" <chr...@demon.co.uk>
> : Perhaps you underestimate the micro's.
[ snip ]
> You are misinformed, at Aegon money is at steak. Also no GM likes to
> lose from a computer since this is bad for his reputation. The next days
> it's in all dutch papers, then the loss is in all the chess magazines
> worldwide. That's not funny for a GM. Computers are suppose to lose
> from a GM... :)
> Also I have seen the GM's sweat while playing and their face wasn't
> exactly happy when they lost. I think you are wrong, you should come
> this year to Aegon and taste it yourself. Aegon is dead serious chess.
: Ed, there's always a danger with so many programs playing that you
: can just cherry pick the best results, or best results series.
: Suppose you were to cooect up all the results from say the last three
: years of all micros and dedicated machines that were within, say, 150
: elo's of the top of the ssdf, and then generate a 'combined' cptr-human
: elo ......
: What then ?
Hi Chris,
Hope you had a good holiday.
The above subject was not meant to "highlight" how incredible good the
Pc programs are but when you were on your ski's we here had a giant
discussion about Deep Blue speculating about the playing strength.
After that the discussion moved to the PC chess programs direction. In
the discussion I chose sides for defending the Pc programs and Bob for
Deep Blue and we had some very good discussions.
So don't see all the Aegon PC highlights as some kind of blurb but as
a part of the speculative discussion Deep Blue versus PC programs
concerning playing strength.
To answer your question if you take the results of Quest (Fritz), Rebel,
Mchess and Hiarcs of the last 2 years the average elo is above 2400.
- Ed Schroder -
: Chris Whittington
mclane
Ed Schroder <rebc...@xs4all.nl> wrote:
>From: "Torstein Hall" <tors...@eunet.no>
>: I find this rating talk to be a little funny. If you put Rebel up
>: against a IM with 2465 in rating (Rebel 8 rating on SSDF) I would be
>: willing to bet quite a lot, that it would lose this match with time
>: controls set 2h 40 moves.
>: This puts the talk about rating improvments into its right position.
>: It only counts for play agains computers as its only measured agains
>: computers.
>: Would you put Rebel 8 up against an IM in such a match and calculate
>: its rating next. I guess it would end up around in the area of
>: 2200-2300. Anyway it would have been very interesting to see such a
>: match.
>Perhaps you underestimate the micro's.
They do. They do. Each year we are having MAN vs. Machine in Den Haag,
with 50 humans vs. 50 programs. And each year the PC results are
brilliant. They can related to Deep Blue's results.
But the difference is: the PC's use only SLOW hardware !!!
You cannot say that these results are randomn because AEGON-tournament
is held each year and always 6 games.
Also we they do many blitz-games there, in the evenings.
They all show us the same: the micro's are strong as hell.
>Look at AEGON (man versus machine) Rebel and other micro's played very
>well against strong IM's and GM's.
>ELO TPR for Rebel the last 3 years...
>AEGON 1996 : 2530
>AEGON 1995 : 2460
>AEGON 1994 : 2470
>That's not in the 2200-2300 range but clearly above.
>It's also not a lucky shot since it's from the last 3 years and quite
>steady... The same counts for other micro's especially Fritz (1996)
>and Hiarcs (1995).
>AEGON 1996 (6 rounds)
>Rebel - GM Christianssen 1-0
>Rebel - GM Vaganian 1/2 - 1/2
>Rebel - IM Hoeksema 1/2 - 1/2
>Playing level 40 in 1:30:00
Jose Lopez Jr.
On 23 Feb 1997 04:14:24 GMT, hy...@crafty.cis.uab.edu (Robert Hyatt)
wrote:
>No... computer learning is *so* primitive. Crafty learns that book lines
>are good or bad, and it learns that a particular position is good or
>bad... but it will take forever to learn how to beat kasparov. The chess
>tree is *so* large...
>
>Humans learn different things... for example, what weakness a program has,
>and then they play to this weakness over and over... Not the same identical
>position, or opening line, but the same "theme." That's a very difficult
>thing to even think about how to do...
>
>
Correct! Hopefully, Kasparov will not play the same games in which he
clearly won on the previous match. I really don't think he would do
this and I know the Deep Blue team has ran those games Deep Blue lost
to Kasparov and found Deep Blue not making those same moves and
hopefully better ones. It would be embarrassing if these same games
be played and Deep Blue make the same wrong moves. I know Kasparov
knows how Deep Blue thinks (weakness) on those particular games and
may use that same line of moves against it on the first rematch game
to see if there was any improvement on Deep Blue's part.
Chris Whittington
--
http://www.demon.co.uk/oxford-soft
Ed Schroder <rebc...@xs4all.nl> wrote in article
<85670673...@gln01-04.dial.xs4all.nl>...
> From: "Chris Whittington" <chr...@demon.co.uk>
>
> > : Perhaps you underestimate the micro's.
>
> [ snip ]
>
> > You are misinformed, at Aegon money is at steak. Also no GM likes to
> > lose from a computer since this is bad for his reputation. The next
days
> > it's in all dutch papers, then the loss is in all the chess magazines
> > worldwide. That's not funny for a GM. Computers are suppose to lose
> > from a GM... :)
>
> > Also I have seen the GM's sweat while playing and their face wasn't
> > exactly happy when they lost. I think you are wrong, you should come
> > this year to Aegon and taste it yourself. Aegon is dead serious chess.
>
>
> : Ed, there's always a danger with so many programs playing that you
> : can just cherry pick the best results, or best results series.
>
> : Suppose you were to cooect up all the results from say the last three
> : years of all micros and dedicated machines that were within, say, 150
> : elo's of the top of the ssdf, and then generate a 'combined' cptr-human
> : elo ......
>
> : What then ?
>
>
> Hi Chris,
>
> Hope you had a good holiday.
>
> The above subject was not meant to "highlight" how incredible good the
> Pc programs are but when you were on your ski's we here had a giant
> discussion about Deep Blue speculating about the playing strength.
Yes, I saw. There were more than 1000 messages in the seven days I was away
:(
Too many.
>
> After that the discussion moved to the PC chess programs direction. In
> the discussion I chose sides for defending the Pc programs and Bob for
> Deep Blue and we had some very good discussions.
I read some posts saying that PC programs were making great leaps forward,
just like Deng Chaio Ping.
If this is true then there must be some technical basis. Where is this
great new programming algorithm ?
Also Thorsten tells me that in his tests he still thinks the programs are
playing the same.
Of course, I don't to take the side of Bob, so it must be that DB is
standing still the same way. A few more processors here, An MMX and PP 200
there .....
Chris Whittington
Ingo Althoefer
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
> I still believe, that if the best micros could play in the FIDE
> circuit, we'd be seeing ratings of 2300-2350... maybe a *very* rare
> 2400, but I doubt it at present.
It might be interesting to compare two different settings for a tournament
between human players and computers.
Assume six human players A,B,C,D,E,F and six different computers a,b,c,d,e,f.
Let them play six rounds.
Setting (i): every human is playing once against every computer.
Setting (ii): human A plays a 6-games match against computer a,
human B plays a 6-games match against computer b,
...
human F plays a 6-games match against computer f.
Assume that your humans have ratings around 2400 Fide Elo, and that the
computers have ratings around 2400 SSDF Elo. Then I believe that setting (i)
would show a much better performance of the computers than setting (ii), as
humans learn very quickly from identical opponents ( like in (ii) ) and
different computers play rather differently.
Probably average computer performances of 2450 in (i) and of 2300 in (ii)
were realistic. ( If 8- instead of 6-games matches were played, the computer's
performance might even drop below 2300, as humans adapt better in longer
matches. )
Also different ways of payment may yield different types of results.
There are at least three pure possibilities for matches between (professional)
human and computer:
1. The human gets a fixed amount independently of the outcome.
2. Only the win/draw/loss status of the m a t c h as a whole counts:
In case of a match loss nothing is paid, for a match draw x $ are given.
For a match win the human gets 2*x $.
3. The payment is coupled purely linear to the success. So, if the human
makes m points he gets m*y $, where y $ is his reward for 1 point.
I think that payment according to 1. makes it easier ( but probably not easy
in an absolute sense ) to get a good result against a professional player.
Comparing 2. and 3. is difficult. When 2. is applied, it will be very hard
to beat the human, whereas it may be rather easy to get slightly less than
50 percent of the points. Also it may be easy to get a very clear win if you
are on the winning road once.
As examples see for instance some 2-games matches in active chess:
Kasparov - Genius 1.5 : 0.5 ( Cologne, May 1995 )
Kasparov - Fritz 1.5 : 0.5 ( London, December 1995 )
Karpov - Fritz 1.5 : 0.5 ( Den Haag, April 1996 )
( in all three matches the human had won the first game )
and two 4-games matches under tournament conditions:
IM David Levy - Deep Blue I 0.0 : 4.0 ( London, 1990 ? )
GM Bent Larsen - Deep Blue II 2.5 : 1.5 ( Copenhagen, 1992 ? )
( also Larsen had won the first game; 3 draws followed )
Under application of 3. match effects are rather excluded. In every
game the human is fighting for y $, almost independently of the outcomes
of the previous games.
When you are interested in getting reliable Elo ratings from your matches,
3. should be preferred to 2. .
Often mixtures of the three basic arrangements are applied. For instance:
In the 1996 DB-Kasp match 1. and 2. were mixed. There was a fixum of
100,000 $. In case of a match draw Kasparov would have got another 150,000 $.
In case of a match win ( and this really happened ) he got another 300,000 $.
In the 1997 DB-Kasp match 1. and 2. are mixed analogously. The fixum is
400,000 $. For a match draw another 150,000 $ are given. For a match win
another 300,000 $ are given.
( So only the fixum has changed from 1996 to 1997 ! )
In my match with Double-Fritz & Boss against GM Timoshchenko 1. and 3. were
mixed. There were 8 games. In case of at most 2.5 points the grandmaster
would have got 5*x $. In case of n points with n >= 2.5 he would have got
2*n*x $. ( x being some positive number. ) Timoshchenko made 3.5 points,
so he received 7*x $.
Ingo Althoefer.
brucemo
Jose Lopez Jr. wrote:
> Correct! Hopefully, Kasparov will not play the same games in which he
> clearly won on the previous match. I really don't think he would do
> this and I know the Deep Blue team has ran those games Deep Blue lost
> to Kasparov and found Deep Blue not making those same moves and
> hopefully better ones. It would be embarrassing if these same games
> be played and Deep Blue make the same wrong moves. I know Kasparov
> knows how Deep Blue thinks (weakness) on those particular games and
> may use that same line of moves against it on the first rematch game
> to see if there was any improvement on Deep Blue's part.
It is doubtful that this would happen, as the people who work on Deep Blue
can just change the opening book so these lines will be avoided. Perhaps
they will try to find improvements in these lines, but of course Kasparov
could also do this. I think the Deep Blue guys will let the machine play
Kasparov, and won't get into this sort of opening book battle.
bruce
Robert Hyatt
mclane (mcl...@prima.ruhr.de) wrote:
: Ed Schroder <rebc...@xs4all.nl> wrote:
: >From: "Torstein Hall" <tors...@eunet.no>
: >: I find this rating talk to be a little funny. If you put Rebel up
: >: against a IM with 2465 in rating (Rebel 8 rating on SSDF) I would be
: >: willing to bet quite a lot, that it would lose this match with time
: >: controls set 2h 40 moves.
: >: This puts the talk about rating improvments into its right position.
: >: It only counts for play agains computers as its only measured agains
: >: computers.
: >: Would you put Rebel 8 up against an IM in such a match and calculate
: >: its rating next. I guess it would end up around in the area of
: >: 2200-2300. Anyway it would have been very interesting to see such a
: >: match.
: >Perhaps you underestimate the micro's.
: They do. They do. Each year we are having MAN vs. Machine in Den Haag,
: with 50 humans vs. 50 programs. And each year the PC results are
: brilliant. They can related to Deep Blue's results.
: But the difference is: the PC's use only SLOW hardware !!!
: You cannot say that these results are randomn because AEGON-tournament
: is held each year and always 6 games.
: Also we they do many blitz-games there, in the evenings.
: They all show us the same: the micro's are strong as hell.
I didn't say they weren't. I see this performance from the micros every
day on ICC. What I do say is that I don't believe that there is a micro
program that plays what I would consider IM-class chess at present. They
are not terribly far away from it. At blitz many are easily what I would
call super-GM players. I can post some interesting results from Crafty vs
GM's to show what I mean.
One thing that is often overlooked, however, is that in general, GM's play
"chess". When they want to, they can play a new and different game called
"anti-silicon chess" that is very damaging to computer programs. To see what
I mean, pop over to ICC and match your favorite program against "greg1" and
see what happens. He's a problem. And the longer the time control, the bigger
the problem. At blitz he's very strong, because he loves to attack, and
programs are simply lousy at understanding the long-range implications of some
positions. At longer time controls, he gets better, because he's also tactically
very strong, and makes fewer tactical mistakes to lose won positions, as he does
on occasion in very quick blitz games.
: >Look at AEGON (man versus machine) Rebel and other micro's played very
: >well against strong IM's and GM's.
: >ELO TPR for Rebel the last 3 years...
: >AEGON 1996 : 2530
: >AEGON 1995 : 2460
: >AEGON 1994 : 2470
: >That's not in the 2200-2300 range but clearly above.
Maybe and maybe not. TPR is not Elo. It's an estimate. We've seen lots
of high TPR results over the year, but the actual USCF ratings have always
been lower. We need real data, from programs playing in real tournaments,
for many games. That would give a real indication, which we simply don't
get at present...
: >It's also not a lucky shot since it's from the last 3 years and quite
Jean-Peter Fendrich
Robert Hyatt wrote:
>
> I agree that all of us can do well against IM's. And GM's... if you look
> at the computer vs GM event on ICC, 5 programs (Ferret, Rebel, Crafty,
> Genius and I think Fritz but I'm not sure) played 5 GM's in game/30. The
> worse computer had a better result than the best GM. Do I believe that any
> of those programs are really better than the GM's? No.
>
> at playing chess. This means that they expend a lot of mental energy and
> time to prepare for an event, because there's money on the line. In this
> GM/computer event, there's no incentive for them to show off their latest pet
> opening innovation, and *every* one of them has a bag full of these as they
> will readily tell you if you are bold enough to ask. They won't show you, but
> they will admit they have some.
>
> or in other matches like the chess.net match, but that's not quite the same
> thing as getting a real IM to sit down across the table and play real chess,
> at the level he's capable of. If I believed that IM's play their best chess
> all the time, I'd claim a 2500+ Elo for Crafty, because I have a couple of
> thousand games against a particularly strong IM, at long time controls, like
> 15 30, which is a pretty long game. However, I know they don't play their very
> best, any more than I'd be able to play my very best at a local chess club
> Wednesday nite tourney.
to 'full' time games.
>
> The DB/Kasparov match is a horse of a different color, however, because it is
> a real game, at a real GM-type tourney time control, and with enough money on
> the line that I'd certainly go into hibernation and prepare were I Kasparov,
> simply because I've never had $700,000 offered to me at any time in my life.
> *That* is one heck of an incentive...
>
> be seeing ratings of 2300-2350... maybe a *very* rare 2400, but I doubt it at
> at present. And I do believe that will slowly inch up. It's just that we see
> so much of the SSDF 2450, 2500 numbers that they begin to appear to be absolute
> numbers, when they aren't...
I think you overestimate the effect of this. Most of these guys acts
very
professional, they just come and do their work playing as good as
possible.
The difference can't be that huge.
There are a number of games between humans and programs and most of them
are supporting the SSDF level of ratings. Often are the performance
ratings of the chess programs better than the SSDF rating as the case
from the AEGON tournament. This shouldn't be too surprisingly because
the
level of the SSDF list is adjusted according to results from tounament
games or games from tournament conditions with humans.
There is one *big but* however...
I think most people overlook the fact that humans perform very different
against computers the first times we play them compared to later on when
we really learned how the beasts are playing. Our ability to adapt makes
a real difference here. Learning abilities in the programs of today,
doesn't change this in any significant way.
There are two levels of adaption:
- Adapting in general to the way most programs are playing. Now I
believe that strong chess players have done their homework on this.
- Adapting to a very specific chess program. Most of us knows about
the effects here.
I am quite convinced of that the SSDF level is a fairly good estimate
of the performance of the first 5-10 games with the same human. After
that the performance will drop considerably. Maybe your 100+ ELO
points is the difference. The point is that most chess programs will
not meet the same GM or IM so many times. Before that, we have new
versions and new models and we are back to the 'zero number of games'
again.
The problem here is, as said many times before, how we should define
the strength of a program. Is it the performance level during it's life
time or is it the theoretical performance after an infinite number
of games against all possible opponents?
Not the same thing, believe me... :-)
--
J-P Fendrich
Mark Rawlings
Here's an interesting quote from the 4th Quarter 1990 Computer Chess
Reports ("Brains vs. Brawn," page 6):
"... its [Deep Thought's] authors admit that it is probably 300 points
weaker than the best micro programs given equal hardware, but the
great depth of search on its special purpose hardware obscures its
ignorance. If this 300 point figure is correct, then a program as
good as the latest Mephisto running on Deep Thought hardware would be
in the class with Kasparov and Karpov, but it may be that the Mephisto
program is incompatible with such hardware."
I guess this was the thinking 6 years ago!
Mark
brucemo
Who was the author of this article?
bruce
Robert Hyatt
Jean-Peter Fendrich (j...@vd.volvo.se) wrote:
I believe it can, from experience. It's doubtful that these guys have
sat across the table from one of these programs and played it a serious
game or three. When they do, they pick up on things such as (1) it doesn't
recognize long-range threats against the king; (2) they don't recognize the
potential danger of an outside passed pawn (many programs don't evaluate
this for reasons unknown to me); (3) they underestimate the value of a
knight, or overestimate the value of a bishop "just because"... and the IM
carefully shows them that a bishop is not always better, or that the bishop
pair is not always better, etc... This is a long list that is very hard to
solve without taking many years to do so. The problem is, each takes a lot
of time to fix, but only a little time to expose their absence... Find an
IM that plays computers a lot (ADOLF [Brian Hartman] on ICC is a good
example) and ask him to give you a few obvious weaknesses of computers. It
won't be tactics, obviously, but you'll get some insight..
: There are a number of games between humans and programs and most of them
: are supporting the SSDF level of ratings. Often are the performance
: ratings of the chess programs better than the SSDF rating as the case
: from the AEGON tournament. This shouldn't be too surprisingly because
: the
: level of the SSDF list is adjusted according to results from tounament
: games or games from tournament conditions with humans.
No it isn't... It was adjusted in 1991, but not since. That was apparently
to take into account about 300 human/computer games. But that was 1991. This
is 6 years and lots of hardware and software improvements later...
: There is one *big but* however...
: I think most people overlook the fact that humans perform very different
: against computers the first times we play them compared to later on when
: we really learned how the beasts are playing. Our ability to adapt makes
: a real difference here. Learning abilities in the programs of today,
: doesn't change this in any significant way.
: There are two levels of adaption:
: - Adapting in general to the way most programs are playing. Now I
: believe that strong chess players have done their homework on this.
: - Adapting to a very specific chess program. Most of us knows about
: the effects here.
: I am quite convinced of that the SSDF level is a fairly good estimate
: of the performance of the first 5-10 games with the same human. After
: that the performance will drop considerably. Maybe your 100+ ELO
: points is the difference. The point is that most chess programs will
: not meet the same GM or IM so many times. Before that, we have new
: versions and new models and we are back to the 'zero number of games'
: again.
Here we are in agreement. The computer will give anyone a "new look" for a
while, but the new wears off in a hurry... That's exactly why I think that should
any of the current micro programs (Crafty included, BTW) play the FIDE circuit,
the ratings would be a lot lower than most would guess. I still think the right
number is maybe 2300. *maybe*...
: The problem here is, as said many times before, how we should define
: the strength of a program. Is it the performance level during it's life
: time or is it the theoretical performance after an infinite number
: of games against all possible opponents?
: Not the same thing, believe me... :-)
Not at all. However, we certainly don't want to define this rating after a
few games, or even a few dozen. I suspect that should Crafty play at next
year's Pan Am, and everyone knows beforehand, the result will be a bit
different from this year. Now they know they'd better prepare. This year it
was a "lark"...
And when they prepare, things will change..
: --
: J-P Fendrich
Robert Hyatt
: I guess this was the thinking 6 years ago!
: Mark
Not surprising at all. They've been open in that much of what they do depends
on a big NPS value. Singular extensions is one idea.
However, don't forget that DB hardware is superior to DT hardware, in that they
did add a lot to the hardware evaluation, things like mobility, square control,
and other things that DT hardware didn't do. They learned a lot before they did
the new stuff. And the speed gap is even more drastic to boot...
Rolf Tueschen
hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>There are two issues to consider. The first is that most GM's make a living
>at playing chess. This means that they expend a lot of mental energy and
>time to prepare for an event, because there's money on the line. In this
>GM/computer event, there's no incentive for them to show off their latest pet
>opening innovation,
snip
>The DB/Kasparov match is a horse of a different color, however, because it is
>a real game, at a real GM-type tourney time control, and with enough money on
>the line that I'd certainly go into hibernation and prepare were I Kasparov,
>simply because I've never had $700,000 offered to me at any time in my life.
>*That* is one heck of an incentive...
---------------------------------------------------------------------------------------------------------------
IMHO interesting thoughts. But there are still some *floors* to watch. There
are some people who don't like that always finding new aspects and they reply
*stop your psycho analysing ...*. Science however must go on.
So I'd like to give some comment on the above.
The argument *GMs have to earn a living --> no money --> no incentive* is
correct but not the whole truth. There are other possible motives to take part
in such events machine vs men. At least the man who has to earn a living could
earn some public relations advantages ... You have to stay in the business.
Please don't forget the possible incentive of inofficial starting money
[Antrittsgeld] as well known in sports like tennis. In tennis a star plays the
first round at least even if he's not prepared to play the whole tournament.
In tennis one could read the like in Steffi Graf's father's trial.
Today's machine vs men IMHO is not regarded as serious playing among top GMs.
That's a bit fun (NOTE: each chess player loves to play chess, look at skittles
and blitz over night in hotel lobby :)), a bit good living (nice hotels) and
other preferences ...
Question of *the* scores I want to discuss with DB vs Kasparov.
IMHO a player like Kasparov always looks at possible sources for earning a
living. With computer chess he has found a very interesting field.
-- Play matches for real big money
-- Entering possible marketing strategies (and PR) with the company (INTEL, now
IBM)
-- PR out of multimedia around the world.
Now think of this: what would happen if Kasparov won 100% in each and every
match vs computer? Do you think companies would still invest a whole lot of
money? Would you further think we as the interested public would appreciate
such *stupid* farces?
If Kasparov wanted it he could *kill* DB 100%. Ok let's skip fatigue or similar
stuff. Even in the middle of the night in a hotel lobby Kasparov would crush
the monster machine at tournament time controls. In march Kasparov has more
difficulties to *invent* some suspense that stirs the chessworld's interest.
Possible rematches respectively.
It's true. Someone wrote about sort of non-sportmanship of DB and team. They
hide their player, they prepare on a well known human player and they try
simply to take maximum advantage out of this surprise setting. If they'll lose
then Elo statistics is not bad either because of K's 2900. :)
This reminds me of my own famous career when I got these smashing phantasies
after learning some book lines I hoped [:)] my opponent didn't know about ...
Fom 1900 Elo on however I detected that almost all others of equal strength had
also *learned* *my* lines. I therefore quickly learned:
Chess is not entirely descripted with the parameter *gambling*.
Today I know that even at 2900 one could have much fun in gambling ...
Roger Davis
Just a quick question: Do Deep Blue's deep pockets include money for
consulting with GMs that might suggest improvements in its play? Or guide
it toward certain openings in which Kaspy is weak. Every great player has a
great coach, right? What about Deep Blue?
Roger Davis
Robert Hyatt
Roger Davis (raj...@ix.netcom.com) wrote:
: Just a quick question: Do Deep Blue's deep pockets include money for
: Roger Davis
Rumor has had Joel Benjamin working with them, and at least a couple of others
at various time... No absolute knowledge about this on my part, although it was
posted here a couple of times by others..
Robert Hyatt
Rolf Tueschen (TUESCHEN.MEDIZ...@t-online.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
: snip
Boy, you're good. You can now figure out exactly what Kasparov's motives are,
what my motives are, and just about everyone else (marsland, levy, friedel,
and so forth.) With this much gullibility, you must own several bridges and
a few plots of land out in the desert... maybe some Edsel stock... stock in
the Russian space shuttle...
: It's true. Someone wrote about sort of non-sportmanship of DB and team. They
A.Mader
Robert Hyatt wrote:
>
> Ed Schroder (rebc...@xs4all.nl) wrote:
> : From: shi...@nntp.best.com (Joe McCaughan)
>
> : : Hmmmm....
>
> : : Things seem to be getting a little quiet out there...
>
> : : Ed? Anything to say?
>
> : Hi,
>
> : In principal I have said everything what is on my mind and I don't like
> : repetitions, they as you know only end in a draw... :)
>
> : Bob and me only. Where is the rest of you?
>
> : thoughts. There are hundreds of people here all with opinions about
> : this subject.
>
>
> : I like to know for example your opinions about breakthroughs. I am
> : quite happy with a 25 ELO gain in one year on the same hardware and I
> : call 50 ELO due to software improvements only simply a breakthrough.
>
> : Going from 1600 to 1650 is nothing compared to 2410 to 2460.
>
> 1500 players exactly the same as a spread of 50 between to 2600 players. It
> simply gives the expected # of wins for each in a match.
>
The ELO formula is simple but not that simple. There are two ways of
computing ELO ratings. One way is to use a variable factor that becomes
less and less the higher the rating gets. If the Swedes use this formula
then Ed is right. The second way is to use a fixed factor. If the Swedes
use this formula then Ed is wrong.
Of course you are right that you cannot compare SSDF ratings with FIDE
ratings, but that's not the point. If a program is able to win >60 out
of 100 games against different strong opponents, I would call this a
breakthrough.
> : Going from 1600 to 1650 is no breakthrough.
> : Going from 2410 to 2460 is a breakthrough.
>
> If you were on the very upper end of the *real* Elo scale, like, say a Kasparov,
> you are correct, because a 50 point gain is difficult. But only because you don't
> get anything for beating players 400 points below you. And if you are on the top,
> you are only playing players below you. And if you are Kasparov, most are 2-300
> points below you. Even one loss costs you 32 points. it might take 32 wins to
> pick that back up and break even.
When you have 2800 ELO and you use a variable factor when computing ELO,
it is very hard to win ELO points but it is also very hard to lose ELO
points. Draws only have a minor effect , the difference has to be very
big to influence the rating. When you play agains opponents 300 point
below you, you probably will make 85 out of 100. I think this is true
with Kasparow. :-)
>
> *But*, on the SSDF, this isn't true. Look how the programs are bunched up on
> the top, with over a dozen within 100 points of each other. No one is way out
> front, and, as a result, they don't suffer from the "compression" problem as
> above.
>
This is true, but according to the ELO formula it should not make much
(any?) difference. It simply isn't possible to lead with a big
difference to the 2nd place. Even Kasparow "only" leads with 20 points
(or something like that, I don't know the list by heart).
> I see this all the time on ICC. Crafty stays in the top 10 or so at blitz all
> the time. It's had a peak ovf 2850, but it's nearly impossible to stay there,
> because it is playing opponents rated at 2200 also. And one loss = -32 points,
> while one win = 0. Against a real master, it would have to win *every* game
> to stick up at that lofty pinnacle. It won't happen. On occasion, a GM will
> go crazy and play it 100 games, and lose 3/4 of them, and drive it up to such
> a rating, but when he goes away, the normal rating pool simply drags it right
> back down because there's not many rated over 2600, that play actively on ICC.
>
It simply isn't possible to lead with a big difference to the 2nd place.
Even Kasparow "only" leads with 20 points (or something like that, I
don't know the list by heart). This is an effect of the ELO formula.
> There's far move 2300-2400 IM's, but if a 2800 player beats a 2300 player he
> gets nothing. He loses a lot when he draws or loses however.
Yes, that's right! See above!
Best wishes
Andreas Mader
Komputer Korner
Torstein Hall wrote:
>
> Would you put Rebel 8 up against an IM in such a match and calculate its
> rating next. I guess it would end up around in the area of 2200-2300.
> Anyway it would have been very interesting to see such a match.
>
> --
> Torstein Hall
> tors...@eunet.no
> http://login.eunet.no/torshall/sjakk.html
>
I wonder why there has never been a 6 game match between Rebel 8 and an
IM? Maybe the the upcoming Hiarcs 6 should do it as I predict a win
for the Hiarcs 6 program/machine against any IM.
--
Komputer Korner
The inkompetent komputer.
A.Mader
Ed Schroder wrote:
>
> From: hy...@crafty.cis.uab.edu (Robert Hyatt)
>
> In my previous posting I said the following:
>
> I like to know for example your opinions about breakthroughs. I am
> quite happy with a 25 ELO gain in one year on the same hardware and I
> call 50 ELO due to software improvements only simply a breakthrough.
>
> Going from 1600 to 1650 is nothing compared to 2410 to 2460.
> Going from 2410 to 2460 is a breakthrough.
> is the happiest man (woman) in the world!
>
>
> Maybe he is right, but to my best knowledge it's much more difficult
> to make an ELO jump from 2410 to 2460 than from 1600 to 1650.
>
> I know it is in computer chess, but to me this also counts for humans.
>
>
> - Ed Schroder -
Can one of the Swedish "Gurus" please tell me (us), if they use a fixed
or a variable factor? With a fixed factor the ratings will become higher
and higher and maybe we will have a SSDF rating of 3000 in the year
2010. With a variable factor the differences will become less and less
and it would not be possible to measure minor improvements at all. The
third possibility is to lower the niveau of the SSDF list form time to
time, but I don't like this, too....
Best wishes
Andreas Mader
Robert Hyatt
Komputer Korner (kor...@netcom.ca) wrote:
: Torstein Hall wrote:
: >
: The inkompetent komputer.
You just don't know the right IM. There are plenty that play like GM's, but
just haven't gotten that third norm yet. One that comes to mind that would
be a difficult challenge would be IM Danny Kopec (don't misinterpret this, I
am not saying Danny is one norm from a GM title). There are others like Brian
Hartman.
Robert Hyatt
A.Mader (andrea...@siemens.at) wrote:
: Robert Hyatt wrote:
: >
: > Ed Schroder (rebc...@xs4all.nl) wrote:
: > : From: shi...@nntp.best.com (Joe McCaughan)
: >
: > : : Hmmmm....
: >
: > : : Things seem to be getting a little quiet out there...
: >
: > : : Ed? Anything to say?
: >
: > : Hi,
: >
: > : In principal I have said everything what is on my mind and I don't like
: > : repetitions, they as you know only end in a draw... :)
: >
: > : But what I really have missed is that sofar it was a discussion between
: > : Bob and me only. Where is the rest of you?
: >
: > : Bob and I are just 2 persons with only 2 brains and 2 opinions and
: > : thoughts. There are hundreds of people here all with opinions about
: > : this subject.
: >
: > : So write them, I like to hear them...
: >
: > : quite happy with a 25 ELO gain in one year on the same hardware and I
: > : call 50 ELO due to software improvements only simply a breakthrough.
: >
: > : Going from 1600 to 1650 is nothing compared to 2410 to 2460.
: > There you are wrong. The Elo formula computes a spread of 50 between two
: > 1500 players exactly the same as a spread of 50 between to 2600 players. It
: > simply gives the expected # of wins for each in a match.
: >
: The ELO formula is simple but not that simple. There are two ways of
: computing ELO ratings. One way is to use a variable factor that becomes
: less and less the higher the rating gets. If the Swedes use this formula
: then Ed is right. The second way is to use a fixed factor. If the Swedes
: use this formula then Ed is wrong.
: Of course you are right that you cannot compare SSDF ratings with FIDE
: ratings, but that's not the point. If a program is able to win >60 out
: of 100 games against different strong opponents, I would call this a
: breakthrough.
:
: > : Going from 1600 to 1650 is no breakthrough.
: > : Going from 2410 to 2460 is a breakthrough.
: >
: > If you were on the very upper end of the *real* Elo scale, like, say a Kasparov,
: > you are correct, because a 50 point gain is difficult. But only because you don't
: > get anything for beating players 400 points below you. And if you are on the top,
: > you are only playing players below you. And if you are Kasparov, most are 2-300
: > points below you. Even one loss costs you 32 points. it might take 32 wins to
: > pick that back up and break even.
: When you have 2800 ELO and you use a variable factor when computing ELO,
: it is very hard to win ELO points but it is also very hard to lose ELO
: points. Draws only have a minor effect , the difference has to be very
: big to influence the rating. When you play agains opponents 300 point
: below you, you probably will make 85 out of 100. I think this is true
: with Kasparow. :-)
Yes, but all the rating schemes I see used, 85 out of 100 would drop you
like a rock. for 85 of those games *no* points, for 15, you lose 32 per
whack until you get within 400 of the person beating you. The only way
to go is down...
I haven't seen a formula that would let a very high rated player lose few
points to a low-rated player, unless you toss in the "Glicko" rating method,
although it still has the upper bound problem, at least the implementations
I have seen...
: >
: > *But*, on the SSDF, this isn't true. Look how the programs are bunched up on
: > the top, with over a dozen within 100 points of each other. No one is way out
: > front, and, as a result, they don't suffer from the "compression" problem as
: > above.
: >
: This is true, but according to the ELO formula it should not make much
: (any?) difference. It simply isn't possible to lead with a big
: difference to the 2nd place. Even Kasparow "only" leads with 20 points
: (or something like that, I don't know the list by heart).
RIght, but in a typical tournament, he's commonly 200 points higher than the
average rating for the next 20-30 players. That's a pretty good handicap. Of
course, his performance justifies his rating, and to keep it there against the
downward pressure of losing a game shows just how strong he really is..
: > I see this all the time on ICC. Crafty stays in the top 10 or so at blitz all
: > the time. It's had a peak ovf 2850, but it's nearly impossible to stay there,
: > because it is playing opponents rated at 2200 also. And one loss = -32 points,
: > while one win = 0. Against a real master, it would have to win *every* game
: > to stick up at that lofty pinnacle. It won't happen. On occasion, a GM will
: > go crazy and play it 100 games, and lose 3/4 of them, and drive it up to such
: > a rating, but when he goes away, the normal rating pool simply drags it right
: > back down because there's not many rated over 2600, that play actively on ICC.
: >
: It simply isn't possible to lead with a big difference to the 2nd place.
: Even Kasparow "only" leads with 20 points (or something like that, I
: don't know the list by heart). This is an effect of the ELO formula.
right, here we agree. Actually it is "possible"... but not very "probable"...
: > There's far move 2300-2400 IM's, but if a 2800 player beats a 2300 player he
eric fitch
Well, Joel Benjamin was certainly hired last time to work on the DB
opening book. Don't know if he still is. I aslo recall that back in the
Deep Thought days, Karpov was brought in as a trainer for a while. Think
there's a good chance that they are still doing this...
Tord Kallqvist Romstad
mclane (mcl...@prima.ruhr.de) wrote:
: What shall * I * say ?? Whenever I say something I am attacked
: BECAUSE I write.
: In your thread many of your sentences could have been out of MY mouth.
: So I read anything you wrote with great amusement and thought:
: Thay always try to damage Chris reputation when they have no further
: arguments/points against him.
: I think: they always could say: Oh - this mclane is a nobody, he
: counts nothing. He is not a programmer. He is maybe crazy.
Thorsten, I want you to know that I regard you and Chris as two of the
most valuable contributors to this group.
My belief is that Deep Blue would win easily against all current
microcomputer programs. It seems that most people in the "anti-DB camp"
base their beliefs only on the worst games DT/DB ever played (like the
DT-Fritz game from Hong Kong, or the last game of the Kasparov match).
This seems unfair to me. If you choose the worst few games Rebel, Hiarcs,
Genius and CST ever played, and based your opinions on these games only,
all those programs would seem very weak. We all play awful games occasionally,
some of us more often than others. Even Kasparov has made moves most of
us would be ashamed of.
Tord
Chris Whittington
--
http://www.demon.co.uk/oxford-soft
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
<5en44s$8...@juniper.cis.uab.edu>...
> Ed Schroder (rebc...@xs4all.nl) wrote:
> : From: shi...@nntp.best.com (Joe McCaughan)
>
> : : Hmmmm....
>
> : : Things seem to be getting a little quiet out there...
>
> : : Ed? Anything to say?
>
>
> : Hi,
>
> : In principal I have said everything what is on my mind and I don't like
> : repetitions, they as you know only end in a draw... :)
>
> : But what I really have missed is that sofar it was a discussion between
> : Bob and me only. Where is the rest of you?
>
> : Bob and I are just 2 persons with only 2 brains and 2 opinions and
> : thoughts. There are hundreds of people here all with opinions about
> : this subject.
>
> : So write them, I like to hear them...
>
> : I like to know for example your opinions about breakthroughs. I am
> : quite happy with a 25 ELO gain in one year on the same hardware and I
> : call 50 ELO due to software improvements only simply a breakthrough.
>
> : Going from 1600 to 1650 is nothing compared to 2410 to 2460.
>
> There you are wrong. The Elo formula computes a spread of 50 between two
> 1500 players exactly the same as a spread of 50 between to 2600 players.
It
> simply gives the expected # of wins for each in a match.
Total bollocks.
Ed is absolutely right. Elo points at the low end of the scale are very
easy to earn. At the top end, very difficult.
This is nothing to do with statistics (as you suggest) but with the
acquisition of chess knowledge.
What is Karpov's grade right now ? 2700 or so ? How difficult would it be
for him to get a grade of 2750 ? How difficult would it be for him to
acquire new chess knowledge and insights represented by these 2700 to 2750
elos ? Bloody difficult.
What was Karpov's grade aged 6 ? 1300 or so ? How difficult was it for him
to get a grade of 1350 ? Bloody easy. He just learnt about the bishop pair
or something.
Bits of new knowledge at the lower end are worth lots. Its game winning
knowledge. Bits of knowledge at the top end are worth only a little. You
needs lots of bits to get a game winning advantage.
Statistic-materialistic-mechanistic twaddle may look good, it may sound
good, it may do you good, but it remains twaddle.
Chris Whittington
>
> : Going from 1600 to 1650 is no breakthrough.
> : Going from 2410 to 2460 is a breakthrough.
>
> If you were on the very upper end of the *real* Elo scale, like, say a
Kasparov,
> you are correct, because a 50 point gain is difficult. But only because
you don't
> get anything for beating players 400 points below you. And if you are on
the top,
> you are only playing players below you. And if you are Kasparov, most
are 2-300
> points below you. Even one loss costs you 32 points. it might take 32
wins to
> pick that back up and break even.
>
> *But*, on the SSDF, this isn't true. Look how the programs are bunched
up on
> the top, with over a dozen within 100 points of each other. No one is
way out
> front, and, as a result, they don't suffer from the "compression" problem
as
> above.
>
> I see this all the time on ICC. Crafty stays in the top 10 or so at
blitz all
> the time. It's had a peak ovf 2850, but it's nearly impossible to stay
there,
> because it is playing opponents rated at 2200 also. And one loss = -32
points,
> while one win = 0. Against a real master, it would have to win *every*
game
> to stick up at that lofty pinnacle. It won't happen. On occasion, a GM
will
> go crazy and play it 100 games, and lose 3/4 of them, and drive it up to
such
> a rating, but when he goes away, the normal rating pool simply drags it
right
> back down because there's not many rated over 2600, that play actively on
ICC.
>
> There's far move 2300-2400 IM's, but if a 2800 player beats a 2300 player
he
> gets nothing. He loses a lot when he draws or loses however.
>
> The bottom line is as you distance yourself from the pack, rating changes
are
> much more hard to produce as you suggest. But on the SSDF, no one is
distanced
> +200 from the rest. or even +100 from the rest. The top of the rating
pool is
> quite compressed, making the "spread" calculation very accurate, but also
+50
> at the top is just as significant and difficult to obtain as +50 in the
middle.
>
> : Just ask any IM, if he (she) makes such an ELO jump the IM in question
> : is the happiest man (woman) in the world!
>
> But not particularly more happy than a regular master that jumps +50, or
> an expert that jumps +50... I agree that +50 represents some work on the
> IM's part, to be sure. But notice also that he's playing in a big pool
of
> players with reliable ratings based on the Elo formula. The SSDF rating
pool
> is much smaller, with rating that are distorted by incest. The
difference
> between two programs gives an indication of the expected win/loss ratio,
but
> the actual rating doesn't mean a lot, other than when compared to another
program
> for the spread. Elo's main "claim to fame" is the "spread" between two
players,
> not the "rating" of a single player. You could add +300 to everyone's
rating,
> and the rating pool would still be stable at that new rating level for
everyone.
>
>
> : I have read Bob's comments on that and besides the fact I find his
> : comments very unfounded I like to hear yours.
>
> : Are Bob and I the only ones here in RGCC?
>
> : I hope not...
>
> : - Ed Schroder -
>
>
>
Robert Hyatt
Chris Whittington (chr...@demon.co.uk) wrote:
: --
: http://www.demon.co.uk/oxford-soft
: Total bollocks.
If you look at what I wrote, I presented two points here. 1. Mathematically,
points at the upper end are hard to earn, if you are at the upper end of all
players, because you only gain points if you play against someone close to you
(within roughly 400 points depending on the formula.) If everyone is close to
you, as in the case of the SSDF, then it's not so difficult to gain points,
*from a mathematical point of view*. 2. gaining points at the low end is
easier from a programming point of view, because you probably have almost no
knowledge and very poor search down there. Simple additions close gaping holes
and boost your rating. At the upper end, all the "easy" things are done, and
therefore it's hard work to find something to add that helps.
All that said, A jump in 50 points by Kasparov would be quite a jump, because he
is so far above everyone that it would take a *remarkable* improvement in his
playing skill to accomplish this. *much more* improvement, IMO, than what is
required to gain 50 points on the SSDF, because the top programs are bunched up
within 50-100 points of each other.
There's plenty of evidence that shows that small changes produce exaggerated
rating changes when self-play is used. That was my point with the upper end of
the SSDF. A rating spread of "R" between two programs simply indicates the
winning ratio between the two programs, and nothing else. Not that against
humans they'd both do about the same... rather that against each other, one
has something that the other one hasn't... and quite often a single such piece
of knowledge is enough to produce a big rating change (big=30-50 points.)
Humans are much different than computers. Since most computers are pretty
"constant" in what they do, a minor change can produce significant gains. When,
in reality, if they were both tossed into a 100-round tournament with a bunch
of strong humans, their ratings might end up identical. That's the point here,
the rating spread indicates win/loss probabilities. I simply think that there's
a significant difference between when two program versions do against each other,
and what they'd do against humans. Were this not true, Crafty would be well over
3000 now. Several changes I've made have resulted in lopsided margins of wins
over the older version, yet made very nominal changes on ICC against the group
of humans it regularly plays against (GM's and IM's mostly). Which of the following
would you trust more:
11.1 beats 10.18 at a ratio of 3:1 (+200 due to a few gross bugs removed.)
11.2 beats 11.1 at a ratio of 1.5:1.
11.3 beats 11.2 at a ratio of 1.2:1.
and then 11.3 beats the human population at a rate of 3:1, while
10.19 beat humans at a rate of 2.95:1. You can obviously draw two
different conclusions from the above data, which is pretty much real
results for Crafty. Yet it's actual skill level (ICC rating) has not
changed a lot since Jakarta. It peaked at 2850, more commonly stays
around 2700, but was back at 2800 last week. And yet I know beyond a
shadow of a doubt that 11.17 outplays 10.18 (Jakarta) by a significant
amount on my machine, mostly due to bug fixes and much faster code. But
it's still playing at about the same performance level against upper level
humans. Difficult to interpret. I tend to trust the latter for how well
it plays, and trust the former to understand how much programming progress
I have made in fixing things. However the latter is what counts... because
that's the group (humans) that "most" of us aspire to beat. I'd suspect that
there are some that only want to top the SSDF list. I'd like to top the FIDE
list. :)
: This is nothing to do with statistics (as you suggest) but with the
: acquisition of chess knowledge.
Note I gave two cases. In one case, you are right. but mathematically,
you are wrong, because if you reach a point +400 over the next best
player, your rating will never climb. No matter how much chess knowledge
you gain..
: What is Karpov's grade right now ? 2700 or so ? How difficult would it be
: for him to get a grade of 2750 ? How difficult would it be for him to
: acquire new chess knowledge and insights represented by these 2700 to 2750
: elos ? Bloody difficult.
: What was Karpov's grade aged 6 ? 1300 or so ? How difficult was it for him
: to get a grade of 1350 ? Bloody easy. He just learnt about the bishop pair
: or something.
I agree, chess-wise...
: Bits of new knowledge at the lower end are worth lots. Its game winning
: Chris Whittington
: >
: >
: >
A.Mader
Robert Hyatt wrote:
>
> : Andreas Mader wrote:
> :
> : When you have 2800 ELO and you use a variable factor when computing ELO,
> : it is very hard to win ELO points but it is also very hard to lose ELO
> : points. Draws only have a minor effect , the difference has to be very
> : big to influence the rating. When you play agains opponents 300 point
> : below you, you probably will make 85 out of 100. I think this is true
> : with Kasparow. :-)
>
> Yes, but all the rating schemes I see used, 85 out of 100 would drop you
> like a rock. for 85 of those games *no* points, for 15, you lose 32 per
> whack until you get within 400 of the person beating you. The only way
> to go is down...
>
Due to the ELO formula you get really *nothing* for a win if your
opponent is rated less than 735 points below you, because then you have
a 100percent probability to win.
I agree with you in many ways, but I see a little inconsistency in your
arguments:
The ELO formula tends to "put the players together". It is very hard to
be a few 100 points ahead of the crowd and its also "hard" to be a few
100 points below the others. So isn't it a "breakthrough" when someone
(speak Rebel 8) manages to be ahead nearly 100 points (see SSDF list
from last year!)??
Best wishes
Andreas Mader
Robert Hyatt
A.Mader (andrea...@siemens.at) wrote:
: I agree with you in many ways, but I see a little inconsistency in your
: arguments:
: The ELO formula tends to "put the players together". It is very hard to
: be a few 100 points ahead of the crowd and its also "hard" to be a few
: 100 points below the others. So isn't it a "breakthrough" when someone
: (speak Rebel 8) manages to be ahead nearly 100 points (see SSDF list
: from last year!)??
I think where we are in disagreement is around the following point:
minor changes in a program often produce exaggerated win/loss results
against an unchanged version of the same program. If you use this number
as the basis for your +100, it can be badly misleading. I posted some
psuedo-numbers from the past 6 months of crafty development, where
playing itself would suggest a 200 point rating improvement, but playing
against players on ICC shows very modest, if any, improvement.
I believe that Hsu posted here, in one of his rare posts, that the
program that will play kasparov this year probably will beat the program
that played kasparov last year by a 3:1 or 4:1 margin. Which is well
beyond +200 points. Do we think it will now crush Kasparov? I don't.
I simply think that a speed boost, plus evaluation and extension tuning
could produce this sort of lop-sided margin. But I'd bet the results
(rating wise) against players like Kasparov will be much less pronounced.
We'll see a prime example in a couple of months, that will make most of
this moot. When it's over, I'll ask him about the winning percentage of
the current program over last year's version. Not sure how he'll extrapolate,
maybe on some test positions or something. But it will be interesting.
Now, with that said, the SSDF is a fairly close model of the self-test
idea. Different programs, but they could be the same program with different
eval or search ideas. They are all pretty close to each other based on
the numbers of games played. I simply think that rating changes when you
consider computer vs computer matches is often exaggerated over rating
changes in computer vs human matches. This is partially explained, of
course, because the two rating pools are different. But it's more likely
explained by the inherent differences between computers and humans and
how they evolve over a series of games (humans do of course, computers,
even the learning ones don't or *barely* do...)
Otherwise I think we agree...
mclane
Ed Schroder <rebc...@xs4all.nl> wrote:
>Bob you are totally wrong here.
Yes. Maybe the games on ICC are NOT serious.
The people are NOT sitting FACE TO FACE .
In Den Haag there is a nice atmossphere. And - I guess the GMs like
it. But they try hard not to lose.
These games are not fun. Of course the whole event is a pleasure.
But playing in a tournament face to face with the GMs is not a joke at
all.
>See also the posting of Enrique one of the best postings I ever read in
>RGCC concerning this subject because it's very informative about the
>results and ratings of the micro's against strong human players (mainly
>GM's) I am curious on your reply of this posting, so many facts.
>Till now the victory list of the micro's is much more impressive than
>the list from DB. Wish they would play more...
>- Ed Schroder -
Torstein Hall
Komputer Korner <kor...@netcom.ca> skrev i artikkelen
<33113F...@netcom.ca>...
> Torstein Hall wrote:
> >
>
> > Would you put Rebel 8 up against an IM in such a match and calculate
its
> > rating next. I guess it would end up around in the area of 2200-2300.
> > Anyway it would have been very interesting to see such a match.
> >
> > --
> > Torstein Hall
> > tors...@eunet.no
> > http://login.eunet.no/torshall/sjakk.html
> >
>
>
> I wonder why there has never been a 6 game match between Rebel 8 and an
> IM? Maybe the the upcoming Hiarcs 6 should do it as I predict a win
> for the Hiarcs 6 program/machine against any IM.
> --
> Komputer Korner
>
> The inkompetent komputer.
Thats someting I belive when I see it, not a second before!
Torstein Hall
Torstein Hall
Robert Hyatt <hy...@crafty.cis.uab.edu> skrev i artikkelen
<5es5gv$f...@juniper.cis.uab.edu>...
You simply make a cut off at 350 elo points difference as I have stated
earlier on.
Then there is no problem playing weaker players, or at least you are not
supposed to score 110% only 90% or something.
Komputer Korner
Ed Schroder wrote:
>
snipped
>
> : But know you are number one and anybody is haunting you and rebel.
> : Thats the nice situation Richard always had. How do you feel now?
>
> Hunted... :)
>
> I feel the breath of Mark... :)
>
> - Ed Schroder -
Ed,
I haven't asked Mark about this but I get the feeling that to
achieve the gains that he has made in Hiarcs 6, there must have been
some GM helping in the knowledge department. It jusy plays too
human like and too strong not to have had a GM influence. Also look at
Crafty, Roman is always helping Bob out with Crafty. You must have some
Dutch GM helping out as well. To get that last 30% of knowledge, I
don't see how you can do it without a GM's help.
brucemo
Rolf Tueschen wrote:
> Now think of this: what would happen if Kasparov won 100% in each and every
> match vs computer? Do you think companies would still invest a whole lot of
> money? Would you further think we as the interested public would appreciate
> such *stupid* farces?
>
> If Kasparov wanted it he could *kill* DB 100%. Ok let's skip fatigue or similar
> stuff. Even in the middle of the night in a hotel lobby Kasparov would crush
> the monster machine at tournament time controls. In march Kasparov has more
> difficulties to *invent* some suspense that stirs the chessworld's interest.
Add Kasparov to the list of people you've smeared in here on zero evidence.
bruce
Jean-Peter Fendrich
Well, you believe in this and I am still not convinced believing that
the
effect can't be that huge.
>
> : There are a number of games between humans and programs and most of them
> : are supporting the SSDF level of ratings. Often are the performance
> : ratings of the chess programs better than the SSDF rating as the case
> : from the AEGON tournament. This shouldn't be too surprisingly because
> : the
> : level of the SSDF list is adjusted according to results from tounament
> : games or games from tournament conditions with humans.
>
> No it isn't... It was adjusted in 1991, but not since. That was apparently
> to take into account about 300 human/computer games. But that was 1991. This
> is 6 years and lots of hardware and software improvements later...
The list is adjusted in this way even if it was 1991, that shouldn't be
of
any problem assuming that 2400 humman ELO back in 1991 still means 2400
human ELO. I haven't read or heard that this isn't the case.
>
> : There is one *big but* however...
>
> : I think most people overlook the fact that humans perform very different
> : against computers the first times we play them compared to later on when
> : we really learned how the beasts are playing. Our ability to adapt makes
> : a real difference here. Learning abilities in the programs of today,
> : doesn't change this in any significant way.
> : There are two levels of adaption:
> : - Adapting in general to the way most programs are playing. Now I
> : believe that strong chess players have done their homework on this.
> : - Adapting to a very specific chess program. Most of us knows about
> : the effects here.
> : I am quite convinced of that the SSDF level is a fairly good estimate
> : of the performance of the first 5-10 games with the same human. After
> : that the performance will drop considerably. Maybe your 100+ ELO
> : points is the difference. The point is that most chess programs will
> : not meet the same GM or IM so many times. Before that, we have new
> : versions and new models and we are back to the 'zero number of games'
> : again.
>
> Here we are in agreement. The computer will give anyone a "new look" for a
> while, but the new wears off in a hurry... That's exactly why I think that should
> any of the current micro programs (Crafty included, BTW) play the FIDE circuit,
> the ratings would be a lot lower than most would guess. I still think the right
> number is maybe 2300. *maybe*...
it goes down, I believe :)
> : The problem here is, as said many times before, how we should define
> : the strength of a program. Is it the performance level during it's life
> : time or is it the theoretical performance after an infinite number
> : of games against all possible opponents?
>
> : Not the same thing, believe me... :-)
>
> Not at all. However, we certainly don't want to define this rating after a
> few games, or even a few dozen. I suspect that should Crafty play at next
> year's Pan Am, and everyone knows beforehand, the result will be a bit
> different from this year. Now they know they'd better prepare. This year it
> was a "lark"...
Didn't you miss my point, answering like this? I never claimed that we
should use a few games only...
The lack of adaptability by the programs introduces a lot of strange
effects that we are just scratching the surface of here. It's probably
possible to write a book on how to beat some *specific* chess program.
It's probably possible to train every player with a rating of 1900 and
above to beat the same program. This will not happen but it's possible
to do. It's not possible to train the same peopple to beat a human with
2300 or 2400 or so, at least not in the same 'automatic' manner as
compared to beating the chess program.
Wich rating is the right in this situation? 1900, 2300, 2400?
The performance rating is what the programs have earned sofar. The
theoretical lower limit is out of my (and your) knowledge.
The performance rating during it's lifetime however is maybe possible to
have some vague feeling of and even predict.
>
> And when they prepare, things will change..
I hope so and that's competition!
---
J-P Fendrich
Robert Hyatt
Komputer Korner (kor...@netcom.ca) wrote:
: The inkompetent komputer.
Only one thing wrong... it's not "that last 30% of knowledge"... It's more
likely "that last 95% of knowledge"... :) Ask any IM/GM. And BTW, while I
get lots of GM input (Roman, Lombardy, plus others I won't name because they've
asked me not to bring them into computer chess discussions) it turns out that most
of it is very abstract. Don't think that a GM is going to walk up, peer over
your shoulder, and make you better. How'd you like to be trained to do brain
surgery like that? It's difficult. Particularly when often a GM can't even
explain why he makes a move, "it just feels like the right move." Now if
someone would only add this to C:
if (FeelsRight(move)) Play(move);
Then I might get closer to creating an electroninc GM. As it is, I'll have
to be satisfied with a near IM player and keep working to improve. :)
Mark Rawlings
brucemo <bru...@nwlink.com> wrote:
>Mark Rawlings wrote:
>>
>> Here's an interesting quote from the 4th Quarter 1990 Computer Chess
>> Reports ("Brains vs. Brawn," page 6):
>>
>> "... its [Deep Thought's] authors admit that it is probably 300 points
>> weaker than the best micro programs given equal hardware, but the
>> great depth of search on its special purpose hardware obscures its
>> ignorance. If this 300 point figure is correct, then a program as
>> good as the latest Mephisto running on Deep Thought hardware would be
>> in the class with Kasparov and Karpov, but it may be that the Mephisto
>> program is incompatible with such hardware."
>Who was the author of this article?
>bruce
The senior editor of that issue was Larry Kaufman. I assume he wrote
the article because it wasn't attributed to anyone else.
Mark
Martin Borriss
In article <33113F...@netcom.ca>,
Komputers Korner <kor...@netcom.ca> writes:
>Torstein Hall wrote:
>>
>
>> Would you put Rebel 8 up against an IM in such a match and calculate its
>> rating next. I guess it would end up around in the area of 2200-2300.
>> Anyway it would have been very interesting to see such a match.
>>
>
>IM? Maybe the the upcoming Hiarcs 6 should do it as I predict a win
>for the Hiarcs 6 program/machine against any IM.
You can predict what you want; but this is ridiculous.
>--
>Komputer Korner
>
>The inkompetent komputer.
Yes.
Martin
Feng-Hsiung Hsu
In article <5etm43$i...@boursy.news.erols.com>,
Mark Rawlings <raw...@erols.com> wrote:
>brucemo <bru...@nwlink.com> wrote:
>
>>Mark Rawlings wrote:
>>>
>>> Here's an interesting quote from the 4th Quarter 1990 Computer Chess
>>> Reports ("Brains vs. Brawn," page 6):
>>>
>>> "... its [Deep Thought's] authors admit that it is probably 300 points
>>> weaker than the best micro programs given equal hardware, but the
>>Who was the author of this article?
>>bruce
>The senior editor of that issue was Larry Kaufman. I assume he wrote
>the article because it wasn't attributed to anyone else.
Are you sure that Larry wrote it? He normally checked his sources. The
above statement was unfounded, and we never did say it. The best micro
programs would not run on DT1's hardware. There was some compromise made
in DT1 hardware design, so it was not as good as it could be on a per node
basis. But the limitation came directly from DT1 hardware, and if the
micros want to run on "equal hardware", they would have to take the
same limitation, which includes loss of search efficiency from the
parallel search and the limited hash table access, a less flexible move
generator which probably makes micros' search algorithm useless, and so on.
In the book, "Computers, Chess, and Cognition", we did say, with lots of
uncertainty, that perhaps there was a "100-200" rating points increase to
be gained, but that required going beyond what the micros were doing. The
estimate is probably very far on the high side. A "100-200" rating points
gain at DT1's level, which potentially would have made DT1 a Super
Grandmaster, is extremely hard. A 300 points gain as alleged in the quote
would have made DT1 a challenger to the world throne. Not bloody likely.
It is ancient history, anyway. Larry found out the hard way that DT1
was not just the node rate a few years latter, although the lesson came from
DT2, which did have a better hardware evaluation and used a better search
algorithm.
mclane
Ed Schroder <rebc...@xs4all.nl> wrote:
>: I then think: they cannot do the same GAME with you. It will not work,
>: because you are a historical person, you normally do not overreact,
>: you have success and you are a world-champion
>: the only pity for them is:
>: that you are often Chris' and my opinon.
>True Chris and I have much in common about "the way to go" in computer
>chess concerning playing strength, but there are other opinions too and
>I like to listen to them and learn from them.
Yes, I still learn much about computer-chess. Thats the reason it is
still not boring .... unbelievable. Also the older I get, the better
the old days were tuned in my memory. And the reality gets more and
more fascinating in computer-chess. The decision was right. I hope
this will never change.
Can you imagine, sometimes I get one of my old chess-computers, put
them on and play some games agaist them, maybe vs. the latest
PC-programs, just to see the difference. I really like these old
machines out of wood.
>: E.g. in this case.
>: So - these are my thoughts.
>: I have tried to fight many threads again, my
>> Bob and I are just 2 persons with only 2 brains and 2 opinions and
>> thoughts. There are hundreds of people here all with opinions about
>> this subject.
>: Right, we should ask intelligent guys like Moritz Berger, Alexander
>: Fuchs, Dirk Frickenschmidt, Stefan-Meyer-Kahlen, Enrique Irazoqui,
>: Peter Schreiner, Vincent, Peter Gillgasch, and and and the rest,
>Yes.
>[ snip ]
>: But Ed - what do you want ?
>: In my opinon you want a fight against Deep Blue or another big iron.
>: I am not sure if they will give you the chance, because they could be
>: afraid a little. Your program is at the peak in the moment.
>Oh no!
>Playing against "the beast" would be nice but it is (was) never my
>intention to challenge them. This is silly, they can't afford the risk!
>Suppose, just suppose they lose... IBM is right, a Deep Blue versus PC
>match will never happen.
>I can imagine that my discussion with Bob sounded that I was looking for
>such a match but I was just jumping in the discussion because in my
>opinion Bob overreacted about the strength of Deep Blue.
I have come to the same conclusions/opinons. Bob - now we have you !
One point for us !! :-)
>So I chose
>sides for the poor Pc guys in the underdog position fighting against
>the NPS dream... :)
Right. Only PC programs are able to get Excalibur out of the anvil.
>To me with a lot of good statements from both sides.
>: So they will say: we consider about, have to organize and other stuff
>: like that.
>: In the end, it is too late for Rebel8 to show it's strength.
>: I am looking forward for a fight. I think Rebel is still the strongest
>: program. Than Mchess and than maybe Hiarcs.
>: You have made a big jump with your rebel.
>: It was easy to find out. I did not need much time to find out.
>: So much progress was it.
>: But know you are number one and anybody is haunting you and rebel.
>: Thats the nice situation Richard always had. How do you feel now?
>Hunted... :)
>I feel the breath of Mark... :)
Yes, He and Marty give their best to get you. Where the hell is
Richard ! Richard - we can't hear you breathing !!!
Come on Richard - a new Genius with different playing style. Less
boring and no asymmetrie. Possible ?!
>- Ed Schroder -
mclane
tor...@ifi.uio.no (Tord Kallqvist Romstad) wrote:
>mclane (mcl...@prima.ruhr.de) wrote:
>: What shall * I * say ?? Whenever I say something I am attacked
>: BECAUSE I write.
>: In your thread many of your sentences could have been out of MY mouth.
>: So I read anything you wrote with great amusement and thought:
>: Thay always try to damage Chris reputation when they have no further
>: arguments/points against him.
>: I think: they always could say: Oh - this mclane is a nobody, he
>: counts nothing. He is not a programmer. He is maybe crazy.
>Thorsten, I want you to know that I regard you and Chris as two of the
>most valuable contributors to this group.
Thats really nice and will give me some energy in this dark time
(i feel always very alone when Chris is not sending me NICE versions.
How can I stand the time without a CSTal sacrificing the hell out of
the opponents ???!!! ) Thanks.
>My belief is that Deep Blue would win easily against all current
>microcomputer programs. It seems that most people in the "anti-DB camp"
>base their beliefs only on the worst games DT/DB ever played (like the
>DT-Fritz game from Hong Kong, or the last game of the Kasparov match).
In my opinion we have also many prejudices.
>This seems unfair to me. If you choose the worst few games Rebel, Hiarcs,
>Genius and CST ever played, and based your opinions on these games only,
>all those programs would seem very weak. We all play awful games occasionally,
>some of us more often than others. Even Kasparov has made moves most of
>us would be ashamed of.
Thats true. You are right here. Maybe we need DB to have an enemy to
fight against. I once told Richard how important GENIUS3 was for us.
We really needed Genius3 to kill it. Only to have a strong enemy we
could focus on. To give our best to win against.
We all need borders to set us and our friends free. Without borders or
laws or enemies we could not live.
>Tord
mclane
brucemo <bru...@nwlink.com> wrote:
>Mark Rawlings wrote:
>>
>> Here's an interesting quote from the 4th Quarter 1990 Computer Chess
>> Reports ("Brains vs. Brawn," page 6):
>>
>> "... its [Deep Thought's] authors admit that it is probably 300 points
>> weaker than the best micro programs given equal hardware, but the
>> great depth of search on its special purpose hardware obscures its
>> ignorance. If this 300 point figure is correct, then a program as
>> good as the latest Mephisto running on Deep Thought hardware would be
>> in the class with Kasparov and Karpov, but it may be that the Mephisto
>> program is incompatible with such hardware."
Ha ! Thats what I needed to make my argumentation FIT !!
Thanks Mark. We have to find out who wrote it. Maybe it was me, using
an alias ?! No - just a personal joke for Andreas Mader..... :-)
Jean-Peter Fendrich
A.Mader wrote:
> Can one of the Swedish "Gurus" please tell me (us), if they use a fixed
> or a variable factor? With a fixed factor the ratings will become higher
> and higher and maybe we will have a SSDF rating of 3000 in the year
> 2010. With a variable factor the differences will become less and less
> and it would not be possible to measure minor improvements at all. The
> third possibility is to lower the niveau of the SSDF list form time to
> time, but I don't like this, too....
>
> Best wishes
> Andreas Mader
What factor are you refering to? Is it K in the standard formula?
The standard formula: Rn = R0 + K(W - We)
Rn = new Rating
R0 = current Rating, W = wins (with draw=1/2),
We = expected W according to current ratings
K = the rating point value of one single win
K is set to a high value if you want the most recent performance be of
high value and set to a low value if you want to give more weight to
earlier performance.
This is not neccesary to do when we deal with programs. The assumption
is that they have the same performance during all their life time.
For that reason SSDF gives the same weight to every game whenever it
was played. Dr. ELO produced another formula to be used in this case
wich will utilize the game results better.
--
J-P Fendrich
Rolf Tueschen
brucemo <bru...@nwlink.com> wrote:
>Rolf Tueschen wrote:
>bruce
---------------------------------------------------------------------------------------------------------------------
Buddy, I know what you want so dearly: to censor (extinguish) people who are of
different opinions. You won't succeed, proxy. .)
Rolf Tueschen
hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>Rolf Tueschen (TUESCHEN.MEDIZ...@t-online.de) wrote:
snip
>: If Kasparov wanted it he could *kill* DB 100%. Ok let's skip fatigue or similar
>: stuff. Even in the middle of the night in a hotel lobby Kasparov would crush
>: the monster machine at tournament time controls. In march Kasparov has more
>: difficulties to *invent* some suspense that stirs the chessworld's interest.
>: Possible rematches respectively.
>Boy, you're good. You can now figure out exactly what Kasparov's motives are,
>what my motives are, and just about everyone else (marsland, levy, friedel,
>and so forth.) With this much gullibility, you must own several bridges and
>a few plots of land out in the desert... maybe some Edsel stock... stock in
>the Russian space shuttle...
It' s strange, but I couldn't find Edsel in my companies' books. Should I take
a closer look on this Edsel, Bob?
BTW, why bruce and now you are so upset reading an interesting opinion. Tell me
what are your opinions, Bob. Do you are so upset because some organization is
close to this thrilling rematch? I heard it through the grape vine ...
Did you overlook that my deep sympathies and admiration are on the side of GK?
>: It's true. Someone wrote about sort of non-sportmanship of DB and team. They
>: hide their player, they prepare on a well known human player and they try
>: simply to take maximum advantage out of this surprise setting. If they'll lose
>: then Elo statistics is not bad either because of K's 2900. :)
>: This reminds me of my own famous career when I got these smashing phantasies
>: after learning some book lines I hoped [:)] my opponent didn't know about ...
>: Fom 1900 Elo on however I detected that almost all others of equal strength had
>: also *learned* *my* lines. I therefore quickly learned:
>: Chess is not entirely descripted with the parameter *gambling*.
>:
>: Today I know that even at 2900 one could have much fun in gambling ...
Dave Gomboc
In article <5es544$f...@juniper.cis.uab.edu>,
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote:
#Komputer Korner (kor...@netcom.ca) wrote:
#: Torstein Hall wrote:
#: >
#
#: > Would you put Rebel 8 up against an IM in such a match and calculate its
#: > rating next. I guess it would end up around in the area of 2200-2300.
#: > Anyway it would have been very interesting to see such a match.
#: >
#: > --
#: > Torstein Hall
#: > tors...@eunet.no
#: > http://login.eunet.no/torshall/sjakk.html
#: >
#
#
#: I wonder why there has never been a 6 game match between Rebel 8 and an
#: IM? Maybe the the upcoming Hiarcs 6 should do it as I predict a win
#: for the Hiarcs 6 program/machine against any IM.
#: --
#: Komputer Korner
#
#: The inkompetent komputer.
#
#You just don't know the right IM. There are plenty that play like GM's, but
#just haven't gotten that third norm yet. One that comes to mind that would
#be a difficult challenge would be IM Danny Kopec (don't misinterpret this, I
#am not saying Danny is one norm from a GM title). There are others like Brian
#Hartman.
Speaking of Canadian IMs, what about Yan Teplitsky? CFC 2552.
GM Kevin Spraggett has apparently referred to him as "a GM in IMs clothing".
I played a game against him last Sunday (last round, 1997 Northern
Alberta Open).. I hung on through the first time control but he was
slowly outplaying me all game and won in the second control.
I was pretty happy to even stay alive for a while.. IM Lawrence Day
crushed me in the second round in under two hours!
Fortunately I won my other three games (all sub-2000 players though)
and managed to grab 1st <2200.
Dave Gomboc (CFC 2132)
drgo...@a.stu.athabascau.ca
Tord Kallqvist Romstad
mclane (mcl...@prima.ruhr.de) wrote:
: >: But know you are number one and anybody is haunting you and rebel.
: >: Thats the nice situation Richard always had. How do you feel now?
: >Hunted... :)
: >I feel the breath of Mark... :)
: Yes, He and Marty give their best to get you. Where the hell is
: Richard ! Richard - we can't hear you breathing !!!
Richard will come back, I am sure. Do you remember Mephisto Lyon, Vancouver
and Genius 1? Each new version of Richard's programs showed little or no
improvement compared to the previous version. Then, suddenly, Richard made
a huge leap forward with Genius 2. I am sure the same thing will happen
again.
: Come on Richard - a new Genius with different playing style. Less
: boring and no asymmetrie. Possible ?!
Don't remove the asymmetric search --- reverse it!
Tord
: >- Ed Schroder -
Mader Andreas
Interesting...
Does that mean that theoretically the programs can reach SSDF ELO 3000
or 4000? If this is true then Bob Hyatt is absolutely right with his
statements concerning different rating pools and 'breakthroughs'.
In the formula I am looking at the moment (explanation, how national
Austrian ELO are computed) K is defined as follows:
K = (square (3400 - R0)) / 100000
That means that the theoretical maximum one can reach is 3399 ELO and it
is harder to get additional ELO when you are already rated very good.
And how does all this fit together with the 'different methods' of the
calculation of ELO points? Bob Hyatt wrote that Crafty gets no points
when it wins against a player >400 ELO below. Due to document I am
looking at the difference has to be >735 ELO points ('We' in your
formula becomes 1). Are there different systems with 'wide' and 'small'
ranges? I always thought that 'We' has been defined 'once and forever'?!
Best wishes
Andreas Mader
Robert Hyatt
Mader Andreas (ma...@p6.gud.siemens.co.at) wrote:
I'm not sure what ICC is doing now. When I first started playing there, I
didn't pay attention. Once Crafty hit the 2500 range, it would always drag
back down to 2300-2400. I found that it was the rating difference. +400
seemed to be the 0 win point back then, but Ken Sloan reported that it has
changed as verified by him. 400 points difference gives +1 for a win,
-31 for a loss on ICC at present. My original problem still stands, that
with that kind of loss/gain, and with the fact that ratings are "all over"
the place on ICC (2300GM's, 2700 IM's for example) there's no way for
a -31 +1 point spread to let you go too far...
Jean-Peter Fendrich
> Andreas Mader
This seems to be a peculiar way of computing K, but I'm not
sure of how K is computed in different countries. I think it is usually
set to a fixed value to about 16 (wich in my view would be better).
Isn't FIDE using 10? This indicates a more stable situation within the
pool of very good players.
The problem here is not the rating of programs vs programs. The real
problem is the estimation of human ratings.
Human chess strength is not fixed over time as programs are.
A new human player is often gaining in strength every week or month.
How much should the results from the previous year influence the rating
compared to recently played games? And for results 6 month old?
That's what K is for. If the strengh is fixed over time, the
(fixed) value choosen for K will have no affect at all and that would be
an idela case.
This ideal case is the reality for program vs program games.
That's is why SSDF publishes the margin of errors and
FIDE never will :-)
Bob is perfectly right that different rating pools has different
characteristics. A Swedish rating point is not exactly the same as
an Austrian one and not the same as in USA or FIDE or whatever.
This was the case some years ago and I'm not updated of the situation
today. I would be surprised if it's changed.
The differences are small but they are a fact.
Even if we all used exactly the same value of K, different rating
pools would definitely yet have slightly different levels.
The ordering within each pool is relevant but when comparing between
them problems could occur.
One way is to adjust the level between different pools, like raising
all the Swedish rating with say 50 points to get the same level as in
USA, base on some knowledge of the differences. This is better than
nothing but it's not as good as a lot of games through the barriers
between individuals in different pools.
--
J-P Fendrich
Jean-Peter Fendrich
Robert Hyatt wrote:
> didn't pay attention. Once Crafty hit the 2500 range, it would always drag
> back down to 2300-2400. I found that it was the rating difference. +400
> seemed to be the 0 win point back then, but Ken Sloan reported that it has
> changed as verified by him. 400 points difference gives +1 for a win,
> -31 for a loss on ICC at present. My original problem still stands, that
> with that kind of loss/gain, and with the fact that ratings are "all over"
> the place on ICC (2300GM's, 2700 IM's for example) there's no way for
> a -31 +1 point spread to let you go too far...
This might be the case when you are computing the rating after every
game.
The situation is a slightly better when you compute the rating from a
batch of games, because then you will not sum upp the same error over
and
over again ( a real value of for example 1.3 is rounded to 1.0 for every
game Crafty wins in the first case.)
The situation is much better if you can use all the games played with
the
same Crafty version during it's life time, based on the fact that the
program doesn't vary in strength over time.
This should be possible for every chess program and would result in much
more stable ratings, but I suppose that ICC is not givning you that
service...
--
J-P Fendrich
brucemo
Mader Andreas wrote:
> Interesting...
> Does that mean that theoretically the programs can reach SSDF ELO 3000
> or 4000? If this is true then Bob Hyatt is absolutely right with his
> statements concerning different rating pools and 'breakthroughs'.
There is no minimum or maximum in that standard formula. Negative numbers
are possible.
bruce
Komputer Korner
Dave Gomboc wrote:
>
snipped
>
> Speaking of Canadian IMs, what about Yan Teplitsky? CFC 2552.
> GM Kevin Spraggett has apparently referred to him as "a GM in IMs clothing".
>
> I played a game against him last Sunday (last round, 1997 Northern
> Alberta Open).. I hung on through the first time control but he was
> slowly outplaying me all game and won in the second control.
> I was pretty happy to even stay alive for a while.. IM Lawrence Day
> crushed me in the second round in under two hours!
>
> Fortunately I won my other three games (all sub-2000 players though)
> and managed to grab 1st <2200.
>
> Dave Gomboc (CFC 2132)
> drgo...@a.stu.athabascau.ca
Of course you are right about the difference between strength in
IM's but I still say that Hiarcs 6 on a Pentium Pro will defeat
them all in a 6 game match. The reason is that it is not long
enough for the IM/GM to gain an understanding of its weaknesses.
Of course if the IM practices with it, perhaps he will discover
its weakness but based on the latest training matches that I have
seen, it is difficult to imagine that Hiarcs 6 has weaknesses
except maybe a long term sacrifice against its Kingside. Even so
it is a book learner so after one of these losses, it won't play the
same line again. Hiarcs 6 will definitely take some scalps even at
40/2. We just need the sponsors to pay the IMs for these matches.
Martin Borriss
In article <331674...@netcom.ca>,
Komputer Korner <kor...@netcom.ca> writes:
>
>Of course you are right about the difference between strength in
>IM's but I still say that Hiarcs 6 on a Pentium Pro will defeat
>them all in a 6 game match.
The post you were responding to said that the poster lost to the "weaker" IM
faster than he lost against the "stronger" IM. So what?
>The reason is that it is not long
>enough for the IM/GM to gain an understanding of its weaknesses.
I think I know most weaknesses already without playing a single game or ever
seeing the program.
>Of course if the IM practices with it, perhaps he will discover
>its weakness but based on the latest training matches that I have
>seen, it is difficult to imagine that Hiarcs 6 has weaknesses
Great. The first chess-playing entity without weaknesses. Kasparov has
weaknesses, Hiarcs 6 hasn't.
>except maybe a long term sacrifice against its Kingside. Even so
>it is a book learner so after one of these losses, it won't play the
>same line again. Hiarcs 6 will definitely take some scalps even at
>40/2. We just need the sponsors to pay the IMs for these matches.
Go ahead. If it will be played at some sunny place, I might come too ;)
Martin
>--
>Komputer Korner
>
>The inkompetent komputer.
Chris Whittington
--
http://www.demon.co.uk/oxford-soft
brucemo <bru...@nwlink.com> wrote in article <331672...@nwlink.com>...
You need to be a really good chess player to get a negative elo.
You'ld have to work out ways to force really bad players to win games.
Chris Whittington
>
> bruce
>
Kenneth Sloan
In article <01bc259a$bcdfb4e0$c308...@cpsoft.demon.co.uk>,
Chris Whittington <chr...@demon.co.uk> wrote:
>...
>You need to be a really good chess player to get a negative elo[sic].
>
>You'ld have to work out ways to force really bad players to win games.
Not at all. If it were not for the institution of an artificial floor
(first at 0000, now raised to 0100) the USCF rating pool would have
several (not many, but more than 1) players with negative USCF Ratings.
[aside - since it's well known that USCF = FIDE+500, this would mean
that there are lots of USCF players who would have negative ratings on
the FIDE scale]
The reasons for this are complex - but they do not require intentionally
losing games. The players involved really are that bad...
Personally, I take this as a great compliment to the USCF, since it
demonstrates an unprecedented success at bringing rated tournament play
to the masses.
By the way...let me point out that Elo pegged the current scale at the
2000-2200 level, and remarked in passing that it was desirable to have
"enough ballast" so that ratings did not become negative. Clearly, he
misjudged the extent to which rated chess would penetrate the market,
and the breadth of playing skills now being rated is much more than he
estimated. 2200 is still 2200 (more or less), as Elo intended - but the
mean rating has declined precipitously with the rise of Scholastic (and
pre-Scholastic!) chess in the USA. Some folk persist in pointing to
this drop as evidence of "deflation" - but it's not.
A pop quiz: without consulting any source, please guess the "mean FIDE
rating". Similarly, estimate what it was 5 years ago. Explain, in the
form of a pleasant essay, why this does not indicate that FIDE ratings
are deflating.
Cheerio, pip pip, tallyho
--
Kenneth Sloan sl...@cis.uab.edu
Computer and Information Sciences (205) 934-2213
University of Alabama at Birmingham FAX (205) 934-5473
Birmingham, AL 35294-1170 http://www.cis.uab.edu/info/faculty/sloan/
Jean-Peter Fendrich
Kenneth Sloan wrote:
- snipped -
>
> By the way...let me point out that Elo pegged the current scale at the
> 2000-2200 level, and remarked in passing that it was desirable to have
> "enough ballast" so that ratings did not become negative. Clearly, he
> misjudged the extent to which rated chess would penetrate the market,
> and the breadth of playing skills now being rated is much more than he
> estimated. 2200 is still 2200 (more or less), as Elo intended - but the
> mean rating has declined precipitously with the rise of Scholastic (and
> pre-Scholastic!) chess in the USA. Some folk persist in pointing to
> this drop as evidence of "deflation" - but it's not.
>
> A pop quiz: without consulting any source, please guess the "mean FIDE
> rating". Similarly, estimate what it was 5 years ago. Explain, in the
> form of a pleasant essay, why this does not indicate that FIDE ratings
> are deflating.
>
> Cheerio, pip pip, tallyho
>
> --
> Kenneth Sloan sl...@cis.uab.edu
> Computer and Information Sciences (205) 934-2213
> University of Alabama at Birmingham FAX (205) 934-5473
> Birmingham, AL 35294-1170 http://www.cis.uab.edu/info/faculty/sloan/
I have no idea of the mean FIDE rating but I have som ideas about
what's happening with a rating pool over time. There are a lot of
peculiar things going on affecting the overall rataings.
We can suppose that the total sum of the 'real' stength in the rating
pool should be the same as the total sum of ELO points to be stable.
- New players are introduced and are given an estimated rating.
If this rating is higher than their 'real' strength theese rating
points will be distributed to the others. If the rating is below
his 'real' strength, ELO points will be 'stolen' from the others.
- New players will be stronger. Often several hundred ELO points.
This means that the total 'real' strength in the pool will be
higher than the total amount of ELO pints to distribute within the
pool.
- Old players will drop in strength. This is the other way around
compared to the previous point.
One might think that point 2 and 3 are compensating each other to
nil effect but this doesn't seem to be the case from my point of view.
New players gaining strength will begin to play even more. Old players
losing strength will reduce their number of games and stop playing
before all the 'stolen' ELO points from the beginning will be 'given'
back to the rating pool.
Do you have more?
--
J-P Fendrich
Joe Stella
f...@watson.ibm.com (Feng-Hsiung Hsu) wrote:
>Are you sure that Larry wrote it? He normally checked his sources. The
>above statement was unfounded, and we never did say it. The best micro
>[...]
I don't understand why no one has answered this post at all,
especially considering how everyone says they like to hear about
DB.
Anyway, I would like to say Welcome Back to the Newsgroup, it's
been a long time since you posted here last...
Now, can we convince you to keep going? Like for instance --
>parallel search and the limited hash table access, a less flexible move
>generator which probably makes micros' search algorithm useless, and so on.
>[...]
What do you mean by a "less flexible move generator" and why would it
make the micro's search algorithm "useless"?
>In the book, "Computers, Chess, and Cognition", we did say, with lots of
>uncertainty, that perhaps there was a "100-200" rating points increase to
>be gained, but that required going beyond what the micros were doing.
"Beyond" in what way? Are you talking about parallel search algorithms,
or something else?
Joe Stella
brucemo
Rolf Tueschen wrote:
> >Add Kasparov to the list of people you've smeared in here on zero evidence.
>
> >bruce
>
> Buddy, I know what you want so dearly: to censor (extinguish) people who are of
> different opinions. You won't succeed, proxy. .)
If you are going to criticize people, I would like you to make some effort to make
sure that your criticisms are founded in fact.
My sole point.
bruce
Mark Rawlings
NoJun...@here.please (Joe Stella) wrote:
> f...@watson.ibm.com (Feng-Hsiung Hsu) wrote:
>>Are you sure that Larry wrote it? He normally checked his sources. The
>>above statement was unfounded, and we never did say it. The best micro
>>[...]
>I don't understand why no one has answered this post at all,
>especially considering how everyone says they like to hear about
>DB.
Hsu's post never showed up on my news server. I did find it on
Dejanews, though. Anyway, I'm not sure who actually wrote the
article, but Larry definitely was the Senior Editor.
Mark
Robert Hyatt
brucemo (bru...@nwlink.com) wrote:
: Rolf Tueschen wrote:
: My sole point.
: bruce
Actually, if your unfounded speculation that Kasparov would intentionally throw
games to avoid blowing DB out of the water were true, then *every* computer vs
GM game (Aegon comes to mind) would be suspect. (Bruce did *not* say this of
course, rather he ws responding to this "claim.") I think there was enough
money and prestige on the line that he did the best he could... Otherwise he
certainly would have avoided the last game or two...
Chris Whittington
--
http://www.demon.co.uk/oxford-soft
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
<5feh4n$h...@juniper.cis.uab.edu>...
What's the fuss about ?
Assuming Kasparov is well stronger the DB (which I assume), and, knowing
that there is a big purse for this contest, and, knowing that more contests
will follow on, especially if DB is *near* to 3-3, or 3.5 to 2.5; then
Kasparoc has mighty big incentive (as last time) to frig the result to a
human-Kasparov win at 3.5 to 2.5
No ?
Chris Whittington
>
Torstein Hall
Chris Whittington <chr...@demon.co.uk> skrev i artikkelen
<01bc27eb$db471d80$c308...@cpsoft.demon.co.uk>...
Its always a big incentive to steal a lot of money, making a pirat copy of
someones program, and so on. Its quite another story if you acuse someone
of doing stealing or other bad deads!
--
Torstein Hall
tors...@eunet.no
http://login.eunet.no/torshall/
Rolf Tueschen
hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>brucemo (bru...@nwlink.com) wrote:
>: Rolf Tueschen wrote:
>: > >Add Kasparov to the list of people you've smeared in here on zero evidence.
>: >
>: > >bruce
>: >
>: > Buddy, I know what you want so dearly: to censor (extinguish) people who are of
>: > different opinions. You won't succeed, proxy. .)
>: If you are going to criticize people, I would like you to make some effort to make
>: sure that your criticisms are founded in fact.
>: My sole point.
>: bruce
>Actually, if your unfounded speculation that Kasparov would intentionally throw
>games to avoid blowing DB out of the water were true, then *every* computer vs
>GM game (Aegon comes to mind) would be suspect. (Bruce did *not* say this of
>course, rather he ws responding to this "claim.") I think there was enough
>money and prestige on the line that he did the best he could... Otherwise he
>certainly would have avoided the last game or two...
---------------------------------------------------------------------
People seem to like this strange BH logic otherwise he would prefer to stop
playing dirty. But certainly Tueschen did never speak of *throwing games*. And
*intentionally* doing so neither. The whole idea was a bit more complicated.
Thinking of imputing some suspense could result in various/different outputs.
The meaning was that GK might think of some *special* playing with some
experiments. Trying to understand that monster. Because Kasparov surely loves
computerchess. And he is interested in its progress.
All this came up because of my serious conviction that GK could win against any
machine as he liked. Experiments and suspense don't mean cheating by throwing
games. I would never say that. And Kasparob would never throw a game
intentionally.
But thoughts must be allowed when GK in wathever event against computers always
lost some points. See Fritz, Genius and Deep Blue. Against Fritz however he
played generally blitz.
Now the next stupid BH logic. Said all this about GK it's a bit comical to look
at other events like AEGON. Tueschen didn't say anything about AEGON. And he
wouldn't. Period.
Last note. What the heck BH meant with *enough money and prestige*? How BH
could know about what is *enough* for GK?
Rolf Tueschen
brucemo <bru...@nwlink.com> wrote:
>Rolf Tueschen wrote:
>> >Add Kasparov to the list of people you've smeared in here on zero evidence.
>>
>> >bruce
>>
>> Buddy, I know what you want so dearly: to censor (extinguish) people who are of
>> different opinions. You won't succeed, proxy. .)
>If you are going to criticize people, I would like you to make some effort to make
>sure that your criticisms are founded in fact.
>My sole point.
>bruce
-------------------------------------------------------
In these 8 days you didn't find out difference between opinions, criticisms
and pure facts.
But as I asked elsewhere tell me the wrong/false statements and I'll check
them.
Calling some opinions about GK that could never be proven selfunderstood sort
of *smearing* speaks against your vocabulary not my expressing opinions.
Chris Whittington
--
http://www.demon.co.uk/oxford-soft
Torstein Hall <tors...@eunet.no> wrote in article
<01bc2827$135367c0$1afb47c1@Ptorshall>...
>
>
> Chris Whittington <chr...@demon.co.uk> skrev i artikkelen
> <01bc27eb$db471d80$c308...@cpsoft.demon.co.uk>...
> >
> > --
> > http://www.demon.co.uk/oxford-soft
> >
> > Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
> > <5feh4n$h...@juniper.cis.uab.edu>...
> > > brucemo (bru...@nwlink.com) wrote:
> > > : Rolf Tueschen wrote:
> > >
> > > : > >Add Kasparov to the list of people you've smeared in here on
zero
> > evidence.
> > > : >
> > > : > >bruce
> > > : >
> > > : > Buddy, I know what you want so dearly: to censor (extinguish)
> people
> > who are of
> > > : > different opinions. You won't succeed, proxy. .)
> > >
> > > : If you are going to criticize people, I would like you to make some
> > effort to make
> > > : sure that your criticisms are founded in fact.
> > >
> > > : My sole point.
> > >
> > > : bruce
> > >
> > > Actually, if your unfounded speculation that Kasparov would
> intentionally
> > throw
> > > games to avoid blowing DB out of the water were true, then *every*
> > computer vs
> > > GM game (Aegon comes to mind) would be suspect. (Bruce did *not* say
> > this of
> > > course, rather he ws responding to this "claim.") I think there was
> > enough
> > > money and prestige on the line that he did the best he could...
> > Otherwise he
> > > certainly would have avoided the last game or two...
> > >
> >
> > What's the fuss about ?
> >
> > Assuming Kasparov is well stronger the DB (which I assume), and,
knowing
> > that there is a big purse for this contest, and, knowing that more
> contests
> > will follow on, especially if DB is *near* to 3-3, or 3.5 to 2.5; then
> > Kasparoc has mighty big incentive (as last time) to frig the result to
a
> > human-Kasparov win at 3.5 to 2.5
> >
> > No ?
> >
> > Chris Whittington
>
> Its always a big incentive to steal a lot of money, making a pirat copy
of
> someones program, and so on. Its quite another story if you acuse someone
> of doing stealing or other bad deads!
Come on, this is not stealing. Its a prize fund offered by IBM. Kasparov
can play how he likes, win lose draw, its up to him.
Its just that I remember last time K played DB. I said exactly the same
thing, Kasparov would/did/had powerful incentive to deliberately engineer a
result of 3.5 to 2.5. Nobody flamed me for saying it, we all just
discussed.
Now, Teuschen says it, here in this thread. Everybody jumps on him. What's
the difference: same content, different poster.
So if I have to spell it out. Why is Teuschen getting flamed for saying
something mow, that has been said before without flames ? That's my
rhetorical question and the purpose of my original post.
I think Teuschen is a nutcase, but I'ld also like to see him allowed to
post here without getting attacked. Often he rambles on endlessly about
stuff that was dead and done, this is silly. Sometimes he has some
interesting observations to make. If he gets flamed just for being himself,
then we lose his interesting observations.
Chris Whittington
Robert Hyatt
Chris Whittington (chr...@demon.co.uk) wrote:
: Come on, this is not stealing. Its a prize fund offered by IBM. Kasparov
: can play how he likes, win lose draw, its up to him.
: Its just that I remember last time K played DB. I said exactly the same
: thing, Kasparov would/did/had powerful incentive to deliberately engineer a
: result of 3.5 to 2.5. Nobody flamed me for saying it, we all just
: discussed.
: Now, Teuschen says it, here in this thread. Everybody jumps on him. What's
: the difference: same content, different poster.
: So if I have to spell it out. Why is Teuschen getting flamed for saying
: something mow, that has been said before without flames ? That's my
: rhetorical question and the purpose of my original post.
: I think Teuschen is a nutcase, but I'ld also like to see him allowed to
: post here without getting attacked. Often he rambles on endlessly about
: stuff that was dead and done, this is silly. Sometimes he has some
: interesting observations to make. If he gets flamed just for being himself,
: then we lose his interesting observations.
: Chris Whittington
I hope my comment wasn't taken as a flame. I don't believe Kasparov would
engineer a close result intentionally, particularly knowing that one bad move
in one game can turn a 1 into a zero in short order. However, notice the
"I don't believe"... I don't see any way *anybody* could "know" whether this
is true or not, except for *one* person. :)
Johan Andersson
Is this the same person responsible for "Super Chess" on the
Sinclair ZX Spectrum computer?
Nice to see at least one Spectrum programmer still active!
--
Johan Andersson
Technology Lab
Dept. of Solid State Electronics
Chalmers University of Technology
S-412 96 Gothenburg
Sweden
Telephone: (0)31-772 1873
E-mail: jo...@ic.chalmers.se
HTTP: http://ic1.ic.chalmers.se/~johan