Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: An Experiment that disproves Hyatt's 1000X NPS Theory

Author: Peter Kappler

Date: 14:01:57 09/20/05

Go up one level in this thread


On September 20, 2005 at 16:23:42, Robert Hyatt wrote:

>On September 20, 2005 at 15:53:32, Peter Kappler wrote:
>
>>>>>(1) deep thought (deep blue's direct predecessor) was the first (and only)
>>>>>program to produce a 2650+ performance result, playing games only against GM
>>>>>players, at 40 moves in 2 hours only for the time control.  It did this over 25
>>>>>consecutive games (intervening games could not be ignored if the result was
>>>>>bad).  No other program has yet accomplished this.
>>>>>
>>>>
>>>>This was impressive 10 years ago, but today any commercial program (and probably
>>>>a fair number of amateurs) could easily accomplish this feat running on ordinary
>>>>hardware.
>>>>
>>>
>>>Why haven't we seen it happen?  Note I am not talking about the much faster time
>>>controls we have seen more of lately.  But real 40 moves in 2 hours.  I don't
>>>believe _any_ program today could pull this off on "ordinary" hardware.  They
>>>would be hard-pressed using very high-end (say quad opteron) systems...
>>>
>>
>>There are so many examples it's hard to know where to begin.  I'll just list a
>>few.
>
>None of those address my point.  If you play in N tournaments, and pick one, you
>can get most any sort of TPR you want.  If you play nothing but GM players, and
>I do mean _no_ IM/FM/lower players, and you play 25 consecutive games, counting
>each and every one, a 2650+ is really a daunting task.


Shredder played in the Argentina event in 2003, 2004, and 2005.  I couldn't find
a TPR for the 2004 tourney, which is why I didn't post it.  But, if you add up
the games from all 4 events (it played in 2 events in 2004) it's something like
40 games, with a TPR above 2700.

And the "GMs only" restriction of the Fredkin Prize doesn't make much sense to
me.  I suspect it's actually harder to get the 2650 result against a field of
IMs/FMs.

>
>I won't say none of these programs can do that.  I will say that none have
>_done_ it yet.  The fredkin prize was specifically structured to require 25
>games so that a single tournament could not be used, for the reason given above.
>

What's the point?  Lets turn the tables and say that the "Fredkin II Prize" is
for producing two 2750+ results in a 12-round human round robin.  Then I can say
 that Deep Thought/Blue never did this.


>So your examples, while certainly amazing enough, are not (yet) in the same
>category as what a 1992 special-purpose piece of hardware did.  We are now 15
>years beyond the point where DT2 was put together in 1990.
>

We'll just have to disagree on this.  I think Shredder's results are probably
more impressive.

One thing that is never discussed is how totally unprepared those GMs must have
been in the Fredkin games.  I mean, imagine playing Deep Thought in 1988.
You're a strong GM, but you've NEVER played a game against a computer.  You
probably don't even know the basic strengths/weaknesses of the computer.  Think
about how much this favored Deep Thought.

Also, I have to say from looking at the old Deep Thought games, that it was
CLEARLY weaker than todays micros.  Just analyze a few games and this will
become painfully obvious.


>Secondly, while not being sure, I do not believe all the events given below are
>40/2.  Some certainly are, particularly those Ed was doing in his GM challenge.
>But most tournaments have gone to faster time controls.  And in one of the
>listed cases, the games were mixed.  Two slow, two faster, two faster still,
>etc.

I gave the link to Ed's page *ONLY* in reference to Chess Tiger's performance in
Argentina 2001.  I specifically did not mention the other events because they
used mixed time controls.


> Nobody doubts that computers can produce 3000+ ratings at blitz.  They
>have already done it and have been doing it for several years now.;
>

The results I gave were not blitz.

-Peter


>
>
>>
>>Chess Tiger 14 - Argentina 2001 - 2788 performance rating  (P3-866)
>>http://www.rebel.nl/resu.htm
>>
>>Shredder 7 - Argentina 2003 - 2753 performance rating.
>>http://www.chessbase.com/newsdetail.asp?newsid=1111
>>
>>Shredder 9 - Argentina 2005 - 2758 performance rating.     (P4-3500)
>>http://www.chessbase.com/newsprint.asp?newsid=2538
>>
>>And many more.  (Note this article was written in 2003)
>>http://www.chessbase.com/newsdetail.asp?newsid=1141
>>
>>
>>>>>(3) deep blue was 100x faster than deep thought 2, with more chess knowledge
>>>>>than deep thought 2, and this is the box that beat Kasparov in a 6 game match.
>>>>>
>>>>>It is certainly possible that todays fastest computers, running today's best
>>>>>commercial programs, are playing at an equal level when compared to deep blue,
>>>>>although the Kramnik/etc matches were played at faster time controls generally,
>>>>>than 40 moves in 2 hours.  But at best the best micros of today are maybe as
>>>>>strong as the 1997 deep blue system.  Far from being far superior to it.  Based
>>>>>solely on the observations given above.  Each of which can easily be verified
>>>>>multiple ways...
>>>>
>>>>
>>>>I'm not sure how you reached the conclusion that DB '97 is an upper bound for
>>>>the strength of todays micros.  It certainly doesn't follow from the
>>>>"observations given above".
>>>>
>>>
>>>
>>>The GM results of 2M nodes per seconds, vs the DB hardware 100X faster.  What PC
>>>today could do better than a 2650 performance against GMs at 40/2???
>>>
>>
>>Every single commercial and probably at least the top 5 amateurs.
>>
>>-Peter
>
>
>This is speculation.  Since it has _not_ been done yet.  My point.  DT did it 13
>years ago.  And its "big brother" was 100X faster than that.  Do you _really_
>think it was a patzer???  Or something pretty remarkable for its time (or for
>any time, really)?



This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.