Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: ICC rating study

Author: Robert Hyatt

Date: 18:48:27 04/20/00

Go up one level in this thread


On April 20, 2000 at 17:32:08, James T. Walker wrote:

>On April 20, 2000 at 09:09:18, Robert Hyatt wrote:
>
>>On April 20, 2000 at 01:39:55, Jason Williamson wrote:
>>
>>>On April 19, 2000 at 23:55:06, Robert Hyatt wrote:
>>>
>>>>On April 19, 2000 at 23:53:31, Robert Hyatt wrote:
>>>>
>>>>>As I had mentioned a while back, I have a sack full of quad xeon 550 machines
>>>>>in a beowulf cluster.  While waiting on a few final pieces to arrive, I decided
>>>>>to do what I thought was an interesting test:
>>>>>
>>>>>two identical machines, and I mean _identical_.  Quad xeon 550's, 27 gigs of
>>>>>SCSI disks in a raid-0 (striping) configuration, 512mb of ram, etc.  IE
>>>>>everything is identical, with all the 3-4-5 piece compressed tablebases,
>>>>>same opening books, etc.
>>>>>
>>>>>The only difference was that 'crafty' plays computers and humans, while scrappy
>>>>>only plays humans.  Several of us had postulated over the years that if you only
>>>>>play humans, you can drive your rating through the roof.  Using the same
>>>>>formulas (5 3 blitz or faster, 60 60 standard or faster, or most any bullet)
>>>>>I have been watching the two programs for a month now.  And they seem to
>>>>>hover at the point scrappy == crafty+100, roughly.  Standard has crafty
>>>>>actually higher, but that is because crafty is playing standard against
>>>>>computers, while scrappy is playing very little standard as humans seem to be
>>>>>avoiding that for the most part... and those that do play standard play crafty
>>>>>as it is better known.
>>>>>
>>>>>100 points was a surprise...  as I thought it would be more.  At present crafty
>>>>>is at 31126 and scrappy is at 3219 blitz (which is the most stable ratings
>>>>
>>>>argh:  ^^^^^
>>>>
>>>>
>>>>3126 of course...
>>>>
>>>>31126 won't be reached for maybe 10-20 more years.  :)
>>>>
>>>>
>>>>
>>>>
>>>>since
>>>>>most games are blitz).
>>>>>
>>>>>It seems that not playing computers is _not_ a way to grossly inflate your
>>>>>rating, unless you consider 100 as inflated.  Note that a rating of 3200 is
>>>>>very high, considering that there are not a lot of GM players that are rated
>>>>>even 3000.  I watched scrappy play a 16 game match earlier this week, it won
>>>>>8 games, lost one, then one 7 more, for a 15-1 result (5 3 blitz).  It lost 32
>>>>>rating points for the effort.  :)
>>>>>
>>>>>I am going to continue the experiment until I get the rest of the beowulf
>>>>>hardware (another quad box and a fast ethernet switch to complement the
>>>>>giganet switch).  If you watch the ratings, you will get a feel for the
>>>>>difference in playing only humans and humans + computers...
>>>
>>>
>>>What do you figure your rating gain will be with the beuwolf beast?
>>
>>
>>Difficult to say right now.  I am running on a quad xeon.  I will be able to
>>use 9 machines (total) although really only 8 of them have the giganet inter-
>>connect.  8 times the horsepower ought to give a search at _least_ a factor
>>of 4 faster, which is conservative I hope.  that would be the equivalent of
>>doubling the speed 2 times.  I would think at least 100 rating points, maybe
>>more, but mainly at standard time controls, as distributed computing is not
>>going to be great for blitz/bullet...
>
>Hello Bob,
>That's too bad!  I was looking forward to seeing that thing take over the world
>at blitz!  Anyway it will be interesting to watch.  Can you give me a couple of
>clues as to why the blitz will not work so good?  In laymans terms of course.
>Regards,
>Jim Walker


The main reason is that with a blitz game, I am looking at maybe 10 plies
for a decent search depth, rather than the normal 13-14 in middlegames.  The
10 plies means lots of the parallel searches will be 'small'...  and small is
bad on a machine where the cost of sending a message might easily be longer
than the time required to search the resulting position...

I am not sure it won't do well at blitz, so we will see on ICC one of these
months.  And I am being _very_ conservative with the 4x speedup.  I am sure I
can get that without any trouble.  And I believe that I can possibly get to
close to the speedup I get with a normal parallel search, although there are
some issues like a global hash table that will take some thought...

More as I think about the issues, but right now, I have been so busy ordering
equipment for a new lab, and working on being 'non-hackable' on our local
network, that I simply haven't done much chess in a couple of months...  But I
definitely have not 'quit' or 'retired'...  :)



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.