Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: More on the "bad math" after an important email...

Author: Robert Hyatt

Date: 19:24:34 09/03/02

Go up one level in this thread


On September 03, 2002 at 20:48:08, Gian-Carlo Pascutto wrote:

>On September 03, 2002 at 20:20:48, Roger D Davis wrote:
>
>>Wow, this comment is in exceptionally bad taste. You don't question the
>>scientific integrity of a researcher lightly, particularly in a public forum.
>
>I was responding to this post from Robert:
>
>---quote----
>
>[snip]... But that
>doesn't mean things were fabricated.
>
>But if you want to believe so, feel free.  It doesn't change a thing either
>way...
>
>------------
>
>I don't know what exactly happend with the results. It seems from this thread
>that even Robert doesn't know. Just because of this, no matter how they were
>produced, I think they are questionable.
>
>--
>GCP


Nah, that is utter nonsense, by any measure you care to name.  The _speedup_
was the heart of the DTS results.  DTS was about nothing other than producing
a reasonable speedup on that particular machine, the Cray.  The speedups were
directly calculated from log files and checked and re-checked, even during the
referee process.  The nodes and times were added after the fact.  Nodes I
now remember _had_ to be calculated as there was no way to "observe" them in
the context of a game-like search that doesn't always finish an iteration to
reach a good node output point.  The times I simply don't remember.  But in
the conversation this afternoon, it is certainly possible that computing the
times would have been a natural thing to do as it was far easier and less
error-prone, but I simply don't remember, and won't speculate that it was or
wasn't done...

If I had the logs, I could easily tell.  But I don't, so I can't, so that's
pretty much that.

But the speedup data is _dead_ right.

I think it pretty funny to make an issue about the two values that change the
most in a parallel search, same position, different runs.  :)  There is already
a significant error in reporting _any_ node counts, much less reporting "the"
node count that really doesn't exist in a decent parallel search.



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.