Re: [gnugo-devel] RE: qsort() causes platform dependence

[Heikki Levanto]

On Thu, Nov 21, 2002 at 02:48:53PM -0500, Aguydude wrote:
> [...] but l'd consider using 1.279604943109628 instead
> of 1.3.

I must admit that I have not studied the matter carefully - hardly at all.
But this kind of modification does ring a warning bell in my mind -
replacing one magic number (1.3) with another (1.27...) that is almost the
same, but looks a bit more complex (at least in decimal notation). Possibly
there are some anomalous cases that sort badly with the 1.3 value, but quite
likely there are some other anomalous cases that do as bad with the
"improved" value. They may not be as easily recognized (for us humans), and
thus will be less likely to get attention, but that does not prove that the
new magic number is better in any real sense.

Computer science is a new and fastmoving branch of science, but sorting is
one of the few old well-known problems for which there exists a good number
of relatively old, tried, and well established solutions. New sorts are
proposed often enough, but I haven't heard of anything that could actually
make serious improvements to the known ones.

There are cases where so much can be known of the data that it is possible
to take advantage of that and choose a sort that may not be optimal in the
general case, but will do exceptionally well in this particular one. I do
not see that this should be the case with GnuGo, especially since so much of
the time seems to be spent in doing other than sorts.

Of course, if someone wants to show that a new wonder-sort is so much better
for GnuGo, it is easy to do - show some performance numbers. Until then I'd
rather worry about cutting search nodes, or about a better understanding of
what actually is happening on the board.

Just my git feeling, of course - I'd be glad to be proven wrong.

- Heikki


-- 
Heikki Levanto  LSD - Levanto Software Development   <address@hidden>