C-extension in Python -- returning results

[Keith Farmer]

tatebll at aol.com (Bill Tate) wrote in message news:<cb4ba455.0111130433.6d50b7b4 at posting.google.com>...

> Keith - now that was an interesting problem.  Since you've found the
> source of your problem this may be moot but I would definitely second
> Jive's suggestion to consider using a neural network for this problem

Got my degree in Physics and Astronomy -- I just wish there was better
notational support for operator math, but really, that's probably not
possible for a single-line style?

How efficient would a neural network be in solving this problem?  I
achieve accuracies on the order of 1e-6 to 1e-7 (both rms and max
error) in less than a second on a 1.2GHz Win2k Pro box accessing a
remote database (data transfer of roughly 100k); quite possibly, I
could achieve this within less than half a second, if I fiddle with
the number of samples I take.  Keep in mind that the motion described
is more or less well-behaved (the earth/moon system being an annoying
exception).  As it is, one coordinate changes as a nearly constant
rate, and the period of the other two is much larger (5x or more) than
the time period of the sample.

Another problem (which I won't tackle for this application, though
it'd be interesting just to see how it'd behave) would be to see how
changing the degree of the polynomial affects the accuracy of the fit,
given a large sample size (~5k points) and a constant period of time
(several months, at least).  The benefit of exploring this would be to
reduce the overall dataset even further (18 numbers * n periods versus
a+b+c numbers for 1 combined period -- earth would still be a
problem).

.. this is what I do when I'm unemployed?  Sheesh!