Re: Numerical Methods Programming: was: Four people decided the yadda yadda yadda
On Sun, 2 Mar 2014 14:46:21 -0500
"Steve Litt of Troubleshooters.Com" <littdom@gmail.com> wrote:
> On Sun, 2 Mar 2014 11:31:08 -0700
> ghaverla <ghaverla@materialisations.com> wrote:
>
>
> > Most of the programming I have done is numerical methods,
>
> What language did you use? I've used a little bit of Scheme, and kind
> of liked it for numbers.
>
> When you say numerical methods programming, do you mean this:
>
> http://en.wikipedia.org/wiki/Numerical_analysis
I still have fortran as a user ID on some websites. My M.Eng. project
(mid 80's) was a dynamical system of differential equations where
equations disappeared at random, in VAX FORTRAN. I had doubly linked
lists and garbage collection in FORTRAN to run this stuff. I've done
some in C, some in C++ and a lot in Perl.
> I've often thought of writing a differentiator program in Python, or
> who knows, maybe Scheme, perhaps something that solves y = f(x) type
> equations simply by iterating closer and closer to see where it
> crosses the axes. This gets ever more inviting, because I'm
> continually forgetting more and more of my high school and college
> math.
>
> I'd love to know what you're doing and how you're doing it.
This big project I want to start at Savannah (the nongnu side) is
probably going to be 45-55 Perl modules when it is finished. The data
I am working from is GPS tracklog data, but there is no reason a person
couldn't have trace element analysis of a surface from inside an Auger
microscope (my background is materials science and engineering).
Outliers happen, especially with personal GPS and no differentials (or
postprocessing). Most of the methods for detecting outliers, strictly
speaking, can only be used to detect a single outlier in data. One of
my modules is in implementation of Pierce's Criteria, where I have
tables to detect up to 9 outliers in up to 60 data points. As I am
starting with tens of thousands of points, I have to get down to 60
before I can think about applying this to flag outliers (in GPS, some
errors are "blunders", which can be corrected after the fact). I am
using a median quadtree to partition the data so that it eventually
gets down to 60 or less points. The idea is to surface things
eventually using thin plate splines, but they have a free parameter in
their design. So, I want to wrap the TPS fitting with Leave One Out
Cross Validation. To find the optimal value of the parameter, I am
supercharging Brent's Method for function minimization based on work
Jack Crenshaw (Embedded.com) has done, and adding some ideas of my own
(he wants his new version to work in embedded devices, I want mine to be
general purpose).
Another thing I am trying to set up, is coming up with a course
(professional development) to teach engineers about computational
statistics (Monte Carlo, bootstrap, jackknife, and others).
I'm a bit of an outlier in Materials Science and Engineering.
Gord
Reply to: