| model fitting [message #41163] |
Tue, 28 September 2004 00:48 |
michael.asten
Messages: 1
Registered: September 2004
|
Junior Member |
|
|
I am doing parametric inversion observed data (typically 30 data
points)to a model (typically 8 significant parameters), which I guess
can also be called curve fitting. Noted the many recommendations to
Craig Markwardts mpfit (thanks to all those who have blazed a trail
here).
Specific issues for me are
(1) the "curve" to be fitted resembles a distorted Jo bessel function,
ie has positives, negatives, crossovers, multiple maxima and minima.
The positions of crossovers and maxima/minima on the x-axis carries
most of the physical information controlling the best-fit parameters
sought.
(2) the observed values of the curve have a some systematic
(non-gaussian) errors or biasses, hence I wonder if a minimum
deviation criterion (L1 norm) would be more stable than least squares.
I recall using software with SVD curvefitting algorithms decades ago
which were quite unstable on data of type(1) above having crossovers.
Any general thoughts from the experts out there relating to these
issues? Regarding (2) above, is there a variant of mpfit.pro (or the
idl-supplied curvefit.pro) which allows for the L1 norm?
Warmest thanks for any pointers.
Michael Asten
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
