[Help-gsl] Levenberg-Marquardt misbehaves...

[Federico Zenith]

Hi everybody,
I am working on a fuel cell model. Right now I have implemented a regression 
member function in a "Fuelcell" class (I work in C++ with the static-function 
hack for GSL).

The Regressor( ) should, given some observations, find out the best parameters 
(there are 8, 7 of the cell plus the inductance of the wiring) of a impedance 
test, which is used to extract data useful for cell modelling.

The underlying Fuelcell is quite nasty - there are two brent and one 
steffenson iteration loops below that level. All parameters, to have any 
meaning, must be > 0.

The various _f, _df and _fdf of the LM loop return GSL_EDOM if any parameter 
is < 0, and also print out which parameter is bad and with what value.
What happens is that the algorithm cycles between two values for the 
cathode-side diffusion-layer thickness, i.e. -246.626 and -0.65914 (metres).
This might have to do with the fact that the diffusion layer is not very 
important for the data set I'm working in.

How to I tell LM to stay away from the negative domain?

Cheers,
-Federico

PS- I noticed there are no derivative-free nonlinear regression methods. I 
happen to have a paper from 1978 about DUD, exactly such an algorithm. Is 
does not even seem too difficult to implement. Is there interest to include 
such an algorithm?

pgpfT1eiL3FNx.pgp
Description: PGP signature