Just realized that my post didn't get the correct formatting.
My code is actually working fine. Typical data is confidential, so I can't give you something to play around. The measurement data is really good (not noisy at all), the Y values are taken with great dynamic gain covering about 8 orders of magnitude.
Here my old questions, I hope they get formatted:
* The parameters differ by several magnitudes (up to E+15!) so I get severe numerical errors. Is it possible to 'normalize' the parameters?
* How to apply boundaries for the parameters, is it correct to corce them after each iteration?
* How to deal with the implicit formula? Currently I ignore the fact that Y occures on the right hand side of the equation (I use the measurement data directly). Is it straight-foreward to implement Lev-Marq for data-points with F(x,y)=0?
* In addition to this, literature suggests to use 'orthogonal distance regression' (it takes into account that both X and Y are 'noisy' signals). I haven't found any description of this method.
Here my new insights: I mainly deal with numeric problems. The formula in my first post contains a really nasty part:
exp(K * (X - Y * P2))
where P2 is one of the parameters I'd like to obtain. If the initial guess isn't good, the term inside the brackets evaluates to numbers of several hundereds (for some X-Y pairs), and then raising to the power of e is giving me values in the range of E+100. In this situation all other variations (of the other parameters) is buried under this value. So in some situations, the fit terminates (not exactly matching the data but somehow ok). But my initial guess for the other parameter P1 didn't change at all during the fit (P1 is also about E-15 smaller than P2).
Have a nice weekend!
Felix
