search

FCS-analysis / PyCorrFit

data analysis and fitting software for fluorescence correlation spectroscopy (FCS)
pycorrfit.craban.de
Other
47 stars 13 forks source link

Convergence criteria of the heuristic fitting algorithm #156

Closed weidemann closed 8 years ago

weidemann commented 8 years ago

In 0.9.7 fitting curves with complex models (e.g. T-T-2D-2D) works usually fine, however a considerable fraction of curves take a huge computing time. Sometimes even though the fit was already perfectly placed on the correlation data.

In some cases, for ideal starting values PyCorrFit tries to re-optimize, usually without success or even by deteriorating the fit quality. Thus PyCorrFit cannot always recognize that the fit parameters are close. The cut-off for aborting the fitting cycles may be more tight. Sometimes values are improved at the 10th digit after comma.

Because the progress in batch fitting is not yet visible, a few of such trouble makers can practically prevent batch fitting at all. There are currently no other means than going through the pages one-by-one and eliminate these data.

weidemann commented 8 years ago

Related to this issue I have the impression that the heuristic fit does not tolerate well inactivation of parameters. Sometimes when the fraction is fixed to zero PyCorrFit still varies the corresponding correlation time, even though only a few digits after comma.

weidemann commented 8 years ago

Previous impression might be wrong. I now realized that PyCorrFit is extremely sensitive on choosing inappropriate models. If in doubt this is a very useful indicator :-)