Re: fitting, step

On Tue, Jul 06, 2010 at 17:45 -0400, Maurice Leutenegger wrote:
> However, when I do "fit_counts", it doesn't fit the data, or it barely
> changes the fit parameters. The one clue that something may be wrong is
> that the "step" field is set to zero for all model parameters I want to
> fit, even if I explicitly set it using "set_par". I have checked this
> with XMM and Chandra data, and there, when I load data, define a model,
> and fit, the fitting algorithm proceeds normally, and the "step" field
> of each thawed parameter is non-zero.

I suspect that the value of 'step' is not the problem. The
'step' value is used only when step>0 (zero values are
ignored), and some optimization methods don't use step at all.

Nevertheless, there was a bug in that the set_par was not
modifying the value of 'step'.  This bug has been fixed in
isis-1.6.0-21 which is available from the web page:

   http://space.mit.edu/cxc/isis/download.html

> So somehow this seems to be a problem with this data format.
>
> Maybe this is a problem with "step", or maybe that's a red herring, but
> either way, I can't change it from zero, and I can't fit the data. Any
> suggestions?

The optimization code has no dependence on the format of the
data file, if any.

What is the value of the fit statistic when the parameters
are set at the initial values?  (e.g. do an eval_counts)

How much does the fit statistic change after you make a small
change in one or more fit parameters?

What is the range of fractional errors on the data values?
(e.g. abs(uncertainty)/data_value).  If any of the
uncertainties are "extremely small", then the optimizer may
have trouble.

Thanks,
-John

----
You received this message because you are
subscribed to the isis-users list.
To unsubscribe, send a message to
isis-users-request_at_email.domain.hiddenwith the first line of the message as:
unsubscribe