Problem with errors of random parameter seeds in amptools fit

[aaust]

Running the fit executable with the argument "-r " correctly performs number of fits, each seeded with random parameters, both production coefficients and internal parameters.
However, the initial uncertainties ("APPROXIMATE ERROR" in log file) assigned to these random parameters for the first fit are taken from the initial values specified in the configuration file. For subsequent fits, these uncertainties are taken from the previous fit, even if this fit failed. This causes a correlation between individual fits. In the worst case, one fit fails and the crazy error values cause all other fits to fail as well.

[jrstevenjlab]

We noticed this with the parameter scans as well, where each point in the scan is done in sequence within the fit executable. So the initial uncertainties are carried from one point in the scan to another and if one point in the scan fails then the next point could be influenced. It wasn't clear how these initial uncertainties influenced the fit result but we didn't come to a conclusion how to deal with this that I recall.

[mashephe]

This really an AmpTools/MINUIT issue rather than a halld_sim issue. I think the problem is not with the errors directly but with the step sizes (which are likely based on the error estimates). I don't think MINUIT has any hooks to reset them and there may be a need for an AmpTools workaround. I need to dig a bit into the version of MINUIT that is embedded in AmpTools. It's good to find these issues and fix them.... you can assign this one to me. Matt

…

On Oct 19, 2022, at 3:05 PM, Justin Stevens ***@***.***> wrote: We noticed this with the parameter scans as well, where each point in the scan is done in sequence within the fit executable. So the initial uncertainties are carried from one point in the scan to another and if one point in the scan fails then the next point could be influenced. It wasn't clear how these initial uncertainties influenced the fit result but we didn't come to a conclusion how to deal with this that I recall. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.

[mashephe]

I looked at this a little more, and it looks like minuit has a command mnrset that sets the function value and errors to be undefined. This is probably what is needed. One would need to wrap this call into the MinuitMinimizationManager. Create a member function for MinuitMinimizationManager called resetErrors() and that function should make the following call:

m_fitter.mnrset(1);

Then call resetErrors() whenever the parameters are reset, e.g., from the AmpToolsInterface.