Come analyze HEASARC, IRSA, and MAST data in the cloud! The Fornax Initiative is now welcoming all interested beta users.
improve
search for a better fit near the current one
| Syntax: | improve |
|---|
The improve command tries to escape a local minimum without restarting the fit from scratch. It is the warm-restart counterpart of fit global: starting from the current best fit, it runs a short derivative-free Differential Evolution pass seeded in a shell around the current parameter values, then polishes the best point it finds with the current method. Use it when a fit has converged but you suspect it has settled in a local rather than the global minimum — improve perturbs the parameters outward, re-optimizes, and keeps the new result only if it is better.
improve requires a valid fit first (it warm-starts from the current parameters). Like fit global it is never-regress: if the search does not beat the current statistic the original parameters are restored, so it can only help. It works with every fit method (in earlier versions improve was available only for the Minuit methods; it now uses the same Differential Evolution machinery as fit global for all of them). The run can be interrupted with Ctrl-C, which leaves the pre-command fit in place.
The difference from fit global is the scope of the search.
improve explores a local shell about the current fit, so it
is cheaper and is meant for refining a fit or hopping a nearby barrier;
fit global searches the full soft-limit box of every
parameter from scratch and is the tool for a genuinely global, multi-modal
problem. As with fit global, any additional competitive
minima within
of the best fit are reported and are
available in PyXspec as Fit.globalBasins.
Example:
XSPEC> fit // Levenberg-Marquardt converges to a local minimum. XSPEC> improve // Differential Evolution shell about that fit, then a polish; if a // deeper minimum is found the parameters move there, otherwise they // are left unchanged.