Re: [Gretl-users] MLE Advanced
On Fri, 2 Sep 2016, Sven Schreiber wrote:
Am 02.09.2016 um 18:15 schrieb Mario Flórez Porras:
> Beforehand, thank you very much for answer.
>
> I know how to set the tolerance parameter. What I want to know if
> possible is how gretl calculate whether my model mets the tolerance?
I don't think gretl is doing anything special here, it should compare (a
number based on) the latest likelihood value update to the one before, and if
the increment is still too large relative to the convergence criterion, then
it would go on. Then what might happen is that the maximum number of
iterations is reached before convergence occurs.
See also the 'set' options 'bfgs_maxiter' and 'bfgs_maxgrad'.
That's right. We break out of the BFGS loop if the change in the
maximand falls below a given relative tolerance. Specifically the
test is
fabs(fmax - f) <= reltol * (fabs(fmax) + reltol)
where fabs() gives absolute value; fmax is the greatest value of the
maximand to date and f is the current value; and reltol is the
tolerance value.
However, breaking out of the maximization loop in this way is not a
sufficient condition for convergence; we also require that the norm
of the gradient is "small enough" -- as Sven says, see bfgs_maxgrad.
Allin Cottrell