I don't know exactly how the convergence criterion is calculated, but it is usually based on a quadratic form that g'*inv(D)*g where g is the gradient and D is the Hessian (or equivalent depending on algorithm). Basically the sum of the (squared) elements of the gradient are being weighted by their estimated precisions. I bet that is how it is computed. The non convergence example is pretty bad. the norm (of the gradient?) is huge.

Lee

On Fri, Sep 2, 2016 at 11:15 AM, Mario Flórez Porras <mario.florez.porras@gmail.com> wrote:

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?

El 1/09/2016 a las 8:43 a. m., Sven Schreiber escribió:

Am 25.08.2016 um 23:42 schrieb Mario Florez Porras:

Hello,dear Gretl Community.

I'm doing a nonlinear estimation via the maximum likelihood method. I
want to do a batch processing of data but in some cases the estimation
don't satisfise tolerance (1,81899e-012), don't met the convergence
criterion. Does anyone knows how it is calculated whether or not the
model meets tolerance?

Apparently nobody wants to answer in the holiday season....

There is a settable option using the "set" command for the MLE tolerance. See the "set" documentation under the section "Numerical methods" in the built-in command reference; as in:

set bfgs_toler <some-value>

You may have to also choose the optimizer for ML explicitly before the mle block, as in:

set optimizer BFGS

Some remarks to the masters of the documentation: First I don't see a "newton_toler" option and I don't immediately understand why this is absent. Also the way to adjust the tolerance setting might be mentioned in the MLE chapter of the guide, I didn't see it there (but may have missed it), I only saw a mention of choosing the optimizer.

OK, but after all these technicalities, I'd say the default tolerance values are reasonable, and if you get a problem there you're better off transforming your data instead (normalizing etc.).

good luck,
sven

_______________________________________________
Gretl-users mailing list
Gretl-users@lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users

_______________________________________________
Gretl-users mailing list
Gretl-users@lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users

Lee Adkins
Professor of Economics and Department Head
Economics and Legal Studies in Business
Oklahoma State University

Email: lee.adkins@okstate.edu

Fax: 405-744-5180

Phone: 405-744-5196

URL: www.learneconometrics.com