Optimization Methods

[diggetybo]

Hey everyone,

I'm dealing with binary response models somewhat regularly and I'm trying to understand different convergence methods. Eviews seems to use newton-marquardt steps. To learn more about this I ran some search engine queries but returned very little, and what was returned was some pay on view pdf files or some extremely advanced proof of the procedure.

Can anyone explain what (if any) trade-offs or differences there are with marquardt steps and some other maximum likelihood iterative approaches, for instance, stochastic gradient ascent?

I'm just curious about the differences (if any) on a purely conceptual basis. I don't need any proofs or fancy derivations.