Consistency of Penalized Likelihood Estimation for Mixture-of-Experts
15 Dezembro 2011 ·
- Quem: Eduardo Mendes -- University of Chicago
- Onde: FGV, sala 317
- Quando: 15 de Dezembro de 2011 às 16:00h
We consider the estimation of the number of components on a mixture-of-experts (ME) model, using penalized maximum likelihood. We show that, under mild conditions on the mixture densities, a large class of penalization criteria enjoy the consistency property. This wide class of methods encompasses the Bayesian Information Criterion (BIC), which is the most frequently used method in the ME literature. We close the paper with a few examples of densities and penalty functions satisfying the assumptions.
Nota para visitantes
A presença é gratuíta e não exige confirmação. A FGV não permite a entrada de homens vestindo bermuda ou chinelo.
Tags: