Bayesian Statistics Seminar
North Carolina State University

presents

Dr. Hedibert Lopes

University of Chicago

"Particle Learning and Smoothing"

ABSTRACT

This paper provides novel particle learning (PL) methods for sequential parameter learning and smoothing in state space models with non-normal errors, non-linear observation equations, and non-linear state evolutions. The methods extend existing particle methods by incorporating unknown parameters, utilizing sufficient statistics, for the parameters and/or the states, and allowing for nonlinearities in the state and/or observation equation. We also show how to solve the state smoothing problem, integrating out parameter uncertainty. Previously, the only approach available for this marginal smoothing problem is MCMC. We show that our algorithms outperform MCMC, as well existing particle filtering algorithms such as the mixture Kalman filter.

This is a joint work with C. Carvalho, M. Johannes and N. Polson

Tuesday, November, 18, 2008

4:00 - 5:00 pm

208 Patterson Hall

           Back to BSWG Home