14:00

G.O. Roberts, University of Warwick

An introduction to gradient based MCMC algorithms

I will give an introduction to MCMC algorithms which try to converge quickly by taking advantage of local derivative information. Many of the algorithms discussed will be motivated by related continuous time processes such as diffusions. So the talk will first introduce diffusions which have desired stationarity properties, and go on to show how MCMC methods can be built out of approximations for these diffusions.

15:30

Sylvain Le Corff, Université Paris-Sud, CNRS, Orsay, France

Some recent results on the identifiability and the estimation of nonparametric hidden Markov models

In this talk, we are interested in recent advances for the inference of nonparametric hidden Markov models. In this setting, the Markovian states are only available partially through an observation sequence. We focus on nonparametric regression models with latent regressors (general state space) and model with nonparametric conditional distributions of the observations given the states (finite state space). We first present some identifiability results that allow to estimate these models based on the observation sequence under various assumptions. Then, we propose several estimation procedures (Expectation Maximization, kernel methods, spectral methods) along with convergence results.

The performance of these algorithms are illustrated with simulated data and real data sets (oscillation processes).