Lunds Tekniska Högskola

Denna sida på svenska This page in English


Här hittar du samlade händelser framöver inom LTHs tekniska och naturvetenskapliga områden, många med bäring på kommunikationsteknik. Till höger har du direktlänk till kalendern för Elektro- och Informationsteknik.

Kalendarium för kommunikationsteknik

Statistics Seminar, "Maximum likelihood for state-space models using an ABC sequential Monte Carlo filter", Umberto Picchini, Lund


Tid: 2017-11-17 13:15 till: 14:00

State-space models (SSMs) are an important class of partially observed models, involving a latent (unobserved) component. Inference for SSMs has generated a very large amount of literature in the past four decades, both for state and parameter inference. Here we consider maximum likelihood estimation for the parameters of general nonlinear/non-Gaussian SSMs. However, as we illustrate, this is coupled with the problem of performing inference for the latent state of the SSM. We consider a stochastic approximation expectation-maximization (SAEM) algorithm to maximize the parameters likelihood function with the novelty of using approximate Bayesian computation (ABC) within SAEM. The task is to provide each iteration of SAEM with a filtered state of the system, and this is achieved using an ABC sampler for the hidden state, based on particle filters methodology. It is shown that the resulting SAEM-ABC algorithm can be calibrated to return accurate inference, and in some situations it can outperform a version of SAEM using the "bootstrap filter". We show how the 
simple bust still widely used bootstrap filter  can produce poor sampled paths, especially when few particles are employed. Instead a simple modification using ABC assigns high weights only to very "important" particles, allowing better paths to be selected, this in turn producing less biased parameter estimation.

[joint work with Adeline Samson, Universite Grenoble Alpes, published on "Computational Statistics"]



Följande disputationer är på gång:


27 Nov: Iman Vakili. Opponent: Prof. Andrea Neto