Title : Approximate Bayesian inference & model learning using particle filtering
Speaker : Saikat Saha, Linkoping University, Sweden
Time and Date : 4:00 pm, January 24, 2011
Venue : Room 201, Mechanical Engineering
Abstract
-------
Optimal estimation problems for general state space models do not typically admit a closed form solution. However, modern Monte Carlo methods have opened the door to solve such complex inference problems. Particle filters (aka sequential Monte Carlo) are a popular class of such Monte Carlo based Bayesian algorithms, which solve these estimation problems numerically in a recursive manner. Currently, this is a very active area of research in applied statistics. In this talk, I will start with a very brief introduction to Particle filter. Subsequently, I will present a selection of my recent research works which include
(a) Sequential learning of Gaussian noise statistics: In many scenarios, the noise/disturbance parameters, might not be known a priori and should be estimated on the run. We propose an efficient method to learn the Gaussian noise parameters using a new Rao-Blackwellized particle filtering approach.
(b) Parameter estimation for a general state space model from short observation data: one popular approach for parameter estimation is the maximum likelihood (ML) or associated Expectation maximization (EM) algorithm. Although ML (therefore, EM) is an asymptotically efficient estimator (unbiased with minimum variance), for short data records, where ML is indeed incapable of achieving its asymptotic optimality, estimator based on minimum MSE may be preferable. We propose a new particle based method in this direction.
(c) Estimation of stochastic volatility and associated parameters from stock data using Heston model with jumps.
Speaker Bio: Dr Saikat Saha is a postdoctoral researcher in the Electric Engineering Department at Linkoping University. He received a B.E. from Bengal Engineering and Science University, Shibpur; a MSc (Eng) from IISc, and a PhD from the University of Twente. Dr. Saha's research interests include particle filter/smoother for Bayesian inference, signal processing, and rare event simulation; parameter estimation, stochastic volatility, jumps models in finance and energy economics.