On the use of Particle Filters for Lagrangian Prediction
Arthur J. Mariano, Mike Chin
RSMAS, U. of Miami
(Abstract received 08/14/2009 for session C)
ABSTRACT
Particle Filter is a class of Sequential Monte Carlo methods that estimate the Probability Density Function (PDF) of the state variables of your system using data, prior PDF estimates, Bayes theorem, a Markov assumption, and quasi-random (re-)sampling. Particle Filters have very low developmental cost, but the efficiency of algorithm is problem dependent. At the heart of most of the algorithms is how to sample and resample the PDF (or a parameterization of it) so that peaks in multi-modal PDFs are resolved and PDF estimates are not too diffuse. Standard statistical, sampling-based particle filtering methods have not worked as well as expected for some geophysical prediction problems. A number of researchers have recently reported very good results for predicting Lagrangian trajectories using particle filtering methods. These recent results suggest that a more dynamical-based (re-)sampling and weighting strategy may be needed for geophysical prediction problems. A number of suggested sampling approaches will be presented for Lagrangian prediction problems.
<< Previous Abstract | C101 | C102 | C103 | C104 | C401 | C402 | C403 | Next Abstract >>