[7-2]

A targeted implicit particle filter

Javier Amezcua (U. of Reading) , Peter Jan van Leeuwen (U. of Reading)

We present a new particle filter algorithm -The Targeted Implicit Particle Filter (TIPF)- which ensures (i) that particles occupy regions of high posterior probability in state space and (ii) that particles have equal weights by construction (hence avoiding degeneracy). It combines successful ideas from the Equivalent Weights Particle Filter (setting target weights for particles) and the Implicit Particle Filter (an efficient implicit sampling procedure). In its simplest implementation, the TIPF does not require any resampling.

We show successful results from the application of this method to (i) a linear-Gaussian model (which allows a comparison against analytic Bayesian estimators), and the Lorenz 96 model with 40 and 400 variables.