Issues and Suggestions for Particle Filtering in High Dimensions

Mark Berliner
Department of Statistics, Ohio State University

Though Monte Carlo or ensemble based approximations to Bayesian sequential updating are well-known and readily implemented, in principle, they may exhibit problematic behavior in very high-dimensional settings. Specifically, the weights associated with ensemble members tend to over- concentrate on a few or even one ensemble member, leading to inefficient, highly variable results. Suggestions for dealing with this problem are discussed. The basic idea is to develop usable weights that avoid over-concentration, yet maintain reasonable value in indicating high and low probability ensemble members. To that end, we seek procedures based on the "informative" part of the data. Some potential strategies are reviewed. This is joint work with Chris Wikle.

Last Update: September 29, 2005