The equivalent-weights particle filter

Melanie Ades, University of Reading

The majority of data assimilation schemes rely on linearity assumptions. However as the resolution and complexity of both the numerical models and observations increases, these linearity assumptions become less appropriate. Particle filters are a nonlinear data assimilation method that avoid the need for such assumptions and hence can represent the full posterior pdf. Unfortunately standard particle filters suffer from filter degeneracy which makes them inapplicable in high dimensional systems. The equivalent weights particle filter is an adaptation to the standard particle filter which theoretically gives a representation of the true posterior pdf, regardless of the number of observations or the dimension of the system. It uses proposal densities to ensure that particles are both close to the observations at analysis time and have equivalent significance when estimating the posterior pdf.

Here the formulation of the equivalent-weights particle filter and its relation to both the standard particle filter and optimal proposal density is presented. Detail is given about the two constituent parts of the scheme and how the information from the observations is used. It is demonstrated that filter degeneracy does not occur with the scheme in a 65,000 variable barotropic vorticity model and so consideration is given as to how well the scheme is able to represent the true posterior pdf. I show that the representation of the posterior pdf is not very sensitive to the number of particles used but that performance measures like rank histograms can be significantly affected by the tuning parameters of the scheme. I conclude with potential directions for future research.

Thursday March 20, 2014

FL2-1001 Small Seminar Room

11:00am