［ 2017年03月01日 ］
RIKEN International Symposium on Data Assimilation 2017
"A brief review of localization for particle filters"
Many particle filters rely on sequential importance sampling, a technique that can be shown to require ensemble sizes that are exponentially large in the system dimension for certain classes of high-dimensional problems with many observations. Localization, in which the update of ensemble members at any location depends only on a spatially local subset of the new observations, has proven extremely useful for ensemble Kalman filters and is one possible approximation that may reduce or avoid the difficulty of particle filtering for high-dimensional problems. I will review both the obstacles to sequential importance sampling in high dimensions and a subset of proposed algorithms for localization in particle filters.