Implicit sampling for data assimilation

Abstract:

Applications of filtering and data assimilation arise in engineering, geosciences, weather forecasting, and many other areas where one has to make estimations or predictions based on uncertain models supplemented by a stream of noisy data. For nonlinear problems, filtering can be very expensive because the number of particles required can be catastrophically large. We will present a nonlinear filtering scheme that is based on implicit sampling, a new sampling technique related to a chainless Monte Carlo method. This sampling strategy generates a particle (sample) beam which is focused on the high probability region of the target probability density function and the focusing makes the number of particles required manageable even if the state dimension is large. Several examples will be given.