skip to content

Sequential Bayesian Inference in high-dimensional geoscience applications

Presented by: 
Peter Jan van Leeuwen
Tuesday 9th January 2018 - 14:30 to 15:30
INI Seminar Room 1
Applications of sequential Bayesian Inference in the geosciences, such as atmosphere, ocean, atmospheric chemistry, and
land-surface, are characterised by high dimensions, nonlinearity, and complex relations between system variables.
While Gaussian-based approximations such as Ensemble Kalman Filters and Smoothers and global variational
methods have been used quite extensively in this field, numerous problems ask for methods that can handle
strong nonlinearities. In this talk I will discuss recent progress using particle filters.

Three main areas of active research in particle filtering can be distinguished, exploring localisation,
exploring proposal densities, and exploring (optimal) transportation (and mergers of these ideas are on the
horizon). In localisation the idea is to split the high-dimensional problem in several smaller problems
that then need to be stitched together in a smart way. The first approximate applications of this methodology
have just made it to weather prediction, showing the exponentially fast developments here. However,
the ‘stitching’ problem remains outstanding. The proposal density methodology discussed next might be
fruitful to incorporate here.

In the proposal density approach one tries to evolve states in state space such that they obtain very similar weights
in the particle filter. Challenges are, of course, the huge dimensions, but these also provide opportunities via
the existence of typical sets, which lead to preferred parts of state space for the particles. Recent attempts to exploit
typical sets will be discussed.

Finally, we will discuss recent progress in (optimal) transportation. The idea here is that a set of prior particles
has to be transformed to a set of posterior particles. This is an old problem in optimal transportation. However,
the optimality condition poses unnecessary constraints, and by relaxing the optimality constraint we are able to
formulate new efficient methods. Specifically, by iteratively minimising the relative entropy between the probability
density of the prior particles and the posterior a sequence of transformations emerges for each particle that seems
to be tractable even for very high dimensional spaces. A new idea is to explore localisation to obtain a more
accurate description of the target posterior, but without the stitching issues mentioned above.

So far, model reduction techniques, emulation, and machine learning techniques have been unsuccessful for
these high-dimensional state estimation problems, but I’m keen to further understand the possibilities and limitations.
The video for this talk should appear here if JavaScript is enabled.
If it doesn't, something may have gone wrong with our embedded player.
We'll get it fixed as soon as possible.
University of Cambridge Research Councils UK
    Clay Mathematics Institute London Mathematical Society NM Rothschild and Sons