Session: Ensemble Kalman filter

[rossmcv]

The ensemble Kalman filter was introduced in Evans (1994) and has become an important technique in high dimensional forecasting. The method has connections to approximate Bayesian computation (Nott et al, 2012), McKean-Vlaslov diffusions (Del Moral and Tugaut, 2016). I would be interested in hearing of anyone's experience in the application of the ensemble Kalman filter.

BTW, I have no experience with this.

Del Moral and Tugaut (2016) On the stability and the uniform propagation of chaos properties of Ensemble Kalman-Bucy filters, arXiv:1605.09329 -- to appear in Annals of Applied Probability.

Evensen (1994) Sequential data assimilation with a nonlinear quasi-geostrophic model using Monte Carlo methods to forecast error statistics, Journal of Geophysical research, 99, 10143–10162

Nott, Marshall and Ngoc (2012), The ensemble Kalman filter is an ABC algorithm, Statistics and Computing, 22, 1273-1276.

[jesse-jesse]

I don't know if this is relevant, but Lewis Mitchell from Uni Adelaide presented this at the recent BOM + ACEMS workshops . https://github.com/ACEMS/ACEMS-BOM2017/blob/master/Mitchell_modelErrorTalk2017.pdf
There is this paper that goes with the talk. - https://github.com/ACEMS/ACEMS-BOM2017/blob/master/Mitchell_QJRMS_2015.pdf. Maybe Lewis has some knowledge to share on this topic

[lewismath]

Hey Ross, Jesse forwarded your question on to me. I did my PhD on ensemble Kalman filtering, albeit from a more dynamical systems/geosciences point of view -- I looked at the application of particular flavours of the EnKF (Evensen's original paper there is just the first of many many variants) to systems with multiscale chaotic dynamics, and ways of controlling various numerical issues that arise with EnKFs in high-dimensional geophysical systems.

I didn't discover ABC and its friends until a couple of years ago, so I don't know all that much about the connections to EnKF, outside of some googling which turned up that same Nott et al paper (which I should really read one day). But I agree that there appear to be numerous connections, and I've been thinking (particularly after attending the ACEMS/BoM workshop and ACEMS review last month) that EnKF fits quite well within ACEMS themes.

Now, I'm not sure what exactly to do with all this information at this stage. But I can at least point you towards a couple of good references, in the atmospheric/geosciences literature anyway...

• the best book is the one by Eugenia Kalnay (https://www.amazon.com/Atmospheric-Modeling-Data-Assimilation-Predictability/dp/0521796296)
• a great review article just appeared a few days ago, by some very good people (inc Evensen): https://arxiv.org/abs/1709.02798
• if you use MATLAB, then Pavel Sakov has good code for a bunch of different EnKF variants and a bunch of different (deterministic) dynamical models to go with Evensen's book (http://enkf.nersc.no), here: http://enkf.nersc.no/Code/EnKF-Matlab/

You could probably do worse if you wanted to learn than having a play around with that code and reading up as necessary. Thinking out loud now: I believe Sakov still lives in Australia (Tasmania?) -- maybe a good first move is to invite him to an ACEMS meeting (like an ECR retreat) to talk about this stuff!

[lewismath]

Oh, TIL there was a data assimilation summer school in 2016, and Pavel Sakov is now at the BoM in Melbourne: http://data-assimilation.com/speaker/pavel-sakov-meteorological-bureau-melbourne-australia/

Seems like an obvious invite to me!