Speaker
Dr
Jonathan ROUGIER
Description
ROUGIER: Data assimilation with uncertain static parameters
Abstract:
Using the link between variational methods and maximum
likelihood, I explore in a non-technical way the conditions
under which data assimilation produces consistent
estimators, and show that with a perfect model these
conditions distinguish clearly between learning about the
state vector and learning about the static parameters
('calibration'). Interestingly -- perhaps paradoxically --
the situation for calibration improves when the model is
imperfect, and correctly modelled as such. These results
are only a sketch in response to the discussions we have
been having about imperfect models, and the pros and cons of
stochastic models.