Speaker
Josie König
Description
Solving data assimilation (DA) problems often involves expensive forward model simulations to compute the expected output for multiple initial conditions. Since this quickly becomes infeasible in high dimensions, reduced models can play an important role in making such computations tractable. In particular, model reduction methods from control (systems) theory are well established for obtaining efficient reduced models while preserving the map from control input to observed output. In this talk, we present a new interpretation of the Bayesian inverse problem, which is essential in DA, as a control problem. We then discuss how established systems-theoretic methods for model order reduction can be adapted to this new setting.
