Oberseminar "Mathematik des Maschinellen Lernens und Angewandte Analysis" - Dr. Philipp Wacker

Continuous stochastic filtering provides a mathematically sound recipe for assimilating a data stream for the objective of inferring the state (or parameters) of a latent time-continuous stochastic process. We will talk about a way of generalising this setting to a scenario where we can allocate a fixed computational budget to time periods that promise more informative data. Techniques from optimal control give us a computationally convenient way of obtaining derivatives of common objective functions mediating the "expected utility" of future data for a proposed sensor budget, which allows efficient gradient-based optimisation of sensor allocation.