ACDC Seminar by Kasper Bågmark on the 15th of October

On the 15th of October at 16.00, Kasper Bågmark will give an Amsterdam Dynamics Seminar. We will go for drinks at Bar Boele after the talk.

The talk will take place in the Maryam seminar room (9A-46).

Title:  Nonlinear filtering based on density approximation and deep BSDE prediction

Abstract: The problem of estimating the probability density of a continuous state given noisy measurements is called the filtering problem. In the case when the system of states and observations is nonlinear the problem cannot be solved analytically (except in a few special cases). Classical methods, namely particle filters, suffer under the curse of dimensionality in the underlying dimension of the state space. Deep learning is a powerful tool in creating scalable approximations for similar problems. In the last few years we have examined methods that exploits the Fokker--Planck equation that is related to state equations governed by Stochastic Differential Equations (SDE). This work concerns using the Deep Backward SDE approach to solve the Fokker--Planck equation. The obtained approximation, after training, is a fast online filter. We examine the convergence of the method analytically and empirically, and show numerical examples compared to the bootstrap particle filter, in particular we demonstrate the method on a toy example up to state dimension 100.

You can subscribe to the ACDC calendar using this link