Natural Gradient Gaussian Filtering
Organizer
Speaker
Wenhan Cao
Time
Wednesday, July 10, 2024 3:00 PM - 3:30 PM
Venue
理科楼A-304
Abstract
Recent Gaussian filtering algorithms typically involve two steps: (1) linearizing the model, and (2) performing the Kalman filter update. However, linearization often fails to adequately handle high nonlinearity, non-Gaussianity, and measurement outliers. In this work, we demonstrate that the update step in Bayesian filtering can be viewed as a variational problem. When parameterized with a Gaussian distribution, this variational problem transforms into a parameter optimization problem that can be effectively solved using natural gradient descent. This approach significantly enhances filtering accuracy in nonlinear and non-Gaussian systems, as well as in the presence of measurement outliers, while maintaining an acceptable computational burden.