Non-Factorised Variational Inference in Dynamical Systems

12/14/2018

∙

by   Alessandro Davide Ialongo, et al.

∙

University of Cambridge

∙

PROWLER.io

∙

4

∙

share

We focus on variational inference in dynamical systems where the discrete time transition function (or evolution rule) is modelled by a Gaussian process. The dominant approach so far has been to use a factorised posterior distribution, decoupling the transition function from the system states. This is not exact in general and can lead to an overconfident posterior over the transition function as well as an overestimation of the intrinsic stochasticity of the system (process noise). We propose a new method that addresses these issues and incurs no additional computational costs.