Skip to main content

AS Seminar: Bayesian transfer learning: FPD-optimal knowledge transfer between linear state processes with uniformly distributed noise


A filter estimates values of unobservable quantity called state. The state is time-varying, which is modelled by time evolution model. The value of the state is transformed into observation (output of the system), which is modelled by observation model. Both the models are linear and their noises are distributed uniformly.

We consider two filters, source and target, estimating the state processes driven stochastically by a common physical process behind. The only information, which target modeller yields from the source, is a predictor (state of data). In this case of incomplete modelling, we use fully probablistic design as a method to transfer information about the estimated quantity in the case of unknown model and observations (data) of the source.

We present theoretical foundations and the results for uniform distribution and orthotopic approximation of the distribution support, comparison with a complete modelling, studies of robustness to modelling mismatch, problems and perspectives.

Submitted by neuner on