Finite sample issues in system identification, change detection and filtering
Models of dynamical systems are commonly used in many fields of science and engineering. In order to make proper use of a model, it is important to know the uncertainties associated with it. An evaluation of the model quality must necessarily be based on a finite number of data points from the true system. In this project we have developed methodologies: Leave-out Sign dominant Correlation Regions (LSCR) and Sign Perturbed Sums (SPS), which stand on a solid theoretical footing in a finite sample context and which deliver useful evaluations of the model uncertainties. In particular, it delivers a probabilistically guaranteed confidence set for the true system parameters for any finite number of data points under very weak assumptions about the noise processes affecting the system. The LSCR and SPS principles are very powerful and they also find applications in filtering and fault detection, and filtering and detection algorithms with guaranteed statistical properties for any finite number of data points that have been developed.
Current research is focused on further developments of the methods and their properties for different types of linear and nonlinear models and identification algorithms (e.g. least squares, prediction error methods, instrumental variables), efficient numerical implementation of the methods and applications of the principles in different areas.
Leader: Erik Weyer
Staff: Erik Weyer
Collaborators: Marco Campi (University of Brescia, Italy), Balazs Csaji (MTA SZTAKI: Institute for Computer Science and Control, Hungarian Academy of Sciences), Algo Care (MTA SZTAKI: Institute for Computer Science and Control, Hungarian Academy of Sciences)
Electrical & Electronic Engineering
Networks and data in society
Signal processing; signals and systems; System identification