Methods for network identification robust with respect to uncertainties in a topology
Donatello Materassi, Murti V. Salapaka · 2016
Summary form only given. The paper tackles the problem of identifying an individual transfer function in a network of linear dynamical systems in the presence of loops under the assumptions that (i) only a subset of the nodes is observable, and (ii) data are being passively recorded (i.e. it is not possible to intervene on any part of the system by actively injecting an input). Such a scenario is often encountered in the study of many naturally occurring systems and is also motivated by operating networked systems where the injection of an external signal might lead to undesired disruptions. Sufficient conditions on which signals should be observable to guarantee the identifiability of a desired link are provided. The enabling result is the fact that the notion of d-separation for graphs (see [1], [2]) implies a notion of independence for networks of linear dynamical systems as well [3], allowing one to import identification methodologies developed in different areas. This talk exemplifies how to import methodologies from the domain of graphical models to solve control and identification problems.