Learning the Network of a Linear Dynamical System from Ambient Fluctuations

Speaker : Deepjyoti Deka
Los Alamos National Lab
Date: 06/09/2017
Time: 2:00 pm - 3:00 pm
Location: LINCS Seminars room

Abstract

Linear Dynamical Systems are used to model several evolving network processes in biology, physical systems as well as financial networks. Estimation of the topology of a linear dynamical system is thus of interest for learning and control in diverse domains. This talk presents a useful framework for topology estimation for general linear dynamical networks (loopy or radial) using time-series measurements of nodal states. The learning framework utilizes multivariate Wiener filtering to unravel the interaction between fluctuations in states at different nodes and identifies operational edges by considering the phase response of the elements of the multivariate Wiener filter. The benefit derived by considering samples from ambient dynamics v/s steady state measurements will be described along with extensions, discussion of sample and computational complexity and open questions. In particular, we discuss one application related to topology estimation in power grids using samples of nodal voltage phase angles collected from the swing equations.