Special EECS Colloquium

Friday, February 20, 2004
306 Soda Hall, Hewlett-Packard Auditorium
3:00-4:00 p.m.

Dr. Kameshwar Poolla

Mechanical Engineering Dept., UC Berkeley


System Identification for Non-parametric Structured Nonlinear Systems




In many system identification problems, we are forced to deal with non-parametric representations of system components. For example, static nonlinear elements are expanded using basis functions, and noise models are often chosen as unstructured coloring filters. Identification of such systems is very sensitive to the choice and number of basis functions. More complete basis choices yield high variance, while succinct bases yield high estimation bias. These identification problems are compounded when attempts are made to incorporate a priori structural information. In this talk we offer a novel approach for identification of structured systems that contain naturally non-parametric elements. Our technique involves a new dispersion metric of static nonlinear maps, together with Kalman smoothing problems. Structural constraints are readily incorporated in our methods. We discuss identifiability, convergence and computational aspects of our methods and also illustrate our techniques using an aero-elasticity modeling problem.


Professor Poolla received his Ph.D. from the University of Florida at Gainseville. He is currently a Professor in the Mechanical Engineering Department at UC Berkeley. His research interests lie in Robust multivariable control system synthesis, adaptive feedback systems, time-varying systems, control of flexible structures.