Speaker
Description
We propose an approach for the L-infinity model reduction of descriptor systems based on smooth optimization techniques. A direct application of smooth optimization techniques for L-infinity model reduction does not seem suitable, as they converge linearly at best for this highly nonsmooth problem, and require the computation of the costly L-infinity norm objective too many times. Instead, we replace the original system with a surrogate system of smaller order interpolating the original system, minimize the L-infinity error objective for the small system, and refine the small system based on this minimization. The numerical results on benchmark examples illustrate that approach leads to locally optimal reduced order systems with respect to the L-infinity objective, and can deal with systems of order a few ten thousands effectively.
