Young Investigators Seminar
◥Title: SDP Relaxation for Optimum Input/Experiment Design for System Identification
◥Speaker: Dr. Kwang-Ki Kim (recently graduated from the University of Illinois at Urbana-Champaign, Aerospace Engineering)
◥Date: 9 April, 2013 (Tue) 4PM
◥Place: Seminar room 1 (#1101) @ W1-3 Bldg.
Obtaining accurate models is valuable for predictive monitoring, state estimation, and control systems design.
In the 1970s--80s many approaches were proposed for the design of the inputs to a system to maximize the quality of the model fit to the resulting experimental data. Interest in such input design problems, also known in the literature as experiment design or experimental design, has resurged due to advances in optimization theory and algorithms with increasing computational power, especially in convex optimization, and model-based control. The objective of optimal input design is to compute a sequence of input signals in the time domain or an input spectrum in the frequency domain such that the covariance matrix of the parameter estimation error is minimized for a properly selected performance measure. Widely used optimization problems for experimental design for system identification are known to be nonconvex and NP-hard, and the problem data such as the Fisher information matrix depend on the true system parameters that are unknown a priori. To overcome such difficulties, this talk considers convex relaxations to compute suboptimal solutions for the original nonconvex problems and parameter adaptation methods to update Fisher information matrix with the estimated parameter that is time-varying and obtained from the available observables. The presented strategy of iterative input design is to compute an optimal input sequence within a finite horizon and use a receding horizon scheme for computing suboptimal input sequences. This design procedure can be considered as an iterative approximate dynamic programming for maximizing the performance of parameter estimation. It is also shown that similar design procedures can be extended to (a) frequency response estimation for which mixed time- and frequency-domain constraints over various types of estimation qualities are explicitly considered and (b) problems of optimal input design for process gain estimation of a class of structured large-scale systems.
*For international students, attendance of this seminar will be considered as attendance of department seminar during regular semester.