Closed-Loop Optimal Experiment Design: Solution via Moment Extension

• Published in: IEEE transactions on automatic control Vol. 60; no. 7; pp. 1731 - 1744
• Main Authors: Hildebrand, Roland, Gevers, Michel, Solari, Gabriel Elias
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.07.2015; The Institute of Electrical and Electronics Engineers, Inc. (IEEE); Institute of Electrical and Electronics Engineers
• Subjects: Approximation methods; Automatic; Closed-loop identification; Engineering Sciences; Joints; Manganese; Mathematics; MIMO; Optimal experiment design; Optimization and Control; Polynomials; Power spectral density; Signal and Image processing; Transfer functions; Vectors; power spectral density; Closed-loop identification; convex programming; optimal experiment design; moment method; Power spectral density; Optimal experiment design
• ISSN: 0018-9286; 1558-2523
• DOI: 10.1109/TAC.2015.2400662

We consider optimal experiment design for parametric prediction error system identification of linear time-invariant multiple-input multiple-output systems in closed-loop when the true system is in the model set. The optimization is performed jointly over the controller and the spectrum of the external excitation, which can be reparametrized as a joint spectral density matrix. The optimal solution consists of first computing a finite set of generalized moments of this spectrum as the solution of a semi-definite program. A second step then consists of constructing a spectrum that matches this finite set of optimal moments and satisfies some constraints due to the particular closed-loop nature of the optimization problem. This problem can be seen as a moment extension problem under constraints. Here we first show that the so-called central extension always satisfies these constraints, leading to a constructive procedure for the optimal controller and excitation spectrum. We then show that one can construct a broader set of parametrized optimal solutions that also satisfy the constraints; the additional degrees of freedom can then be used to achieve additional objectives.