On the Sample Complexity of Data-Driven Inference of the L2-Gain

Lately, data-driven control has become a widespread area of research. A few recent big-data based approaches for data-driven control of nonlinear systems try to use classical input-output techniques to design controllers for systems for which only a finite number of (input-output) samples are known. These methods focus on using the given data to compute bounds on the $\mathcal {L}_{2}$ -gain or on the shortage of passivity from finite input-output data, allowing for the application of the small gain theorem or the feedback theorem for passive systems. One question regarding these methods asks about their sample complexity, namely how many input-output samples are needed to get an approximation of the operator norm or of the shortage of passivity. We show that the number of samples needed to estimate the operator norm of a system is roughly the same as the number of samples required to approximate the system in the operator norm.

[1]  Petre Stoica,et al.  Decentralized Control , 2018, The Control Systems Handbook.

[2]  Nikolai Matni,et al.  On the Sample Complexity of the Linear Quadratic Regulator , 2017, Foundations of Computational Mathematics.

[3]  L. Ljung,et al.  Subspace-based multivariable system identification from frequency response data , 1996, IEEE Trans. Autom. Control..

[4]  John Lygeros,et al.  Data-Enabled Predictive Control: In the Shallows of the DeePC , 2018, 2019 18th European Control Conference (ECC).

[5]  Pietro Tesi,et al.  Data-based guarantees of set invariance properties , 2019, IFAC-PapersOnLine.

[6]  Biao Huang,et al.  System Identification , 2000, Control Theory for Physicists.

[7]  Zhuo Wang,et al.  From model-based control to data-driven control: Survey, classification and perspective , 2013, Inf. Sci..

[8]  Frank Allgöwer,et al.  One-Shot Verification of Dissipativity Properties From Input–Output Data , 2019, IEEE Control Systems Letters.

[9]  R. Dudley.  The Sizes of Compact Subsets of Hilbert Space and Continuity of Gaussian Processes , 1967 .

[10]  Bart De Moor,et al.  A note on persistency of excitation , 2005, 2004 43rd IEEE Conference on Decision and Control (CDC) (IEEE Cat. No.04CH37601).

[11]  Frank Allgöwer,et al.  Some problems arising in controller design from big data via input-output methods , 2016, 2016 IEEE 55th Conference on Decision and Control (CDC).

[12]  Frank Allgöwer,et al.  Data-driven inference of passivity properties via Gaussian process optimization , 2019, 2019 18th European Control Conference (ECC).

[13]  Benjamin Recht,et al.  A Tour of Reinforcement Learning: The View from Continuous Control , 2018, Annu. Rev. Control. Robotics Auton. Syst..

[14]  M. Anusha,et al.  Big Data-Survey , 2016 .

[15]  Pietro Tesi,et al.  Formulas for Data-Driven Control: Stabilization, Optimality, and Robustness , 2019, IEEE Transactions on Automatic Control.

[16]  Paulo Tabuada,et al.  Data-driven control for feedback linearizable single-input systems , 2017, 2017 IEEE 56th Annual Conference on Decision and Control (CDC).

[17]  Shai Ben-David,et al.  Understanding Machine Learning: From Theory to Algorithms , 2014 .

[18]  Arseniy Akopyan,et al.  A constructive proof of Kirszbraun’s theorem , 2008 .

[19]  Shun-ichi Azuma,et al.  Efficient Data-Driven Estimation of Passivity Properties , 2019, IEEE Control Systems Letters.

[20]  Frank Allgöwer,et al.  Sampling strategies for data-driven inference of passivity properties , 2017, 2017 IEEE 56th Annual Conference on Decision and Control (CDC).

[21]  Frank Allgöwer,et al.  Determining dissipation inequalities from input-output samples , 2017 .

[22]  P. Olver.  Nonlinear Systems , 2013 .

[23]  Lennart Ljung,et al.  Nonlinear black-box modeling in system identification: a unified overview , 1995, Autom..

[24]  Albert Wilansky,et al.  Topology for Analysis , 1970 .