Nonlinear system identification in Sobolev spaces

Abstract

We consider the problem of deriving from experimental data an approximation of an unknown function, whose derivatives also approximate the unknown function derivatives. Solving this problem is useful, for instance, in the context of nonlinear system identification for obtaining models that are more accurate and reliable than the traditional ones, based on plain function approximation. Indeed, models identified by accounting for the derivatives can provide a better performance in several tasks, such as multi-step prediction, simulation, Nonlinear Model Predictive Control, and control design in general. In this paper, we propose a novel approach based on convex optimization, allowing us to solve the aforementioned identification problem. We develop an optimality analysis, showing that models derived using this approach enjoy suitable optimality properties in Sobolev spaces. The optimality analysis also leads to the derivation of tight uncertainty bounds on the unknown function and its derivatives. We demonstrate the effectiveness of the approach with two numerical examples. The first one is concerned with approximation of an univariate function, the second one with multi-step prediction of the Chua chaotic circuit.

Online Learning of Parameterized Uncertain Dynamical Environments with Finite-sample Guarantees

“…Now we employ our adaptive learning algorithm for the characterization of the uncertain vehicle states and learning of the unknown road-condition parameter e in real time. To do this, we take p = 3 predictors as in (7) with w ≡ 0, and i = 1, e1 = 0, e2 = 0, i = 2, e1 = 10, e2 = 0, i = 3, e1 = 0, e2 = 10.…”

“…The system identification literature broadly encompasses linear [4], [5] and non-linear systems [6], [7], with asymptotic performance guarantees. More recently, finite-sample analysis of identification methods have been proposed for linear systems [8]- [11].…”

“….To tighten the previous upper bound, consider any λ such that λ ∈ [0, T σe ). Then the following bound holds7…”

Online Optimization and Learning in Uncertain Dynamical Environments with Performance Guarantees

“…To deal with such challenges, data-driven approaches have regained attention. These methods are originated from system identification literature [16][17][18], which learn models from sufficient amount of properly collected data. More recently, Willem's fundamental lemma in Behavioral System Theory has been leveraged for system learning [19,20], as well as estimation [20,21] and predictive control [22][23][24][25].…”

NSM Converges to a k-NN Regressor Under Loose Lipschitz Estimates

“…Based purely on collected data-points, tight deterministic bounds on the function values f (x) at unobserved points can be established, making it an interesting modeling alternative. The theoretical foundations of the NSM technique were recently generalized to Hölder continuous functions [14,15], extended to a large class of problems [16], and also to more abstract function spaces [17]. Applications of this methodology can be found in the context of vehicle yaw reference-tracking [18], electrical microgrids scheduling [19], and approximating general linear MPC control laws [20].…”