Regularized techniques, also named as kernel-based techniques, are the major advances in system identification in the last decade. Although many promising results have been achieved, their theoretical analysis is far from complete and there are still many key problems to be solved. One of them is the asymptotic theory, which is about convergence properties of the model estimators as the sample size goes to infinity. The existing related results for regularized system identification are about the almost sure convergence of various hyperparameter estimators. A common problem of those results is that they do not contain information on the factors that affect the convergence properties of those hyperparameter estimators, e.g., the regression matrix. In this article, we tackle problems of this kind for the regularized finite impulse response model estimation with the empirical Bayes (EB) hyperparameter estimator and filtered white noise input. In order to expose and find those factors, we study the convergence in distribution of the EB hyperparameter estimator, and the asymptotic distribution of its corresponding model estimator. For illustration, we run Monte Carlo simulations to show the efficacy of our obtained theoretical results.

In this article, we consider linear quadratic (LQ) control problem with process and measurement noises. We analyze the LQ problem in terms of the average cost and the structure of the value function. We assume that the dynamics of the linear system is unknown and only noisy measurements of the state variable are available. Using noisy measurements of the state variable, we propose two model-free iterative algorithms to solve the LQ problem. The proposed algorithms are variants of policy iteration routine where the policy is greedy with respect to the average of all previous iterations. We rigorously analyze the properties of the proposed algorithms, including stability of the generated controllers and convergence. We analyze the effect of measurement noise on the performance of the proposed algorithms, the classical off-policy, and the classical Q-learning routines. We also investigate a model-building approach, inspired by adaptive control, where a model of the dynamical system is estimated and the optimal control problem is solved assuming that the estimated model is the true model. We use a benchmark to evaluate and compare our proposed algorithms with the classical off-policy, the classical Q-learning, and the policy gradient. We show that our model-building approach performs nearly identical to the analytical solution and our proposed policy iteration based algorithms outperform the classical off-policy and the classical Q-learning algorithms on this benchmark but do not outperform the model-building approach.

The paper contains a discussion of earlier work on Total Model Errors and Model Validation. It is maintained that the recent change of paradigm to kernel based system identification has also affected the basis for (and interest in) giving bounds for the total model error.

In this article, a unified identification framework called constrained subspace method for structured state-space models (COSMOS) is presented, where the structure is defined by a user-specified linear or polynomial parametrization. The new approach operates directly from the input and output data, which differs from the traditional two-step method that first obtains a state-space realization followed by the system-parameter estimation. The new identification framework relies on a subspace inspired linear regression problem which may not yield a consistent estimate in the presence of process noise. To alleviate this problem, the linear regression formulation is imposed by structured and low-rank constraints in terms of a finite set of system Markov parameters and the user specified model parameters. The nonconvex nature of the constrained optimization problem is dealt with by transforming the problem into a difference-of-convex optimization problem, which is then handled by the sequential convex programming strategy. Numerical simulation examples show that the proposed identification method is more robust than the classical prediction-error method initialized by random initial values in converging to local minima, but at the cost of heavier computational burden.

System identification is a mature research area with well established paradigms, mostly based on classical statistical methods. Recently, there has been considerable interest in so called kernel-based regularisation methods applied to system identification problem. The recent literature on this is extensive and at times difficult to digest. The purpose of this contribution is to provide an accessible account of the main ideas and results of kernel-based regularisation methods for system identification. The focus is to assess the impact of these new techniques on the field and traditional paradigms.

n/a

The problem of estimating continuous-time model parameters of linear dynamical systems using sampled time-domain input and output data has received considerable attention over the past decades and has been approached by various methods. The research topic also bears practical importance due to both its close relation to first principles modelling and equally to linear model-based control design techniques, most of them carried in continuous time. Nonetheless, as the performance of the existing algorithms for continuous-time model identification has seldom been assessed and, as thus far, it has not been considered in a comprehensive study, this practical potential of existing methods remains highly questionable. The goal of this brief paper is to bring forward a first study on this issue and to factually highlight the main aspects of interest. As such, an analysis is performed on a benchmark designed to be consistent both from a system identification viewpoint and from a control-theoretic one. It is concluded that robust initialization aspects require further research focus towards reliable algorithm development. (C) 2019 Elsevier Ltd. All rights reserved.

This paper proposes a new method to identify nonlinear state-space systems from heterogeneous datasets. The method is described in the context of identifying biochemical/gene networks (i.e., identifying both reaction dynamics and kinetic parameters) from experimental data. Simultaneous integration of various datasets has the potential to yield better performance for system identification. Data collected experimentally typically vary depending on the specific experimental setup and conditions. Typically, heterogeneous data are obtained experimentally through 1) replicate measurements from the same biological system or 2) application of different experimental conditions such as changes/perturbations in biological inductions, temperature, gene knock-out, gene over-expression, etc. We formulate here the identification problem using a Bayesian learning framework that makes use of “sparse group” priors to allow inference of the sparsest model that can explain the whole set of observed heterogeneous data. To enable scale up to large number of features, the resulting nonconvex optimization problem is relaxed to a reweighted Group Lasso problem using a convex-concave procedure. As an illustrative example of the effectiveness of our method, we use it to identify a genetic oscillator (generalized eight species repressilator). Through this example we show that our algorithm outperforms Group Lasso when the number of experiments is increased, even when each single time-series dataset is short. We additionally assess the robustness of our algorithm against noise by varying the intensity of process noise and measurement noise.

Identification of structured state-space (gray-box) model is popular for modeling physical and network systems. Due to the non-convex nature of the gray-box identification problem, good initial parameter estimates are crucial for successful applications. In this paper, the non-convex gray-box identification problem is reformulated as a structured low-rank matrix factorization problem by exploiting the rank and structured properties of a block Hankel matrix constructed by the system impulse response. To address the low-rank optimization problem, it is first transformed into a difference-of-convex (DC) formulation and then solved using the sequentially convex relaxation method. Compared with the classical gray-box identification methods like the prediction-error method (PEM), the new approach turns out to be more robust against converging to non-global minima, as supported by a simulation study. The developed identification can either be directly used for gray-box identification or provide an initial parameter estimate for the PEM. (C) 2018 Elsevier Ltd. All rights reserved.

The local approach to linear parameter varying (LPV) system identification consists in interpolating individually estimated local linear time invariant (LTI) models corresponding to fixed values of the scheduling variable. It is shown in this paper that, without any global structural assumption of the considered LPV system, individually estimated local state-space LTI models do not contain sufficient information for determining similarity transformations making them coherent. It is possible to estimate these similarity transformations from input-output data under appropriate excitation conditions. (C) 2017 Published by Elsevier Ltd.

Inspired by ideas taken from the machine learning literature, new regularization techniques have been recently introduced in linear system identification. In particular, all the adopted estimators solve a regularized least squares problem, differing in the nature of the penalty term assigned to the impulse response. Popular choices include atomic and nuclear norms (applied to Hankel matrices) as well as norms induced by the so called stable spline kernels. In this paper, a comparative study of estimators based on these different types of regularizers is reported. Our findings reveal that stable spline kernels outperform approaches based on atomic and nuclear norms since they suitably embed information on impulse response stability and smoothness. This point is illustrated using the Bayesian interpretation of regularization. We also design a new class of regularizers defined by "integral" versions of stable spline/TC kernels. Under quite realistic experimental conditions, the new estimators outperform classical prediction error methods also when the latter are equipped with an oracle for model order selection. (C) 2016 Elsevier Ltd. All rights reserved.

An established method for grey-box identification is to use maximum-likelihood estimation for the nonlinear case implemented via extended Kalman filtering. In applications of (nonlinear) model predictive control a more and more common approach for the state estimation is to use moving horizon estimation, which employs (nonlinear) optimization directly on a model for a whole batch of data. This paper shows that, in the linear case, horizon estimation may also be used for joint parameter estimation and state estimation, as long as a bias correction based on the Kalman filter is included. For the nonlinear case two special cases are presented where the bias correction can be determined without approximation. A procedure how to approximate the bias correction for general nonlinear systems is also outlined. (C) 2015 Elsevier Ltd. All rights reserved.

Anomaly detection in large populations is a challenging but highly relevant problem. It is essentially a multi-hypothesis problem, with a hypothesis for every division of the systems into normal and anomalous systems. The number of hypothesis grows rapidly with the number of systems and approximate solutions become a necessity for any problem of practical interest. In this paper we take an optimization approach to this multi-hypothesis problem. It is first shown to be equivalent to a non-convex combinatorial optimization problem and then is relaxed to a convex optimization problem that can be solved distributively on the systems and that stays computationally tractable as the number of systems increase. An interesting property of the proposed method is that it can under certain conditions be shown to give exactly the same result as the combinatorial multi-hypothesis problem and the relaxation is hence tight.

This paper gives an overview of the identificationof linear systems. It covers the classical approach ofparametric methods by means of maximum likelihood andpredicion error methods, as well all classical non-parametricmethods through spectral analysis. It also covers very recenttechniques dealing with convex formulations by regularizationof FIR and ARX models, as well as new alternatives tospectral analysis, through local linear models. An example of identification of aircraft dynamics illustrates the approaches.

This paper proposes a general convex framework for the identification of switched linear systems. The proposed framework uses over-parameterization to avoid solving the otherwise combinatorially forbidding identification problem, and takes the form of a least-squares problem with a sum-of-norms regularization, a generalization of the ℓ₁-regularization. The regularization constant regulates the complexity and is used to trade off the fit and the number of submodels.

Intrigued by some recent results on impulse response estimation by kernel and nonparametric techniques, we revisit the old problem of transfer function estimation from input-output measurements. We formulate a classical regularization approach, focused on finite impulse response (FIR) models, and find that regularization is necessary to cope with the high variance problem. This basic, regularized least squares approach is then a focal point for interpreting other techniques, like Bayesian inference and Gaussian process regression. The main issue is how to determine a suitable regularization matrix (Bayesian prior or kernel). Several regularization matrices are provided and numerically evaluated on a data bank of test systems and data sets. Our findings based on the data bank are as follows. The classical regularization approach with carefully chosen regularization matrices shows slightly better accuracy and clearly better robustness in estimating the impulse response than the standard approach - the prediction error method/maximum likelihood (PEM/ML) approach. If the goal is to estimate a model of given order as well as possible, a low order model is often better estimated by the PEM/ML approach, and a higher order model is often better estimated by model reduction on a high order regularized FIR model estimated with careful regularization. Moreover, an optimal regularization matrix that minimizes the mean square error matrix is derived and studied. The importance of this result lies in that it gives the theoretical upper bound on the accuracy that can be achieved for this classical regularization approach.

Model-based engineering becomes more and more important in industrial practice. System identification is a vital technology for producing the necessary models, and has been an active area of research and applications in the automatic control community during half a century. At the same time, increasing demands require the area to constantly develop and sharpen its tools. This paper deals with how system identification does that by amalgamating concepts, features and methods from other fields. It describes encounters with four areas in systems theory and engineering: Networked Systems, Particle Filtering Techniques, Sparsity and Compressed Sensing, and Machine Learning. The impacts on System Identification methodology by these encounters are described and illustrated.

The particle filter has become an important tool in solving nonlinear filtering problems for dynamic systems. This correspondence extends our recent work, where we proved that the particle filter converges for unbounded functions, using L⁴-convergence. More specifically, the present contribution is that we prove that the particle filter converge for unbounded functions in the sense of Lp-convergence, for an arbitrary p ≥ 2.

Segmentation of time-varying systems and signals into models whose parameters are piecewise constant in time is an important and well studied problem. Here it is formulated as a least-squares problem with sum-of-norms regularization over the state parameter jumps. a generalization of L1-regularization. A nice property of the suggested formulation is that it only has one tuning parameter, the regularization constant which is used to trade-off fit and the number of segments.

The standard continuous time state space model with stochastic disturbances contains the mathematical abstraction of continuous time white noise. To work with well defined, discrete time observations, it is necessary to sample the model with care. The basic issues are well known, and have been discussed in the literature. However, the consequences have not quite penetrated the practice of estimation and identification. One example is that the standard model of an observation, being a snapshot of the current state plus noise independent of the state, cannot be reconciled with this picture. Another is that estimation and identification of time continuous models require a more careful treatment of the sampling formulas. We discuss and illustrate these issues in the current contribution. An application of particular practical importance is the estimation of models based on irregularly sampled observations.

This paper treats several aspects relevant to identification of continuous-time Output error (OE) models based on sampled data. The exact method for doing this is well known both for data given in the time and frequency domains. This approach becomes somewhat complex, especially for non-uniformly sampled data. We study various ways to approximate the exact method for reasonably fast sampling. While an objective is to gain insights into the non-uniform sampling case, this paper only gives explicit results for uniform sampling.

A promising method for estimation of the time-delay in continuous-time linear dynamical systems uses the phase of the all-pass part of a discrete-time model of the system. We have discovered that this method can sometimes fail totally and we suggest a method for avoiding such failures.

Regressor selection can be viewed as the first step in the system identification process. The benefits of finding good regressors before estimating complex models are especially clear for nonlinear systems, where the class of possible models is huge. In this article, a structured way of using the tool analysis of variance (ANOVA) is presented and used for NARX model (nonlinear autoregressive model with exogenous input) identification with many candidate regressors.

The basic nonlinear filtering problem for dynamical systems is considered. Approximating the optimal filter estimate by particle filter methods has become perhaps the most common and useful method in recent years. Many variants of particle filters have been suggested, and there is an extensive literature on the theoretical aspects of the quality of the approximation. Still a clear-cut result that the approximate solution, for unbounded functions, converges to the true optimal estimate as the number of particles tends to infinity seems to be lacking. It is the purpose of this contribution to give such a basic convergence result for a rather general class of unbounded functions. Furthermore, a general framework, including many of the particle filter algorithms as special cases, is given.

This paper presents a parameter estimation of continuous-time polytopic models for a linear parameter varying (LPV) system. The prediction error method of linear time invariant (LTI) models is modified for polytopic models. The modified prediction error method is applied to an LPV aircraft system that has flight velocity as the varying parameter and stability and control derivatives (SCDs) as the model parameters. In an identification simulation, the polytopic model is more suitable for expressing the behaviors of the LPV aircraft system than the LTI model from the viewpoints of time and frequency responses. The SCDs of the initial polytopic model are adjusted to fit the model output to the output-data obtained from the LPV aircraft system.

Nonlinear systems can be approximated by linear time-invariant (LTI) models in-many ways. Here, LTI models that are optimal approximations in the mean-square error sense are analyzed. A necessary and sufficient condition on the input signal for the optimal LTI approximation of an arbitrary nonlinear finite impulse response (NFIR) system to be a linear finite impulse response (FIR) model is presented. This condition says that the in ut should be separable of a certain order, i.e., that certain conditional expectations should be,P linear. For the special case of Gaussian input signals, this condition is closely related to a generalized version of Bussgang's classic theorem about static nonlinearities. It is shown that this generalized theorem can be used for structure identification and for the identification of generalized Wiener-Hammerstein systems.

A general framework for estimating nonlinear functions and systems is described and analyzed in this paper. Identification of a system is seen as estimation of a predictor function. The considered predictor function estimate at a particular point is defined to be affine in the observed outputs, and the estimate is defined by the weights in this expression. For each given point, the maximal mean-square error (or an upper bound) of the function estimate over a class of possible true functions is minimized with respect to the weights, which is a convex optimization problem. This gives different types of algorithms depending on the chosen function class. It is shown how the classical linear least squares is obtained as a special case and how unknown-but-bounded disturbances can be handled. Most of the paper deals with the method applied to locally smooth predictor functions. It is shown how this leads to local estimators with a finite bandwidth, meaning that only observations in a neighborhood of the target point will be used in the estimate. The size of this neighborhood (the bandwidth) is automatically computed and reflects the noise level in the data and the smoothness priors. The approach is applied to a number of dynamical systems to illustrate its potential.

This paper addresses the problem of identification of hybrid dynamical systems, by focusing the attention on hinging hyperplanes and Wiener piecewise affine autoregressive exogenous models, in which the regressor space is partitioned into polyhedra with affine submodels for each polyhedron. In particular, we provide algorithms based on mixed-integer linear or quadratic programming which are guaranteed to converge to a global optimum. For the special case where the estimation data only seldom switches between the different submodels, we also suggest a way of trading off between optimality and complexity by using a change detection approach.

In this contribution, the variance properties of a two-step ARX estimation scheme are discussed. An expression for the covariance of the final low-order model is calculated and it is shown how this covariance can be minimised (at least for high-model orders). The implication of the results is that identification of the dynamics of a system can very easily be performed with standard linear least squares (two times), even if the measurement noise is heavily colored. A numerical example is included, where this two-step method gives a variance which is close (but not equal) to the Cramèr-Rao lower bound. Moreover, the point estimate of the covariance is close to the one obtained through Monte Carlo simulations.

This contribution describes a common family of estimation methods for system identification, viz, prediction-error methods. The basic ideas behind these methods are described. An overview of typical model structures to which they can be applied is also given, as well as the most fundamental asymptotic properties of the resulting estimates.

This note deals with the problem of estimating the variance of an undermodeled model. Undermodeling means that the model class used is not flexible enough to describe the underlying system. The proposed solution to the problem is an algorithm that is based on the bootstrap. A simulation example shows that the variance estimates based on the proposed algorithm are in very good agreement with Monte Carlo simulations.

In this paper the effect of some weighting matrices on the asymptotic variance of the estimates of linear discrete time state space systems estimated using subspace methods is investigated. The analysis deals with systems with white or without observed inputs and refers to the Larimore type of subspace procedures. The main result expresses the asymptotic variance of the system matrix estimates in canonical form as a function of some of the user choices, clarifying the question on how to choose them optimally. It is shown, that the CCA weighting scheme leads to optimal accuracy. The expressions for the asymptotic variance can be implemented more efficiently as compared to the ones previously published.

In this contribution we examine certain variance properties of model reduction. The focus is on L₂ model reduction, but some general results are also presented. These general results can be used to analyze various other model reduction schemes. The models we study are finite impulse response (FIR) and output error (OE) models. We compare the variance of two estimated models. The first one is estimated directly from data and the other one is computed by reducing a high order model, by L₂ model reduction. In the FIR case we show that it is never better to estimate the model directly from data, compared to estimating it via L₂ model reduction of a high order FIR model. For OE models we show that the reduced model has the same variance as the directly estimated one if the reduced model class used contains the true system.

The so-called accelerated convergence is an ingenuous idea to improve the asymptotic accuracy in stochastic approximation (gradient based) algorithms. The estimates obtained from the basic algorithm are subjected to a second round of averaging, which leads to optimal accuracy for estimates of time-invariant parameters. In this contribution, some simple calculations are used to get some intuitive insight into these mechanisms. Of particular interest is to investigate the properties of accelerated convergence schemes in tracking situations. It is shown that a second round of averaging leads to the recursive least-squares algorithm with a forgetting factor. This also means that in case the true parameters are changing as a random walk, accelerated convergence does not, typically, give optimal tracking properties.

Asymptotic variance expressions are analyzed for models that are identified on the basis of closed-loop data. The considered methods comprise the classical `direct' method, as well as the more recently developed indirect methods, employing coprime factorized models, dual Youla/Kucera parametrizations and the two-stage approach. The variance expressions are compared with the open-loop situation, and evaluated in terms of their relevance for subsequent model-based control design.

Expressions for the variance of an estimated frequency function are necessary for many issues in model validation and experiment design. A general result is that a simple expression for this variance can be obtained asymptotically as the model order tends to infinity. This expression shows that the variance is inversely proportional to the signal-to-noise ratio frequency by frequency. Still, for low order models the actual variance may be quite different. This has also been pointed out in several recent publications. In this contribution we derive an exact expression for the variance, which is not asymptotic in the model order. This expression applies to a restricted class of models: AR-models, as well as fixed pole models with a polynomial noise model. It brings out the character of the simple approximation and the convergence rate to the limit as the model order increases. It also provides nonasymptotic lower bounds for the general case. The calculations are illustrated by numerical examples.

A new method for closed-loop identification that allows fitting the model to the data with arbitrary frequency weighting is described and analyzed. Just as the direct method, this new method is applicable to systems with arbitrary feedback mechanisms. This is in contrast to other methods, such as the indirect method and the two-stage method, that assume linear feedback. The finite sample behavior of the proposed method is illustrated in a simulation study.

We consider the situation where a non-linear physical system is identified from input-output data. In case no specific physical structural knowledge about the system is available, parameterized grey-box models cannot be used. Identification in black-box type of model structures is then the only alternative, and general approaches like neural nets, neuro-fuzzy models, etc., have to be applied. However, certain non-structural knowledge about the system is sometimes available. It could be known, e.g., that the step response is monotonic, or that the steady-state gain curve is monotonic. The main question is then how to utilize and maintain such information in an otherwise black-box framework. In this paper we show how this can be done, by applying a specific fuzzy model structure, with strict parametric constraints. The usefulness of the approach is illustrated by experiments on real-world data.

The determination of the right resolution parameter when estimating frequency functions of linear systems is a trade-off between bias and variance. Traditional non-parametric approaches, like `window-closing' employ a global resolution parameter - the window width - that is tuned by ad hoc methods, usually visual inspection of the results. This paper suggests a method that tunes such parameters by an automatic procedure. A further benefit is that the tuning can be performed locally, i.e., that different resolutions can be used in different frequency bands. The ideas are based on local polynomial regression and a data-driven bandwidth selector. The advantages of the proposed method are illustrated in numerical examples.

Identification of systems operating in closed loop has long been of prime interest in industrial applications. The problem offers many possibilities, and also some fallacies, and a wide variety of approaches have been suggested, many quite recently. The purpose of the current contribution is to place most of these approaches in a coherent framework, thereby showing their connections and display similarities and differences in the asymptotic properties of the resulting estimates. The common framework is created by the basic prediction error method, and it is shown that most of the common methods correspond to different parameterizations of the dynamics and noise models. The so-called indirect methods, e.g., are indeed “direct” methods employing noise models that contain the regulator. The asymptotic properties of the estimates then follow from the general theory and take different forms as they are translated to the particular parameterizations. We also study a new projection approach to closed-loop identification with the advantage of allowing approximation of the open-loop dynamics in a given, and user-chosen frequency domain norm, even in the case of an unknown, nonlinear regulator.

Data evaluation and classification have been made on measurements by an electronic nose on the headspace of samples of different types of bacteria growing on petri dishes. The chosen groups were: Escherichia coli, Enterococcus sp., Proteus mirabilis, Pseudomonas aeruginosa, and Staphylococcus saprophytica. An approximation of the response curve by time was made and the parameters in the curve fit were taken as important features of the data set. A classification tree was used to extract the most important features. These features were then used in an artificial neural network for classification. Using the ‘leave-one-out’ method for validating the model, a classification rate of 76% was obtained

The problem of tracking time-varying properties of a signal is studied. The somewhat contradictory notion of “time-varying spectrum” and how to estimate the “current” spectrum in an on-line fashion is discussed. The traditional concepts and relations between time and frequency resolution are crucial for this problem. We introduce two definitions for the time resolution of filters, essentially measuring the effective number of past data that are used to form the estimate. In, for example, wavelet transform techniques, frequency-dependent time resolutions are used so that fewer data are used at higher frequencies, thus enabling faster tracking of high-frequency components (at the price of worse frequency resolution). The main contribution of the paper is to show how this same feature can be introduced when estimating spectra via a time-varying, autoregressive model of the signal. This is achieved by a special choice of nominal covariance matrix for the underlying parameter changes.

The problem of assessing the quality of a given, or estimated, model is a central issue in system identification. Various new techniques for estimating bias and variance contributions to the model error have been suggested in recent literature. In this contribution, classical model validation procedures are placed at the focus of our attention. We discuss the principles by which we reach confidence in a model through such validation techniques and also how the distance to a “true” description can be estimated this way. In particular, we stress how the typical model validation procedure gives a direct measure of the model error of the model test without referring to its ensemble properties. Several model error bounds are developed for various assumptions about the disturbances entering the system.

A new identification algorithm which identifies low complexity models of infinite-dimensional systems from equidistant frequency-response data is presented. The new algorithm is a combination of the Fourier transform technique with the recent subspace techniques. Given noisefree data, finite-dimensional systems are exactly retrieved by the algorithm. When noise is present, it is shown that identified models strongly converge to the balanced truncation of the identified system if the measurement errors are covariance bounded. Several conditions are derived on consistency, illustrating the trade-offs in the selection of certain parameters of the algorithm. Two examples are presented which clearly illustrate the good performance of the algorithm.

Using data from extensive vibrational tests of the new Saab 2000 aircraft, a combined method for vibration analysis is studied. The method is based on a realization algorithm followed by standard prediction error methods (PEM). We find that the realization algorithm gives good initial model parameter estimates that can be further improved by the use of PEM. We use the method to get insights into the vibrational eigenmodes.

A general family of tracking algorithms for linear regression models is studied. It includes the familiar least mean square gradient approach, recursive least squares, and Kalman filter based estimators. The exact expressions for the quality of the obtained estimates are complicated. Approximate, and easy-to-use, expressions for the covariance matrix of the parameter tracking error are developed. These are applicable over the whole time interval, including the transient, and the approximation error can be explicitly calculated.

A nonlinear black-box structure for a dynamical system is a model structure that is prepared to describe virtually any nonlinear dynamics. There has been considerable recent interest in this area, with structures based on neural networks, radial basis networks, wavelet networks and hinging hyperplanes, as well as wavelet-transform-based methods and models based on fuzzy sets and fuzzy rules. This paper describes all these approaches in a common framework, from a user's perspective. It focuses on what are the common features in the different approaches, the choices that have to be made and what considerations are relevant for a successful system-identification application of these techniques. It is pointed out that the nonlinear structures can be seen as a concatenation of a mapping form observed data to a regression vector and a nonlinear mapping from the regressor space to the output space. These mappings are discussed separately. The latter mapping is usually formed as a basis function expansion. The basis functions are typically formed from one simple scalar function, which is modified in terms of scale and location. The expansion from the scalar argument to the regressor space is achieved by a radial- or a ridge-type approach. Basic techniques for estimating the parameters in the structures are criterion minimization, as well as two-step procedures, where first the relevant basis functions are determined, using data, and then a linear least-squares step to determine the coordinates of the function approximation. A particular problem is to deal with the large number of potentially necessary parameters. This is handled by making the number of ‘used’ parameters considerably less than the number of ‘offered’ parameters, by regularization, shrinking, pruning or regressor selection.

Semiphysical modelling is often interpreted as an application of system identification where physical insight into the application is used to come up with suitable non-linear transformations of the raw measurements so as to allow for a good model structure. This modelling procedure is less ‘ambitious’ than those used for traditional physical modelling in that no complete physical structure is sought, just suitable inputs and outputs that can be subjected to more or less standard model structures such as linear regressions. In this paper we discuss a semiphysical modelling procedure and various tools supporting it. These include constructive algorithms originating from commutative and differential algebra as well as more informal tools such as the programming environment.

It is a fundamental problem of identification to be able—even before the data have been analyzed—to decide if all the free parameters of a model structure can be uniquely recovered from data. This is the issue of global identifiability. In this contribution we show how global identifiability for an arbitrary model structure (basically with analytic non-linearities) can be analyzed using concepts and algorithms from differential algebra. It is shown how the question of global structural identifiability is reduced to the question of whether the given model structure can be rearranged as a linear regression. An explicit algorithm to test this is also given. Furthermore, the question of ‘persistent excitation’ for the input can also be tested explicitly is a similar fashion. The algorithms involved are very well suited for implementation in computer algebra. One such implementation is also described.

Techniques for sampling of continuous time stochastic processes are presented. To obtain flexible models and well-posed filtering problems, we assume an underlying continuous time innovations model. To sample such a model `averaged sampling' is applied. It is shown that this technique is equivalent to the following two step procedure: Determine by instantaneous (direct) sampling a discrete model for the continuous time process obtained by integrating the original innovations model. Then differentiate the sampled process to remove the discrete pole at z = 1 introduced by the integration. An advantage with this procedure is that one obtains ARMA(n, n) models, while instantaneous sampling only gives ARMA(n, n-1) models. Furthermore, the problem of updating discrete time models, without using a continuous time model, in case of a change of sampling rate - decimation/interpolation - is addressed.

The problem of estimating the transfer function of a linear system, together with the spectral density of an additive disturbance, is considered. The set of models used consists of linear rational transfer functions and the spectral densities are estimated from a finite-order autoregressive disturbance description. The true system and disturbance spectrum are, however, not necessarily of finite order. We investigate the properties of the estimates obtained as the number of observations tends to ∞ at the same time as the model order employed tends to ∞ . It is shown that the estimates are strongly consistent and asymptotically normal, and an expression for the asymptotic variances is also given. The variance of the transfer function estimate at a certain frequency is related to the signal/noise ratio at that frequency and the model orders used, as well as the number of observations. The variance of the noise spectral estimate relates in a similar way to the squared value of the true spectrum.

Most processes of realistic complexity cannot be described by simple linear relationships. An alternative to creating high order/non-linear models is to develop 'composite models’, i.e. a collection of simple models along with rules concerning when to use which one. In this paper we describe a method for constructing such composite models from observed data. It is assumed that the dynamics of the process changes with some 'operating-point vector’, which is assumed to be a measurable quantity. Based on input-output measurements and measurements of the operating-point vector, a composite model is constructed which consists of piecewise linear models. Different regions of the operating point space thus give different linear dynamics. The dynamics as well as the region boundaries are determined from the data. The basic idea is to utilize a method from recursive identification, which is able to track slow as well as rapid dynamic changes. A classification procedure is then applied to the models produced by this identification procedure, and finally borders are created between the different classified models. Techniques for supervised pattern recognition are used for the latter step. The whole construction procedure is illustrated with an example.

Tracking time-varying properties is of crucial importance in all adaptive algorithms. In this contribution we study a fairly general algorithm for tracking properties of model parameters that can be described in a linear regression form (including AR models and the like). An explicit expression for the mean square error between the estimated and the true (time-varying) parameter is established. For slow adaptation this expression can be arbitrarily well approximated by a much simpler expression. The treatment differs from other related studies using weak convergence theory, averaging, etc. in that the results are not asymptotic in nature and are applicable also to the transient phase as well as over unbounded time intervals.

A brief outline is given of mainstream system identifications, i.e. the typical approaches, algorithms, and properties in the world of data-based model construction. A number of important problems that are not sufficiently understood are pointed out. Particular attention is given to the problem of how to develop constructive and systematic ways to determine suitable model structures, i.e. to move from parameter estimation to system identification.

To track the time-varying dynamics of a system or the time-varying properties of a signal is a fundamental problem in control and signal processing. Many approaches to derive such adaptation algorithms and to analyse their behaviour have been taken. This article gives a survey of basic techniques to derive and analyse algorithms for tracking time-varying systems. Special attention is paid to the study of how different assumptions about the true system's variations affect the algorithm. Several explicit and semi-explicit expressions for the mean square error are derived, which clearly demonstrate the character of the trade-off between tracking ability and noise rejection.

The problem of tracking time-varying linear systems is discussed. The focus is on the model quality in terms of the mean square error (MSE) between the true (momentary) transfer function and the estimated one. This MSE is thus a function of frequency. The exact expression for the MSE is complicated, but simple expressions that are asymptotic in the model order are developed for model structures of finite impulse response (FIR) character. Simulations verify that these simple expressions are quite reliable and insightful even for moderate model orders. Expressions are developed for three basic adaptation algorithms (recursive identification algorithms), viz. the least-mean-squares algorithm, the recursive least-squares algorithm with exponential forgetting, and a tracking algorithm based on the Kalman filter. The results apply both to slowly time-varying systems and to the model recovery after an abrupt change in the system dynamics.

Methods of identifying bilinear systems from recorded input-output data are discussed in this article. A short survey of the existing literature on the topic is given. ‘Standard’ methods from linear systems identification, such as least squares, extended least squares, recursive prediction error and instrumental variable methods are transferred to bilinear, input-output model structures and tested in simulation. Special attention is paid to problems of stabilizing the model predictor, and it is shown how a time-varying Kalman filter and associated parameter estimation algorithm can deal with this problem.

Estimation of transfer functions of linear systems is one of the most common system identification problems. Several different design variables, chosen by the user for the identification procedure, affect the properties of the resulting estimate. In this paper it is investigated how the choices of prefilters, noise models, sampling interval, and prediction horizon (i.e., the use ofk-step ahead prediction methods) influence the estimate. An important aspect is thai the true system is not assumed to be exactly represented within the chosen model set. The estimate will thus be biased. It is shown how the distribution of bias in the frequency domain is governed by a weighting function, which emphasizes different frequency bands. The weighting function, in turn, is a result of the previously listed design variables. It is shown, e.g., thai the common least-squares method has a tendency to emphasize high frequencies, and that this can be counteracted by prefiltering. It is also shown that, asymptotically, it is only the prediction horizon itself, and not how it is split up into sampling interval times number of predicted sampling instants, that affects this weighting function.

The problem of estimating the transfer function of a linear, stochastic system is considered. The transfer function is parametrized as a black box and no given order is chosen a priori. This means that the model orders may increase to infinity when the number of observed data tends to infinity. The consistency and convergence properties of the resulting transfer function estimates are investigated. Asymptotic expressions for the variances and distributions of these estimates are also derived for the case that the model orders increase. It is shown that the variance of the transfer function estimate at a certain frequency is asymptotically given by the noise-to-signal ratio at that frequency mulliplied by the model-order-to-number-of-data-points ratio.

Many recursive identification methods have been developed for on-line estimation of parameters in dynamical systems. Such methods can also be applied to data for off-line problems, usually by passing the data through the recursive identifier a few times. In this paper such an approach is discussed. Suppose that a recursive identification method is applied to a data string obtained by concatenation of the original finite-data record. It is proved that the recursive identification method will then converge to a (local) minimum of the off-line finite-data identification criterion. We also provide some arguments as to why recursive identification may be a competitive way of performing the minimization. When using the data more than once it is a priori not clear how to treat the initial values and the forgetting factor. The problem is investigated and comparisons are made with an off-line prediction error method. Practical experiments show that the execution time for a repeated recursive algorithm is less than or equal to that of the off-line algorithm, and that the obtained accuracies are comparable.

This paper treats the close conceptual relationships between basic approaches to the estimation of transfer functions of linear systems. The classical methods of frequency and spectral analysis are shown to be related to the well-known time domain methods of prediction error type via a common “empirical transfer function estimate”. Asymptotic properties of the estimates obtained by the respective methods are also described and discussed. An important feature that is displayed by this treatment is a frequency domain weighting function that determines the distribution of bias in case the true system cannot be exactly described within the chosen model set. The choice of this weighting function is made in terms of noise models for time-domain methods. The noise model thus has a dual character from the system approximation point of view.

Identification of black-box transfer function models is considered. It is assumed that the transfer function models possess a certain shift-property, which is satisfied for example by all polynomial-type models. Expressions for the variances of the transfer function estimates are derived, that are asymptotic both in the number of observed data and in the model orders. The result is that the joint covariance matrix of the transfer functions from input to output and from driving white noise source to the additive output disturbance, respectively, is proportional to the inverse of the joint spectrum matrix for the input and driving noise multiplied by the spectrum of the additive output noise. The factor of proportionality is the ratio of model order to number of data. This result is independent of the particular model structure used. The result is applied to evaluate the performance degradation due to variance for a number of typical model uses. Some consequences for input design are also drawn.

A general linear least-squares estimation problem is considered. It is shown how the optimal filters for filtering and smoothing can be recursively and efficiently calculated under certain structural assumptions about the covariance functions involved. This structure is related to an index known as the displacement rank, which is a measure of non-Toeplitzness of a covariance kernel. When a state space type structure is added, it is shown how the Chandrasekhar equations for determining the gain of the Kalman-Bucy filter can be derived directly from the covariance function information; thus we are able to imbed this class of state-space problems into a general input-output framework.

An algorithm is described for the selection of model structure for identifying state-space models of ‘black box’ character. The algorithm receives as ‘input’ a given system in a given parametrization. It is then tested whether this parametrization is suitable (well conditioned) for identification purposes. If not, a better one is selected and the transformation of the system to the new representation is performed. This algorithm can be used as a block both in an iterative, off-line identification procedure, and for recursive, on-line identification. It can be called whenever there is some indication that the model structure is ill-conditioned. It is discussed how the model structure selection algorithm can be interfaced with an off-line identification procedure. A complete procedure is described and tested on real and simulated data.

A fast recursive matrix method for the numerical solution of Fredholm integral equations with stationary kernels is derived. IfN denotes the number of nodal points, the complexity of the algorithm isO(N 2), which should be compared toO(N 3) for conventional algorithms for solving such problems. The method is related to fast algorithms for inverting Toeplitz matrices.

Applications to equations of the first and second kind as well as miscellaneous problems are discussed and illustrated with numerical examples. These show that the theoretical improvement in efficiency is indeed obtained, and that no problems with numerical stability or accuracy are encountered.

The basic techniques of time domain and frequency domain identification, including the maximum entropy methods, are outlined. Then connections and distinctions between the methods are explored. This includes the derivation of some analytic relationships together with a discussion of the restrictions inherent in choosing certain methods, and their ease of use in different experimental conditions. It is concluded that these are complementary rather than competing techniques.

In this paper, we study the factorization and the representation of Fredholm operators belonging to the algebra $\mathcal{R}$ generated by inversion and composition of Toeplitz integral operators. The operators in $\mathcal{R}$ have the interesting property of being close to Toeplitz (in a sense quantifiable by an integer-valued index $\alpha $) and, at the same time, of being dense in the space of arbitrary kernels. By using these properties, we derive a set of efficient algorithms (generalized fast-Cholesky and Levinson recursions) for the factorization and the inversion of arbitrary Fredholm operators. The computational burden of these algorithms depends on how close (as measured by the index $\alpha $) these operators are to being Toeplitz.

We also obtain several important representation theorems for the decomposition of operators in $\mathcal{R}$ in terms of sums of products of lower times upper triangular Toeplitz operators. These results can be used to approximate operators corresponding to noncausal and time-variant systems in terms of operators representing causal and anticausal time-invariant systems, a property that has a large number of potential applications in signal processing problems.

We present a new solution for the fixed interval linear least-squares smoothing of a random signal, finite dimensional or not, inadditive white noise. By using the so-called Sobolev identity of radiative transfer theory, the smoothed estimate for stationary processes is expressed entirely in terms of time-invariant causal and anticausal filtering operations; these are interpreted from a stochastic point of view as giving certain constrained (time-invariant) filtered estimates of the signal. Then by using a recently introduced notion of processes close to stationary, these results are extended in a natural way to general nonstationary processes. From a computational point of view, the representations presented here are particularly convenient, not only because time-invariant filters can be used to find the smoothed estimate, but also because a fast algorithm based on the so-called generalized Krein-Levinson recursions can be used to compute the time-invariant filters themselves.

A certain class of methods to select suitable models of dynamical stochastic systems from measured input-output data is considered. The methods are based on a comparison between the measured outputs and the outputs of a candidate model. Depending on the set of models that is used, such methods are known under a variety of names, like output-error methods, equation-error methods, maximum-likelihood methods, etc. General results are proved concerning the models that are selected asymptotically as the number of observed data tends to infinity. For these results it is not assumed that the true system necessarily can be exactly represented within the chosen set of models. In the particular case when the model set contains the system, general consistency results are obtained and commented upon. Rather than to seek an exact description of the system, it is usually more realistic to be content with a suitable approximation of the true system with reasonable complexity properties. Here, the consequences of such a viewpoint are discussed.

Recursive algorithrms for the solution of linear least-squares estimation problems have been based mainly on state-space models. It has been known, however, that recursive Levinson-Whittle-Wiggins-Robinson (LWR) algorithms exist for stationary time-series, using only input-output information (i.e, covariance matrices). By introducing a way of classifying stochastic processes in terms of an "index of nonstationarity" we derive extended LWR algorithms for nonstationary processes We show also how adding state-space structure to the covariance matrix allows us to specialize these general results to state-space type estimation algorithms. In particular, the Chandrasekhar equations are shown to be natural descendants of the extended LWR algorithm.

Convergence with probability one of a recursive stochastic approximation algorithm is considered. Some extensions of previous results for the Robbins-Monro and the Kiefer-Wolfowitz procedures are given. An inportant feature of the approach taken here is that the convergence analysis can be directly extended to more complex algorithms.

A state-space representation of a dynamical, stochastic system is given. A corresponding model, parametrized in a particular way, is considered and an algorithm for the estimation of its parameters is analysed. The class of estimation algorithms thus considered contains general output error methods and model reference methods applied to stochastic systems. It also contains adaptive filtering schemes and, e.g. the extended least squares method.

It is shown that if a certain transfer function associated with the true system is positive real, then the estimation algorithm converges with probability 1 to a value that gives a correct input-output model.

This paper reviews work on self-tuning regulators. The regulator algorithms, their theory and industrial applications are reviewed. The paper is expository—the major ideas are covered but detailed analysis is given elsewhere.

Recursive algorithms where random observations enter are studied in a fairly general framework. An important feature is that the observations my depend on previous "outputs" of the algorithm. The considered class of algorithms contains, e.g., stochastic approximation algorithm, recursive identification algorithm, and algorithms for adaptive control of linear systems. It is shown how a deterministic differential equation can be associated with the algorithm. Problems like convergence with probability one, possible convergence points and asymptotic behavior of the algorithm can all be studied in terms of this differential equation. Theorems stating the precise relationships between the differential equation and the algorithm are given as well as examples of applications of the results to problems in identification and adaptive control.

It is often necessary in practice to perform identification experiments on systems operating in closed loop. There has been some confusion about the possibilities of successful identification in such cases, evidently due to the fact that certain common methods then fail. A rapidly increasing literature on the problem is briefly surveyed in this paper, and an overview of a particular approach is given. It is shown that prediction error identification methods, applied in a direct fashion will given correct estimates in a number of feedback cases. Furthermore, the accuracy is not necessarily worse in the presence of feedback; in fact optimal inputs may very well require feedback terms. Some practical applications are also described.

Least-squares estimation of the parameters of a vector difference equation model of a dynamic system is studied. A theorem for the convergence and consistency of the least-squares estimate is given that is valid under general feedback conditions.

Mathematical models of real life systems and processes are essential in today’s industrial work. To be able to construct such models is therefore a fundamental skill in modern engineering...

The DMV seminar "Stochastische Approximation und Optimierung zufalliger Systeme" was held at Blaubeuren, 28. 5. -4. 6. 1989. The goal was to give an approach to theory and application of stochas tic approximation in view of optimization problems, especially in engineering systems. These notes are based on the seminar lectures. They consist of three parts: I. Foundations of stochastic approximation (H. Walk); n. Applicational aspects of stochastic approximation (G. PHug); In. Applications to adaptation :ugorithms (L. Ljung). The prerequisites for reading this book are basic knowledge in probability, mathematical statistics, optimization. We would like to thank Prof. M. Barner and Prof. G. Fischer for the or ganization of the seminar. We also thank the participants for their cooperation and our assistants and secretaries for typing the manuscript. November 1991 L. Ljung, G. PHug, H. Walk Table of contents I Foundations of stochastic approximation (H. Walk) §1 Almost sure convergence of stochastic approximation procedures 2 §2 Recursive methods for linear problems 17 §3 Stochastic optimization under stochastic constraints 22 §4 A learning model; recursive density estimation 27 §5 Invariance principles in stochastic approximation 30 §6 On the theory of large deviations 43 References for Part I 45 11 Applicational aspects of stochastic approximation (G. PHug) §7 Markovian stochastic optimization and stochastic approximation procedures 53 §8 Asymptotic distributions 71 §9 Stopping times 79 §1O Applications of stochastic approximation methods 80 References for Part II 90 III Applications to adaptation algorithms (L.

Methods of recursive identification deal with the problem of building mathematical models of signals and systems on-line, at the same time as data is being collected. Such methods, which are also known as adaptive algorithms or sequential parameter estimation methods, may be applied to a wide spectrum of online adaptive systems, such as devices for signal processing, prediction, or control and are useful for modeling systems in general. For example, they can be used to analyze the demand for power on an electric generating grid and help the grid adjust to continually changing power needs, or applied to the changing conditions of a papermaking plant, or to monitoring pollution in a river.

This book provides a comprehensive and systematic framework for developing, describing, and analyzing such recursive algorithms. It has been carefully designed and organized to meet the needs of readers with different objectives. With a myriad of algorithms now in use, it provides a simple and coherent frame of reference for understanding the subject and will serve as a guide to the large number of choices made available by the advent of inexpensive, powerful digital processors.

Readers primarily interested in theory will find a detailed development of convergence analysis and asymptotic distribution results. For graduate students it is a basic introduction to the subject. And for engineers interested in practical applications, the book's earlier theory-oriented chapters are equipped with "user's summaries" that provide direct access to the discussion of practical aspects developed in the final three chapters on implementation and applications.

Boken behandlar grundläggande metoder för beskrivning, analys och dimensionering av reglersystem. Klassiska angreppssätt som rotortmetoden och frekvensanalys med hjälp av Bodediagram diskuteras.

Dessutom introduceras tillståndsvariabler och metoder att analysera känslighet och robusthet hos reglersystem. Grunderna för behandling av olinjära fenomen och digital implementering av regulatorer diskuteras.

Bokens omfattning motsvarar en typisk grundkurs i reglerteknik inom civilingenjörsutbildningen.

A new type of linear kernel smoother is derived and studied. The smoother, referred to as weight determination by manifold regularization, is the solution to a regularized least squares problem. The regularization avoids overfitting and can be used to express prior knowledge of an underlying smooth function. An interesting property ofthe kernel smoother is that it is well suited for systems govern by the semi-supervised smoothness assumption. Several examples are given to illustrate this property. We also discuss why these types of techniques can have a potential interest for the system identification community.

It is a typical situation in modern modeling that a total model is built up from simpler submodels, or modules, for example residing in a model library. The total model could be quite complex, while the modules are well understood and analysed. A procedure to decide global parameter identifiability for such a collection of model equations of differential-algebraic nature is suggested. It is shown how to make use of the natural modularization of the model structure. Basically, global identifiability is obtained if and only if each module is identifiable, and the connecting signals can be retrieved from the external signals, without knowledge of the values of the parameters.

System Identification concerns the problem of building mathematical models of dynamical systems. This involves a fair amount of theory and algorithms. Equally important, though, is the practical side of the methodology. This chapter deals with the issues that are essential to construct a good model in practice. Such issues include the problem of input and experiment design. Typical essential features of the input are discussed as well as examples of commonly used inputs. Next comes the question of how to condition the measured signals. This involves issues of removing trends and disturbances outside of the frequency ranges of interest for the model. The most demanding task is to find a suitable model structure, guided by information in the observed data. A first cut methodology for this is described. For a successful application, it is in the end a matter of combining intuition, and information from various data test. Two applications, a fighter aircraft and a buffer vessel in process industry illustrate the process.

The interplay between the time and frequency domains for linear systems is well known and most useful. In the case of identifying linear models from observed data, this interplay manifests itself in two ways.

The observed data is of course primarily recorded in the time domain (even though the recording equipment may deliver them in the frequency domain). To build linear models from the data we can either do the fitting in the time domain and evaluate the resulting model’s properties in the frequency domain. We can also transfer the data themselves to the frequency domain and fit models directly there. In this contribution we shall consider some aspects of these two possibilities.

Mechanisms for adapting models, filters, decisions, regulators and soon to changing properties of a system or a signal are of fundamental importance in many modern signal processing and control algorithms. In this chapter we give an overview of some basic set-ups and algorithms that are used for this. We pay special attention to the rationale behind the different algorithms, thus distinguishing between "optimal" algorithms and "ad hoc" algorithms. We also give an outline of the basic approaches to performance analysis of adaptive algorithms.

The convergence with probability one, of the parameter estimates obtained from prediction error identification methods, such as the maximum likelihood method, is analysed in this paper. It is shown that under quite weak assumptions on the actual system that has generated the data, the expected value of the identification criterion can be used for the asymptotic analysis of the estimates. In particular, the true system does not have to belong to the set of models over which the search for optimum is made. The implications of this result for consistency analysis and for questions of identifiability, as well as for other related problems are discussed.

Asymptotic theory is one of the core subjects in system identification theory and often used to assess properties of model estimators. In this paper, we focus on the asymptotic theory for the kernel-based regularized system identification and study the convergence in distribution of the generalized cross validation (GCV) based hyper-parameter estimator. It is shown that the difference between the GCV based hyper-parameter estimator and the optimal hyper-parameter estimator that minimizes the mean square error scaled by 1/root N converges in distribution to a zero mean Gaussian distribution, where N is the sample size and an expression of covariance matrix is obtained. In particular, for the ridge regression case, a closed-form expression of the variance is obtained and shows the influence of the limit of the regression matrix on the asymptotic distribution. For illustration, Monte Carlo numerical simulations are run to test our theoretical results.

The R2021b release of the System Identification ToolboxTM for MATLAB contains new features that enable the use of machine learning techniques for nonlinear system identification. With this release it is possible to build nonlinear ARX models with regression tree ensemble and Gaussian process regression mapping functions. The release contains several other enhancements including, but not limited to, (a) online state estimation using the extended Kalman filter and the unscented Kalman filter with code generation capability; (b) improved handling of initial conditions for transfer functions and polynomial models; (c) a new architecture of nonlinear black-box models that streamlines regressor handling, reduces memory footprint and improves numerical accuracy; and (d) easy incorporation of identification apps in teaching tools and interactive examples by leveraging the Live Editor tasks of MATLAB. Copyright (C) 2021 The Authors.

In the overview paper on nonlinear system identification Schoukens and Ljung (2019), it was indicated that reliable expressions to calculate the variance of an estimated nonlinear model are lacking, especially if the disturbing noise is entering the nonlinear regressors. In this study, we provide a better view on the driving mechanisms of the variability of estimated nonlinear models that is due to noise on the output. To do so, we follow a double approach. Firstly, a basic insight on the impact of disturbing noise on the monomial yTh is studied. Next, these insights are used in a case study on the forced Duffing oscilator data, also called the Silver box (Schoukens and Noel, 2016). The following models are studied: Nonlinear autoregressive exogenous models (NARX) using a 2-layer Neural Net (NARX-NN) and a polynomial (NARX-poly) expansion; and polynomial nonlinear state space models (PNLSS). This limited study indicates that an output error criterion (PNLSS, and NARX used in a simulation mode) does better than minimizing the equation error (NARX-NN and NARX-poly in prediction mode). When the signal-to-noise ratio (SNR) drops below 20 dB, the reduction in the error is more than a factor 10. This is a strong indication that, just as for linear identification, it is very important to properly deal with the noise properties in the cost function whenever the SNR of the output measurements drops such that noise becomes more important than structural model errors. Copyright (C) 2021 The Authors.

Deep learning is a topic of considerable interest today. Since it deals with estimating - or learning - models, there are connections to the area of System Identification developed in the Automatic Control community. Such connections are explored and exploited in this contribution. It is stressed that common deep nets such as feedforward and cascadeforward nets are nonlinear ARX (NARX) models, and can thus be easily incorporated in System Identification code and practice. The case of LSTM nets is an example of NonLinear State-Space (NLSS) models. Copyright (C) 2020 The Authors.

This paper studies the asymptotic properties of the hyperparameter estimators including the leave-k-out cross validation (LKOCV) and r-fold cross validation (RFCV), and discloses their relation with the Steins unbiased risk estimators (SURE) as well as the mean squared error (MSE). It is shown that as the number of data goes to infinity, the LKOCV shares the same asymptotic best hyperparameter minimizing the MSE estimator as the SURE does if the input is bounded and the ratio between the training data and the whole data tends to zero. We illustrate the efficacy of the theoretical result by Monte Carlo simulations.

Regularization methods with regularization matrix in quadratic form have received increasing attention. For those methods, the design and tuning of the regularization matrix are two key issues that are closely related. For systems with complicated dynamics, it would be preferable that the designed regularization matrix can bring the hyper-parameter estimation problem certain structure such that a locally optimal solution can be found efficiently. An example of this idea is to use the so-called multiple kernel Chen et al. (2014) for kernel-based regularization methods. In this paper, we propose to use the multiple regularization matrix for the filter-based regularization. Interestingly, the marginal likelihood maximization with the multiple regularization matrix is also a difference of convex programming problem, and a locally optimal solution could be found with sequential convex optimization techniques. (C) 2018, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.

An estimated state-space model can possibly be improved by further iterations with estimation data. This contribution specifically studies if models obtained by subspace estimation can be improved by subsequent re-estimation of the B, C, and D matrices (which involves linear estimation problems). Several tests are performed, which show that it is generally advisable to do such further re-estimation steps using the maximum likelihood criterion. Stated more succinctly in terms of MATLABC (R) functions, ssest generally outperforms n4sid. (C) 2018, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.

Using Maximum Likelihood (or Prediction Error) methods to identify linear state space model is a prime technique. The likelihood function is a nonconvex function and care must be exercised in the numerical maximization. Here the focus will be on affine parameterizations which allow some special techniques and algorithms. Three approaches to formulate and perform the maximization are described in this contribution: (1) The standard and well known Gauss Newton iterative search, (2) a scheme based on the EM (expectation-maximization) technique, which becomes especially simple in the affine parameterization case, and (3) a new approach based on lifting the problem to a higher dimension in the parameter space and introducing rank constraints. (C) 2018, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.

Gray-box identification is prevalent in modeling physical and networked systems. However, due to the non-convex nature of the gray-box identification problem, good initial parameter estimates are crucial for a successful application. In this paper, a new identification method is proposed by exploiting the low-rank and structured Hankel matrix of impulse response. This identification problem is recasted into a difference-of-convex programming problem, which is then solved by the sequential convex programming approach with the associated initialization obtained by nuclear-norm optimization. The presented method aims to achieve the maximum impulse-response fitting while not requiring additional (non-convex) conditions to secure non-singularity of the similarity transformation relating the given state space matrices to the gray-box parameterized ones. This overcomes a persistent shortcoming in a number of recent contributions on this topic, and the new method can be applied for the structured state-space realization even if the involved system parameters are unidentifiable. The method can be used both for directly estimating the gray-box parameters and for providing initial parameter estimates for further iterative search in a conventional gray-box identification setup. (C) 2017, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.

This paper considers the input design of kernelbased regularization methods for LTI system identification. We first derive the Bayesian mean squared error matrix under the Bayesian perspective, and then use some typical scalar measures (e.g., the A-optimality, D-optimality, and E-optimality) as optimization criteria for the input design problem. Instead of directly solving the nonconvex optimization problem, we propose a two-step procedure. The first step is to solve a convex optimization and the second one is to determine the inverse image of a quadratic map. Both of these two steps can be solved efficiently by the proposed method and hence all the globally optimal inputs are found. In particular, we show that for some kernels, the optimal input under the D-optimality has an explicit expression.

Network reconstruction has become particularly important in systems biology, and is now expected to deliver information on causality. Systems in nature are inherently nonlinear. However, for nonlinear dynamical systems with hidden states, how to give a useful definition of dynamic networks is still an open question. This paper presents a useful definition of Boolean dynamic networks for a large class of nonlinear systems. Moreover, a robust inference method is provided. The well-known Millar-10 model in systems biology is used as a numerical example, which provides the ground truth of causal networks for key mRNAs involved in eukaryotic circadian clocks. In addition, as second contribution of this paper, we suggest definitions of linear network identifiability, which helps to unify the available work on network identifiability.

The identification of continuous-time models of dynamical systems based on sampled measurements of input and output signals is a research topic that has received much attention during the past decades. However, a framework for the correct assessment of the performance of various estimation methods, as well as their numerical reliability, is still missing due to a number of benchmarking difficulties, equally applicable to both discrete-and continuous-time identification problems. This paper revisits this topic, reports new numerical results, highlights several fundamental aspects regarding the definition of an appropriate benchmark for the evaluation of continuous-time linear model identification algorithms and discusses several means of addressing the related existing problems.

System identification with regularization methods has attracted increasing attention recently and is a complement to the current standard maximum likelihood/ prediction error method. In this paper, we focus on the kernel-based regularization method and give a spectral analysis of the so-called diagonal correlated (DC) kernel, one family of kernel structures that has been proven useful for linear time-invariant system identification. In particular, using the theory of Bessel functions, we derive the eigenvalues and corresponding eigenfunctions of the DC kernel. Accordingly, we derive the Karhunen-Loeve expansion of the stochastic process whose covariance function is the DC kernel.

The mainstream approach to identication of linear discrete-time models is givenby parametric Prediction Error Methods (PEM). As a rule, the model complexity is unknownand model order selection (MOS) is a key ingredient of the estimation process. A dierentapproach to linear system identication has been recently proposed where impulse responsesare described in a Bayesian framework as zero-mean Gaussian processes. Their covariances aregiven by the so-called stable spline, TC or DC kernels that encode information on regularityand BIBO stability. In this paper, we show that these new kernel-based techniques lead alsoto a new eective MOS method for PEM. Furthermore, this paves the way to the design ofa new impulse response estimator that combines the regularized approaches and the classicalparametric PEM. Numerical experiments show that the performance of this technique is verysimilar to that of PEM equipped with an oracle which selects the best model order by knowingthe true impulse response.

System Identification is about estimating models of dynamical systems from measured input-output data. Its traditional foundation is basic statistical techniques, such as maximum likelihood estimation and asymptotic analysis of bias and variance and the like. Maximum likelihood estimation relies on minimization of criterion functions that typically are non-convex, and may cause numerical search problems. Recent interest in identification algorithms has focused on techniques that are centered around convex formulations. This is partly the result of developments in machine learning and statistical learning theory. The development concerns issues of regularization for sparsity and for better tuned bias/variance trade-offs. It also involves the use of subspace methods as well as nuclear norms as proxies to rank constraints. A quite different route to convexity is to use algebraic techniques manipulate the model parameterizations. This article will illustrate all this recent development.                       

FIR (finite impulse response) model is widely used in tackling the problem of the impulse response estimation with quantized measurements. Its use is, however, limited, in the case when a high order FIR model is required to capture a slowly decaying impulse response. This is because the high variance for high order FIR models would override the low bias and thus lead to large MSE (mean square error). In this contribution, we apply the recently introduced regularized FIR model approach to the problem of the impulse response estimation with binary measurements. We show by Monte Carlo simulations that the proposed approach can yield both better accuracy and better robustness than a recently introduced FIR model based approach.

Change detection has traditionally been seen as a centralized problem. Many change detection problems are however distributed in nature and the need for distributed change detection algorithms is therefore significant. In this paper a distributed change detection algorithm is proposed. The change detection problem is first formulated as a convex optimization problem and then solved distributively with the alternating direction method of multipliers (ADMM). To further reduce the computational burden on each sensor, a homotopy solution is also derived. The proposed method have interesting connections with Lasso and compressed sensing and the theory developed for these methods are therefore directly applicable.

The identification of multiple affine subspaces from a set of data is of interest in fields such as system identification, data compression, image processing and signal processing and in the literature referred to as subspace clustering. If the origin of each sample would be known, the problem would be trivially solved by applying principal component analysis to samples originated from the same subspace. Now, not knowing what samples that originates from what subspace, the problem becomes considerably more difficult. We present a novel convex formulation for subspace clustering. The proposed method takes the shape of a least-squares problem with sum-of-norms regularization over optimization parameter differences, a generalization of the ℓ1-regularization. The regularization constant is used to trade off fit and the identified number of affine subspaces.

Segmentation of time series data is of interest in many applications, as for example in change detection and fault detection. In the area of convex optimization, the sum-of-norms regularization has recently proven useful for segmentation. Proposed formulations handle linear models, like ARX models, but cannot handle nonlinear models. To handle nonlinear dynamics, we propose integrating the sum-of-norms regularization with a least squares support vector machine (LS-SVM) core model. The proposed formulation takes the form of a convex optimization problem with the regularization constant trading off the fit and the number of segments.

Systems today often consist of logic switches working besides continuous physical systems. The demand for novel hybrid system identification algorithms is therefore of growing interest and essential for the development of control algorithms for this type of systems. An important type of hybrid systems is piecewise affine systems. The identification of piecewise affine systems is here tackled by overparametrizing and assigning a regressor-parameter to each of the observations. The regressor parameters are forced to be the same if that not causes a major increase in the fit term. The formulation takes the shape of a least-squares problem with sum-of-norms regularization over regressor parameter differences, a generalization of l1-regularization. The regularization constant is used to trade off fit and the number of partitions of the model.

System Identification for linear systems and models is a well established and mature topic. Identifying nonlinear models is a much more rich and demanding problem area. In this presentation some major approaches and concepts for that are outlined.

The presence of abrupt changes, such as impulsive disturbances and load disturbances, make state estimation considerably more difficult than the standard setting with Gaussian process noise. Nevertheless, this type of disturbances is commonly occurring in applications which makes it an important problem. An abrupt change often introduces a jump in the state and the problem is therefore readily treated by change detection techniques. In this paper, we take a rather different approach. The state smoothing problem for linear state space models is here formulated as a least-squares problem with sum-of-norms regularization, a generalization of the ℓ1-regularization. A nice property of the suggested formulation is that it only has one tuning parameter, the regularization constant which is used to trade off fit and the number of jumps.

Many tracking problems are split into two sub-problems, first a smooth reference trajectory is generated that meet the control design objectives, and then a closed loop control system is designed to follow this reference trajectory as well as possible. Applications of this kind include (autonomous) vehicle navigation systems and robotics. Typically, a spline model is used for trajectory generation and another physical and dynamical model is used for the control design. Here we propose a direct approach where the dynamical model is used to generate a control signal that takes the state trajectory through the waypoints specified in the design goals. The strength of the proposed formulation is the methodology to obtain a control signal with compact representation and that changes only when needed, something often wanted in tracking. The formulation takes the shape of a constrained least-squares problem with sum-of-norms regularization, a generalization of the ℓ1-regularization. The formulation also gives a tool to, e.g. in model predictive control, prevent chatter in the input signal, and also select the most suitable instances for applying the control inputs.

To estimate the past climate, for example the ocean temperature 1000 years ago, one has to turn to naturally occurring climate recorders. There exist a number of climate recorders in nature from which the past temperature can be extracted. However, only a few natural archives are able to record climate fluctuations with high enough resolution so that the seasonal variations can be reconstructed. One such archive is a bivalve shell. The chemical composition of a shell of a bivalve depends on a number of chemical and physical parameters of the water in which the shell was composed. Of these parameters, the water temperature is probably the most important one. It should therefore be possible to estimate the water temperature for the years the shell was built, from measurements of the shell's chemical composition. In this paper, we explore this possibility. We do this by first observing that the chemical compositions lie on a one-dimensional manifold parameterized by the water temperature. This manifold is then utilized in the regression to obtain accurate estimates of past water temperatures.

The Two-Stage Algorithm (TSA) has been extensively used and adapted for the identification of Hammerstein systems. It is essentially based on a particular formulation of Hammerstein systems in the form of bilinearly parameterized linear regressions. This paper has been motivated by a somewhat contradictory fact: though the optimality of the TSA has been established by Bai in 1998 only in the case of some special weighting matrices, the unweighted TSA is usually used in practice. It is shown in this paper that the unweighted TSA indeed gives the optimal solution of the weighted nonlinear least-squares problem formulated with a particular weighting matrix. This provides a theoretical justification of the unweighted TSA, and leads to a generalization of the obtained result to the case of colored noise with noise whitening. Numerical examples of identification of Hammerstein systems are presented to validate the theoretical analysis.

The paper describes additions to the System Identification Toolbox offered by The MathWorks, Inc, that handle the estimation of nonlinear models. Both structured grey-box models and general, flexible black-box models are covered. The idea is that the look and feel of the syntax, and the graphical user interface (GUI) should be as close as possible to the linear case. This presentation is focused on the GUI functionality and the possibilities to simulate the estimated models in Simulink.

We address a problem of estimation of an unknown regression function f at a given point x_0 from noisy observations y_i = f(x_i)+ e_i, ;i =1, ..., n. Here x_i in {cal R}^k are observable regressors and (e_i) are normal i.i.d. (unobservable) disturbances. The problem is analyzed in the minimax framework, namely, we suppose that f belongs to some functional class F, such that its finite-dimensional cut F_n = {f(x_i),fin F,i =0, ..., n, } is a convex compact set. For an arbitrary fixed regression plan X_n =(x_1;...;x_n) we study minimax on F_n confidence intervals of affine estimators and construct an estimator which attains the minimax performance on the class of arbitrary estimators when the confidence level approaches 1.

The prediction-error approach to parameter estimation of linear models often involves solving a non-convex optimization problem. In some cases, it is therefore difficult to guarantee that the global optimum will be found. A common way to handle this problem is to find an initial estimate, hopefully lying in the region of attraction of the global optimum, using some other method. The prediction-error estimate can then be obtained by a local search starting at the initial estimate. In this paper, a new approach for finding an initial estimate of certain linear models utilizing structure and the subspace method is presented. The polynomial models are first written on the observer canonical state-space form, where the specific structure is later utilized, rendering least-squares estimation problems with linear equality constraints.

Order selection of linear regression models has been thoroughly researched in the statistical community for some time. Different shrinkage methods have been proposed, such as the Ridge and Lasso regression methods. Especially the Lasso regression has won fame because of its ability to set less important parameters exactly to zero. However, these methods do not take dynamical systems into account, where the regressors are ordered via the time lag. To this end, a modified variant of the nonnegative garrote method will be analyzed.

The Wiener model is a block oriented model having a linear dynamicsystem followed by a static nonlinearity.The dominating approachto estimate the components of this model has been to minimize theerror between the simulated and the measured outputs. We show thatthis will in general lead to biased estimates if there is otherdisturbances present than measurement noise. The implications ofBussgangs theorem in this context are also discussed. For the casewith general disturbances we derive the Maximum Likelihood methodand show how it can be efficiently implemented. Comparisons betweenthis new algorithm and the traditional approach confirm that the newmethod is unbiased and also has superior accuracy.

The Direct Weight Optimization (DWO) approach is a nonparametric estimation approach that has appeared in recent years within the field of nonlinear system identification. In previous work, all function classes for which DWO has been studied have included only continuous functions. However, in many applications it would be desirable also to be able to handle discontinuous functions. Inspired by the bilateral filter method from image processing, such an extension of the DWO framework is proposed for the smoothing problem. Examples show that the properties of the new approach regarding the handling of discontinuities are similar to the bilateral filter, while at the same time DWO offers a greater flexibility with respect to different function classes handled.

The basic least squares method for identifying linear systems has been extensively studied. Conditions for convergence involve issues about noise assumptions and behavior of the sample covariance matrix of the regressors. Lai and Wei proved in 1982 convergence for essentially minimal conditions on the regression matrix: All eigenvalues must tend to infinity, and the logarithm of the largest eigenvalue must not tend to infinity faster than the smallest eigenvalue. In this contribution we revisit this classical result with respect to assumptions on the noise: How much unstructured disturbances can be allowed without affecting the convergence? The answer is that the norm of these disturbances must tend to infinity slower than the smallest eigenvalue of the regression matrix.