In time series analysis, most of the models are based on the assumption of covariance stationarity. However, many time series in the applied sciences show a time-varying second-order structure. That is, variance and covariance, or equivalently the spectral structure, are likely to change over time. Examples may be found in a growing number of fields, such as biomedical time series analysis, geophysics, telecommunications, or financial data analysis, to name but a few. In this thesis, we are concerned with the modelling of such nonstationary time series, and with the subsequent questions of how to estimate their second-order structure and how to forecast these processes. We focus on univariate, discrete-time processes with zero-mean arising, for example, when the global trend has been removed from the data. The first chapter presents a simple model for nonstationarity, where only the variance is time-varying. This model follows the approach of "local stationarity" introduced by [1]. We show that our model satisfactorily explains the nonstationary behaviour of several economic data sets, among which are the U.S. stock returns and exchange rates. This chapter is based on [5]. In the second chapter, we study more complex models, where not only the variance is evolutionary. A typical example of these models is given by time-varying ARMA(p,q) processes, which are ARMA(p,q) with time-varying coefficients. Our aim is to fit such semiparametric models to some nonstationary data. Our data-driven estimator is constructed from a minimisation of a penalised contrast function, where the contrast function is an approximation to the Gaussian likelihood of the model. The theoretical performance of the estimator is analysed via non asymptotic risk bounds for the quadratic risk. In our results, we do not assume that the observed data follow the semiparamatric structure, that is our results hold in the misspecified case. The third chapter introduces a fully nonparametric model for local nonstationarity. This model is a wavelet-based model of local stationarity which enlarges the class of models defined by Nason et al. [3]. A notion of time-varying "wavelet spectrum' is uniquely defined as a wavelet-type transform of the autocovariance function with respect to so-called "autocorrelation wavelets'. This leads to a natural representation of the autocovariance which is localised on scales. One particularly interesting subcase arises when this representation is sparse, meaning that the nonstationary autocovariance may be decomposed in the autocorrelation wavelet basis using few coefficients. We present a new test of sparsity for the wavelet spectrum in Chapter 4. It is based on a non-asymptotic result on the deviations of a functional of a periodogram. In this chapter, we also present another application of this result given by the pointwise adaptive estimation of the wavelet spectrum. Chapters 3 and 4 are based on [6] Computational aspects of the test of sparsity and of the pointwise adaptive estimator are considered in Chapter 5. We give a description of a full algorithm, and an application in biostatistics. In this chapter, we also derive a new test of covariance stationarity, applied to another case study in biostatistics. This chapter is based on [7]. Finally, Chapter 6 address the problem how to forecast the general nonstationary process introduced in Chapter 3. We present a new predictor and derive the prediction equations as a generalisation of the Yule-Walker equations. We propose an automatic computational procedure for choosing the parameters of the forecasting algorithm. Then we apply the prediction algorithm to a meteorological data set. This chapter is based on [2,4]. References [1] Dahlhaus, R. (1997). Fitting time series models to nonstationary processes. Ann. Statist., 25, 1-37, 1997. [2] Fryzlewicz, P., Van Bellegem, S. and von Sachs, R. (2003). Forecasting non-stationary time series by wavelet process modelling. Annals of the Institute of Statistical Mathematics. 55, 737-764. [3] Nason, G.P., von Sachs, R. and Kroisandt, G. (2000). Wavelet processes and adaptive estimation of evolutionary wavelet spectra. Journal of the Royal Statistical Society Series B. 62, 271-292. [4] Van Bellegem, S., Fryzlewicz, P. and von Sachs, R. (2003). A wavelet-based model for forecasting non-stationary processes. In J-P. Gazeau, R. Kerner, J-P. Antoine, S. Metens and J-Y. Thibon (Eds.). GROUP 24: Physical and Mathematical Aspects of Symmetries. Bristol: IOP Publishing (in press). [5] Van Bellegem, S. and von Sachs, R. (2003). Forecasting economic time series with unconditional time-varying variance. International Journal of Forecasting (in press). [6] Van Bellegem, S. and von Sachs, R. (2003). Locally adaptive estimation of sparse, evolutionary wavelet spectra (submitted). [7] Van Bellegem, S. and von Sachs, R. (2003). On adaptive estimation for locally stationary wavelet processes and its applications (submitted).

Le développement des systèmes intelligents pour contrôler la stabilité du véhicule et éviter les accidents routier est au cœur de la recherche automobile. L'expansion de ces systèmes intelligents à l'application réelle exige une estimation précise de la dynamique du véhicule dans des environnements diverses (dévers et pente). Cette exigence implique principalement trois problèmes : ⅰ), extraire des informations non mesurées à partir des capteurs faible coût; ⅱ), rester robuste et précis face aux les perturbations incertaines causées par les erreurs de mesure ou de la méconnaissance de l'environnement; ⅲ), estimer l'état du véhicule et prévoir le risque d'accident en temps réel. L’originalité de cette thèse par rapport à l’existant, consiste dans le développement des nouveaux algorithmes, basés sur des nouveaux modèles du véhicule et des différentes techniques d'observation d'état, pour estimer des variables ou des paramètres incertains de la dynamique du véhicule en temps réel. La première étape de notre étude est le développement de nouveaux modèles pour mieux décrire le comportement du véhicule dans des différentes situations. Pour minimiser les erreurs de modèle, un système d'estimation composé de quatre observateurs est proposé pour estimer les forces verticales, longitudinales et latérales par pneu, ainsi que l'angle de dérive. Trois techniques d'observation non linéaires (EKF, UKF et PF) sont appliquées pour tenir compte des non-linéarités du modèle. Pour valider la performance de nos observateurs, nous avons implémenté en C++ des modules temps-réel qui, embarqué sur le véhicule, estiment la dynamique du véhicule pendant le mouvement. / Enhancing road safety by developing active safety system is the general purpose of this thesis. A challenging task in the development of active safety system is to get accurate information about immeasurable vehicle dynamics states. More specifically, we need to estimate the vertical load, the lateral frictional force and longitudinal frictional force at each wheel, and also the sideslip angle at center of gravity. These states are the key parameters that could optimize the control of vehicle's stability. The estimation of vertical load at each tire enables the evaluation of the risk of rollover. Estimation of tire lateral forces could help the control system reduce the lateral slip and prevent the situation like spinning and drift out. Tire longitudinal forces can also greatly influence the performance of vehicle. The sideslip angle is one of the most important parameter to control the lateral dynamics of vehicle. However, in the current market, very few safety systems are based on tire forces, due to the lack of cost-effective method to get these information. For all the above reasons, we would like to develop a perception system to monitor these vehicle dynamics states by using only low-cost sensor. In order to achieve this objective, we propose to develop novel observers to estimate unmeasured states. However, construction of an observer which could provide satisfactory performance at all condition is never simple. It requires : 1, accurate and efficient models; 2, a robust estimation algorithm; 3, considering the parameter variation and sensor errors. As motivated by these requirements, this dissertation is organized to present our contribution in three aspects : vehicle dynamics modelization, observer design and adaptive estimation. In the aspect of modeling, we propose several new models to describe vehicle dynamics. The existent models are obtained by simplifying the vehicle motion as a planar motion. In the proposed models, we described the vehicle motion as a 3D motion and considered the effects of road inclination. Then for the vertical dynamics, we propose to incorporate the suspension deflection to calculate the transfer of vertical load. For the lateral dynamics, we propose the model of transfer of lateral forces to describe the interaction between left wheel and right wheel. With this new model, the lateral force at each tire can be calculated without sideslip angle. Similarly, for longitudinal dynamics, we also propose the model of transfer of longitudinal forces to calculate the longitudinal force at each tire. In the aspect of observer design, we propose a novel observation system, which is consisted of four individual observers connected in a cascaded way. The four observers are developed for the estimation of vertical tire force, lateral tire force and longitudinal tire force and sideslip angle respectively. For the linear system, the Kalman filter is employed. While for the nonlinear system, the EKF, UKF and PF are applied to minimize the estimation errors. In the aspect of adaptive estimation, we propose the algorithms to improve sensor measurement and estimate vehicle parameters in order to stay robust in presence of parameter variation and sensor errors. Furthermore, we also propose to incorporate the digital map to enhance the estimation accuracy. The utilization of digital map could also enable the prediction of vehicle dynamics states and prevent the road accidents. Finally, we implement our algorithm in the experimental vehicle to realize real-time estimation. Experimental data has validated the proposed algorithm.

This research presents a novel framework for automatic chroma keying and the optimizations for real-time and stereoscopic 3D processing. It first simulates the process of human perception on isolating foreground elements in a given scene by perceptual analysis, and then predicts foreground colours and alpha map based on the analysis results and the restored clean background plate rather than direct sampling. Besides, an object level depth map is generated through stereo matching on a carefully determined feature map. In addition, three prototypes on different platforms have been implemented according to their hardware capability based on the proposed framework. To achieve real-time performance, the entire procedures are optimized for parallel processing and data paths on the GPU, as well as heterogeneous computing between GPU and CPU. The qualitative comparisons between results generated by the proposed algorithm and other existing algorithms show that the proposed one is able to generate more acceptable alpha maps and foreground colours especially in those regions that contain translucencies and details. And the quantitative evaluations also validate our advantages in both quality and speed.

<p dir="ltr">This thesis presents probably approximately correct (PAC) bounds on estimates of the transition kernels of Controlled Markov chains (CMC’s). CMC’s are a natural choice for modelling various industrial and medical processes, and are also relevant to reinforcement learning (RL). Learning the transition dynamics of CMC’s in a sample efficient manner is an important question that is open. This thesis aims to close this gap in knowledge in literature.</p><p dir="ltr">In Chapter 2, we lay the groundwork for later chapters by introducing the relevant concepts and defining the requisite terms. The two subsequent chapters focus on non-parametric estimation. </p><p dir="ltr">In Chapter 3, we restrict ourselves to a finitely supported CMC with d states and k controls and produce a general theory for minimax sample complexity of estimating the transition matrices.</p><p dir="ltr">In Chapter 4 we demonstrate the applicability of this theory by using it to recover the sample complexities of various controlled Markov chains, as well as RL problems.</p><p dir="ltr">In Chapter 5 we move to a continuous state and action spaces with compact supports. We produce a robust, non-parametric test to find the best histogram based estimator of the transition density; effectively reducing the problem into one of model selection based on empricial processes.</p><p dir="ltr">Finally, in Chapter 6 we move to a parametric and Bayesian regime, and restrict ourselves to Markov chains. Under this setting we provide a PAC-Bayes bound for estimating model parameters under tempered posteriors.</p>

À partir des observations Z(n) = {(Xi, Yi), i = 1, ..., n} satisfaisant Yi = f(Xi) + ζi, nous voulons reconstruire la fonction f. Nous évaluons la qualité d'estimation par deux critères : le risque Ls et le risque uniforme. Dans ces deux cas, les hypothèses imposées sur la distribution du bruit ζi serons de moment borné et de type sous-gaussien respectivement. En proposant une collection des estimateurs à noyau, nous construisons une procédure, qui est initié par Goldenshluger et Lepski, pour choisir l'estimateur dans cette collection, sans aucune condition sur f. Nous prouvons ensuite que cet estimateur satisfait une inégalité d'oracle, qui nous permet d'obtenir les estimations minimax et minimax adaptatives sur les classes de Hölder anisotropes. / From the observation Z(n) = {(Xi, Yi), i = 1, ..., n} satisfying Yi = f(Xi) + ζi, we would like to approximate the function f. This problem will be considered in two cases of loss function, Ls-risk and uniform risk, where the condition imposed on the distribution of the noise ζi is of bounded moment and of type sub-gaussian, respectively. From a proposed family of kernel estimators, we construct a procedure, which is initialized by Goldenshluger and Lepski, to choose in this family a final estimator, with no any assumption imposed on f. Then, we show that this estimator satisfies an oracle inequality which implies the minimax and minimax adaptive estimation over the anisotropic Hölder classes.

Die vorliegende Arbeit hat nichtparametrische Schätzmethoden für diskret beobachtete Lévyprozesse zum Gegenstand. Ein Lévyprozess mit endlichen zweiten Momenten und endlicher Variation auf Kompakta wird niederfrequent beobachtet. Die Sprungdynamik wird vollständig durch das endliche signierte Maß my(dx):= x ny(dx) beschrieben. Ein lineares Funktional von my soll nichtparametrisch geschätzt werden. Im ersten Teil werden Kernschätzer konstruiert und obere Schranken für das korrespondierende Risiko bewiesen. Daraus werden Konvergenzraten unter Glattheitsannahmen an das Lévymaß hergeleitet. Für Spezialfälle werden untere Schranken bewiesen und daraus Minimax-Optimalität gefolgert. Der Schwerpunkt liegt auf dem Problem der datengetriebenen Wahl des Glättungsparameters, das im zweiten Teil untersucht wird. Da die nichtparametrische Schätzung für Lévyprozesse starke strukturelle Ähnlichkeiten mit Dichtedekonvolutionsproblemen mit unbekannter Fehlerdichte aufweist, werden beide Problemstellungen parallel diskutiert und die Methoden allgemein sowohl für Lévyprozesse als auch für Dichtedekonvolution entwickelt. Es werden Methoden der Modellwahl durch Penalisierung angewandt. Während das Prinzip der Modellwahl im üblichen Fall darauf beruht, dass die Fluktuation stochastischer Terme durch Penalisierung mit einer deterministischen Größe beschränkt werden kann, ist die Varianz im hier betrachteten Fall unbekannt und der Strafterm somit stochastisch. Das Hauptaugenmerk der Arbeit liegt darauf, Strategien zum Umgang mit dem stochastischen Strafterm zu entwickeln. Dabei ist ein modifizierter Schätzer für die charakteristische Funktion im Nenner zentral, der es erlaubt, die punktweise Kontrolle der Abweichung dieses Objects von seiner Zielgröße auf die gesamte reelle Achse zu erweitern. Für die Beweistechnik sind insbesondere Talagrand-Konzentrationsungleichungen für empirische Prozesse relevant. / This thesis deals with nonparametric estimation methods for discretely observed Lévy processes. A Lévy process X having finite variation on compact sets and finite second moments is observed at low frequency. The jump dynamics is fully described by the finite signed measure my(dx)=x ny(dx). The goal is to estimate, nonparametrically, some linear functional of my. In the first part, kernel estimators are constructed and upper bounds on the corresponding risk are provided. From this, rates of convergence are derived, under regularity assumptions on the Lévy measure. For particular cases, minimax lower bounds are proved. The rates of convergence are thus shown to be minimax optimal. The focus lies on the data driven choice of the smoothing parameter, which is being considered in the second part. Since nonparametric estimation methods for Lévy processes have strong structural similarities with with nonparametric density deconvolution with unknown error density, both fields are discussed in parallel and the concepts are developed in generality, for Lévy processes as well as for density deconvolution. The choice of the bandwidth is realized, using techniques of model selection via penalization. The principle of model selection via penalization usually relies on the fact that the fluctuation of certain stochastic quantities can be controlled by penalizing with a deterministic term. Contrarily to this, the variance is unknown in the setting investigated here and the penalty term is hence itself a stochastic quantity. It is the main concern of this thesis to develop strategies to dealing with the stochastic penalty term. The most important step in this direction will be a modified estimator of the unknown characteristic function in the denominator, which allows to make the pointwise control of this object uniform on the real line. The main technical tools involved in the arguments are concentration inequalities of Talagrand type for empirical processes.

Die Schätzung von Quantilen und verwandten Funktionalen wird in zwei inversen Problemen behandelt: dem klassischen Dekonvolutionsmodell sowie dem Lévy-Modell in dem ein Lévy-Prozess beobachtet wird und Funktionale des Sprungmaßes geschätzt werden. Im einem abstrakteren Rahmen wird semiparametrische Effizienz im Sinne von Hájek-Le Cam für Funktionalschätzung in regulären, inversen Modellen untersucht. Ein allgemeiner Faltungssatz wird bewiesen, der auf eine große Klasse von statistischen inversen Problem anwendbar ist. Im Dekonvolutionsmodell beweisen wir, dass die Plugin-Schätzer der Verteilungsfunktion und der Quantile effizient sind. Auf der Grundlage von niederfrequenten diskreten Beobachtungen des Lévy-Prozesses wird im nichtlinearen Lévy-Modell eine Informationsschranke für die Schätzung von Funktionalen des Sprungmaßes hergeleitet. Die enge Verbindung zwischen dem Dekonvolutionsmodell und dem Lévy-Modell wird präzise beschrieben. Quantilschätzung für Dekonvolutionsprobleme wird umfassend untersucht. Insbesondere wird der realistischere Fall von unbekannten Fehlerverteilungen behandelt. Wir zeigen unter minimalen und natürlichen Bedingungen, dass die Plugin-Methode minimax optimal ist. Eine datengetriebene Bandweitenwahl erlaubt eine optimale adaptive Schätzung. Quantile werden auf den Fall von Lévy-Maßen, die nicht notwendiger Weise endlich sind, verallgemeinert. Mittels äquidistanten, diskreten Beobachtungen des Prozesses werden nichtparametrische Schätzer der verallgemeinerten Quantile konstruiert und minimax optimale Konvergenzraten hergeleitet. Als motivierendes Beispiel von inversen Problemen untersuchen wir ein Finanzmodell empirisch, in dem ein Anlagengegenstand durch einen exponentiellen Lévy-Prozess dargestellt wird. Die Quantilschätzer werden auf dieses Modell übertragen und eine optimale adaptive Bandweitenwahl wird konstruiert. Die Schätzmethode wird schließlich auf reale Daten von DAX-Optionen angewendet. / The estimation of quantiles and realated functionals is studied in two inverse problems: the classical deconvolution model and the Lévy model, where a Lévy process is observed and where we aim for the estimation of functionals of the jump measure. From a more abstract perspective we study semiparametric efficiency in the sense of Hájek-Le Cam for functional estimation in regular indirect models. A general convolution theorem is proved which applies to a large class of statistical inverse problems. In particular, we consider the deconvolution model, where we prove that our plug-in estimators of the distribution function and of the quantiles are efficient. In the nonlinear Lévy model based on low-frequent discrete observations of the Lévy process, we deduce an information bound for the estimation of functionals of the jump measure. The strong relationship between the Lévy model and the deconvolution model is given a precise meaning. Quantile estimation in deconvolution problems is studied comprehensively. In particular, the more realistic setup of unknown error distributions is covered. Under minimal and natural conditions we show that the plug-in method is minimax optimal. A data-driven bandwidth choice yields optimal adaptive estimation. The concept of quantiles is generalized to the possibly infinite Lévy measures by considering left and right tail integrals. Based on equidistant discrete observations of the process, we construct a nonparametric estimator of the generalized quantiles and derive minimax convergence rates. As a motivating financial example for inverse problems, we empirically study the calibration of an exponential Lévy model for asset prices. The estimators of the generalized quantiles are adapted to this model. We construct an optimal adaptive quantile estimator and apply the procedure to real data of DAX-options.

"Fluid-structure interaction arises in a wide array of technological applications including naval and marine hydrodynamics, civil and wind engineering and flight vehicle aerodynamics. When a fluid flows over a bluff body such as a circular cylinder, the periodic vortex shedding in the wake causes fluctuating lift and drag forces on the body. This phenomenon can lead to fatigue damage of the structure due to large amplitude vibration. It is widely believed that the wake structures behind the structure determine the hydrodynamic forces acting on the structure and control of wake structures can lead to vibration control of the structure. Modeling this complex non-linear interaction requires coupling of the dynamics of the fluid and the structure. In this thesis, however, the vibration of the flexible cylinder is prescribed, and the focus is on modeling the fluid dynamics in its wake. Low-dimensional iterative circle maps have been found to predict the universal dynamics of a two-oscillator system such as the rigid cylinder wake. Coupled map lattice (CML)models that combine a series of low-dimensional circle maps with a diffusion model have previously predicted qualitative features of wake patterns behind freely vibrating cables at low Reynolds number. However, the simple nature of the CML models implies that there will always be unmodelled wake dynamics if a detailed, quantitative comparison is made with laboratory or simulated wake flows. Motivated by a desire to develop an improved CML model, we incorporate self-learning features into a new CML that is trained to precisely estimate wake patterns from target numerical simulations and experimental wake flows. The eventual goal is to have the CML learn from a laboratory flow in real time. A real-time self-learning CML capable of estimating experimental wake patterns could serve as a wake model in a future anticipated feedback control system designed to produce desired wake patterns. A new convective-diffusive map that includes additional wake dynamics is developed. Two different self-learning CML models, each capable of precisely estimating complex wake patterns, have been developed by considering additional dynamics from the convective-diffusive map. The new self-learning CML models use adaptive estimation schemes which seek to precisely estimate target wake patterns from numerical simulations and experiments. In the first self-learning CML, the estimator scheme uses a multi-variable least-squares algorithm to adaptively vary the spanwise velocity distribution in order to minimize the state error (difference between modeled and target wake patterns). The second self-learning model uses radial basis function neural networks as online approximators of the unmodelled dynamics. Additional unmodelled dynamics not present in the first self-learning CML model are considered here. The estimator model uses a combination of a multi-variable normalized least squares scheme and a projection algorithm to adaptively vary the neural network weights. Studies of this approach are conducted using wake patterns from spectral element based NEKTAR simulations of freely vibrating cable wakes at low Reynolds numbers on the order of 100. It is shown that the self-learning models accurately and efficiently estimate the simulated wake patterns within several shedding cycles. Next, experimental wake patterns behind different configurations of rigid cylinders were obtained. The self-learning CML models were then used for off-line estimation of the stored wake patterns. With the eventual goal of incorporating low-order CML models into a wake pattern control system in mind, in a related study control terms were added to the simple CML model in order to drive the wake to the desired target pattern of shedding. Proportional, adaptive proportional and non-linear control techniques were developed and their control efficiencies compared."

The role of vision as an additional sensing mechanism has received a lot of attention in recent years in the context of autonomous flight applications. Modern Unmanned Aerial Vehicles (UAVs) are equipped with vision sensors because of their light-weight, low-cost characteristics and also their ability to provide a rich variety of information of the environment in which the UAVs are navigating in. The problem of vision based autonomous flight is very difficult and challenging since it requires bringing together concepts from image processing and computer vision, target tracking and state estimation, and flight guidance and control. This thesis focuses on the adaptive state estimation, guidance and control problems involved in vision-based formation flight. Specifically, the thesis presents a composite adaptation approach to the partial state estimation of a class of nonlinear systems with unmodeled dynamics. In this approach, a linear time-varying Kalman filter is the nominal state estimator which is augmented by the output of an adaptive neural network (NN) that is trained with two error signals. The benefit of the proposed approach is in its faster and more accurate adaptation to the modeling errors over a conventional approach. The thesis also presents two approaches to the design of adaptive guidance and control (G&C) laws for line-of-sight formation flight. In the first approach, the guidance and autopilot systems are designed separately and then combined together by assuming time-scale separation. The second approach is based on integrating the guidance and autopilot design process. The developed G&C laws using both approaches are adaptive to unmodeled leader aircraft acceleration and to own aircraft aerodynamic uncertainties. The thesis also presents theoretical justification based on Lyapunov-like stability analysis for integrating the adaptive state estimation and adaptive G&C designs. All the developed designs are validated in nonlinear, 6DOF fixed-wing aircraft simulations. Finally, the thesis presents a decentralized coordination strategy for vision-based multiple-aircraft formation control. In this approach, each aircraft in formation regulates range from up to two nearest neighboring aircraft while simultaneously tracking nominal desired trajectories common to all aircraft and avoiding static obstacles.

No description available.