Cette thèse se focalise sur la reconnaissance automatique de la parole (RAP) robuste au bruit. Elle comporte deux parties. Premièrement, nous nous focalisons sur une meilleure prise en compte des incertitudes pour améliorer la performance de RAP en environnement bruité. Deuxièmement, nous présentons une méthode pour accélérer l'apprentissage d'un réseau de neurones en utilisant une fonction auxiliaire. Dans la première partie, une technique de rehaussement multicanal est appliquée à la parole bruitée en entrée. La distribution a posteriori de la parole propre sous-jacente est alors estimée et représentée par sa moyenne et sa matrice de covariance, ou incertitude. Nous montrons comment propager la matrice de covariance diagonale de l'incertitude dans le domaine spectral à travers le calcul des descripteurs pour obtenir la matrice de covariance pleine de l'incertitude sur les descripteurs. Le décodage incertain exploite cette distribution a posteriori pour modifier dynamiquement les paramètres du modèle acoustique au décodage. La règle de décodage consiste simplement à ajouter la matrice de covariance de l'incertitude à la variance de chaque gaussienne. Nous proposons ensuite deux estimateurs d'incertitude basés respectivement sur la fusion et sur l'estimation non-paramétrique. Pour construire un nouvel estimateur, nous considérons la combinaison linéaire d'estimateurs existants ou de fonctions noyaux. Les poids de combinaison sont estimés de façon générative en minimisant une mesure de divergence par rapport à l'incertitude oracle. Les mesures de divergence utilisées sont des versions pondérées des divergences de Kullback-Leibler (KL), d'Itakura-Saito (IS) ou euclidienne (EU). En raison de la positivité inhérente de l'incertitude, ce problème d'estimation peut être vu comme une instance de factorisation matricielle positive (NMF) pondérée. De plus, nous proposons deux estimateurs d'incertitude discriminants basés sur une transformation linéaire ou non linéaire de l'incertitude estimée de façon générative. Cette transformation est entraînée de sorte à maximiser le critère de maximum d'information mutuelle boosté (bMMI). Nous calculons la dérivée de ce critère en utilisant la règle de dérivation en chaîne et nous l'optimisons par descente de gradient stochastique. Dans la seconde partie, nous introduisons une nouvelle méthode d'apprentissage pour les réseaux de neurones basée sur une fonction auxiliaire sans aucun réglage de paramètre. Au lieu de maximiser la fonction objectif, cette technique consiste à maximiser une fonction auxiliaire qui est introduite de façon récursive couche par couche et dont le minimum a une expression analytique. Grâce aux propriétés de cette fonction, la décroissance monotone de la fonction objectif est garantie / This thesis focuses on noise robust automatic speech recognition (ASR). It includes two parts. First, we focus on better handling of uncertainty to improve the performance of ASR in a noisy environment. Second, we present a method to accelerate the training process of a neural network using an auxiliary function technique. In the first part, multichannel speech enhancement is applied to input noisy speech. The posterior distribution of the underlying clean speech is then estimated, as represented by its mean and its covariance matrix or uncertainty. We show how to propagate the diagonal uncertainty covariance matrix in the spectral domain through the feature computation stage to obtain the full uncertainty covariance matrix in the feature domain. Uncertainty decoding exploits this posterior distribution to dynamically modify the acoustic model parameters in the decoding rule. The uncertainty decoding rule simply consists of adding the uncertainty covariance matrix of the enhanced features to the variance of each Gaussian component. We then propose two uncertainty estimators based on fusion to nonparametric estimation, respectively. To build a new estimator, we consider a linear combination of existing uncertainty estimators or kernel functions. The combination weights are generatively estimated by minimizing some divergence with respect to the oracle uncertainty. The divergence measures used are weighted versions of Kullback-Leibler (KL), Itakura-Saito (IS), and Euclidean (EU) divergences. Due to the inherent nonnegativity of uncertainty, this estimation problem can be seen as an instance of weighted nonnegative matrix factorization (NMF). In addition, we propose two discriminative uncertainty estimators based on linear or nonlinear mapping of the generatively estimated uncertainty. This mapping is trained so as to maximize the boosted maximum mutual information (bMMI) criterion. We compute the derivative of this criterion using the chain rule and optimize it using stochastic gradient descent. In the second part, we introduce a new learning rule for neural networks that is based on an auxiliary function technique without parameter tuning. Instead of minimizing the objective function, this technique consists of minimizing a quadratic auxiliary function which is recursively introduced layer by layer and which has a closed-form optimum. Based on the properties of this auxiliary function, the monotonic decrease of the new learning rule is guaranteed.

Ce travail de thèse concerne la synthèse des observateurs et des lois de commande des systèmes d’ordre fractionnaire. Le document présenté est constitué de 4 chapitres : Le premier chapitre du manuscrit de thèse contient une introduction, traitant les notions mathématiques de base et de stabilités des systèmes d’ordre fractionnaire ainsi qu’une présentation des différentes définitions. Les conditions de stabilités de ces systèmes et quelques exemples de systèmes modélisés par des équations différentielles fractionnaires sont présentés. Dans le deuxième chapitre, nous nous sommes intéressés à la conception de plusieurs types d’observateurs dits d’ordre réduit, d’ordre plein et des observateurs fonctionnels pour les systèmes d’ordre fractionnaire avec et sans retards. Dans le cas où il n’y a pas de retards dans la dynamique du système, des observateurs d’ordres plein et réduit ont été synthétisé afin d’assurer l’estimation des pseudo-états. Dans un deuxième temps, un observateur fonctionnel a été synthétisé dans le cas où le retard est présent dans la dynamique du système. Dans le chapitre 3, nous avons travaillé sur la synthèse d’observateur pour les systèmes d’ordre fractionnaire incertains. Nos contributions sont classées en trois grandes lignes : premièrement, quand le système considéré est affecté par des entrées inconnues, un observateur fonctionnel a été proposé. En deuxième partie, des observateurs H∞ pour les systèmes d’ordre fractionnaire avec et sans retards ont été synthétisés afin d’assurer la stabilité de l’erreur d’observation. Il s’agit en fait de garantir une borne du gain L2 entre l’erreur d’observation et les perturbations non mesurables affectant la dynamique du système : ce gain L2 est aussi appelé norme H∞. Ce chapitre présente aussi la synthèse d’un observateur robuste vis-à-vis des incertitudes de modélisation pour cette classe de systèmes. Les conditions suffisantes de convergence des erreurs d’estimations des pseudo-états obtenues sont établies sous la forme d’un ensemble d’inégalités matricielles LMIs. Le dernier chapitre du manuscrit est consacré à la commande basée sur les différents observateurs obtenus. Nous nous sommes intéressés à la commande basée sur un observateur pour les systèmes d’ordre fractionnaire. Cette commande est basée sur les observateurs proposés dans les chapitres précédents. Des conditions de stabilité et des procédures de synthèse sont présentées / This work focuses on the synthesis of observers and the controller laws for fractional order systems. The presented document consists of 4 chapters: The first chapter of the theses manuscript contains an introduction dealing with the basic mathematical notions and the stability analysis of fractional systems as well as a presentation of the different definitions. The stability conditions of these systems and some examples of systems modeled by fractional differential equations are presented. In the second chapter, we were interested in the design of several types of observers of reduced order, full order, and functional observers for fractional systems with and without delays. In the case where there are no delays in the dynamics of the system, observers of full and reduced orders have been synthesized in order to ensure the estimation of the pseudo-states. In a second step, a functional observer was synthesized in the case where the delay is present in the dynamics of the system. In Chapter 3, we worked on observer synthesis for uncertain fractional order systems. Our contributions are classified into three main lines: first, when the system under consideration is affected by unknown inputs, a functional observer has been proposed. In the second part, H∞ observers for fractional order systems with and without delays have been synthesized to ensure the stability of the estimation error. It is a question of guaranteeing a bound of the L2 gain between the observation error and the non-measurable perturbations affecting the dynamics of the system: this gain L2 is also called H∞ norm. In last part of this chapter, the synthesis of a robust observer with respect to modeling uncertainties for this class of systems is presented. The sufficient conditions of convergence of the estimation errors of the pseudo-states obtained are established in the form of a set of matrix inequalities LMIs. The last chapter of the manuscript is devoted to the command based on the different observers obtained. We were interested in observer-based control for fractional order systems. This command is based on the observers proposed in the previous chapters. Stability conditions and synthesis procedures are presented

En conception de produits ou de systèmes, les approches d'optimisation déterministe sont de nos jours largement utilisées. Toutefois, ces approches ne tiennent pas compte des incertitudes inhérentes aux modèles utilises, ce qui peut parfois aboutir à des solutions non fiables. Il convient alors de s'intéresser aux approches d'optimisation stochastiques. Les approches de conception robuste à base d'optimisation stochastique (Reliablity Based Robust Design Optimization, RBRDO) tiennent compte des incertitudes lors de l'optimisation au travers d'une boucle supplémentaire d'analyse des incertitudes(Uncertainty Anlysis, UA). Pour la plupart des applications pratiques, l'UA est réalisée par une simulation de type Monte Carlo (Monte Carlo Simulation, MCS) combinée avec l’analyse structurale. L'inconvénient majeur de ce type d'approche réside dans le coût de calcul qui se révèle être prohibitif. Par conséquent, nous nous sommes intéressés dans nos travaux aux développements de méthodologies efficaces pour la mise en place de RBRDO s'appuyant sur une analyse MCS. Nous présentons une méthode d'UA s'appuyant sur une analyse MCS dans laquelle la réponse aléatoire est approximée sur une base du chaos polynomial (Polynomial Chaos Expansion, PCE). Ainsi, l'efficacité de l'UA est grandement améliorée en évitant une trop grande répétition des analyses structurales. Malheureusement, cette approche n'est pas pertinente dans le cadre de problèmes en grande dimension, par exemple pour des applications en dynamique. Nous proposons ainsi d'approximer la réponse dynamique en ne tenant compte que de la résolution aux valeurs propres aléatoires. De cette façon, seuls les paramètres structuraux aléatoires apparaissent dans le PCE. Pour traiter le problème du mélange des modes dans notre approche, nous nous sommes appuyés sur le facteur MAC qui permet de le quantifier. Nous avons développé une méthode univariable permettant de verifier quelle variable générait un mélange de modes de manière à le réduire ou le supprimer. Par la suite, nous présentons une approche de RBRDO séquentielle pour améliorer l'efficacité et éviter les problèmes de non-convergence présents dans les approches de RBRDO. Dans notre approche, nous avons étendu la stratégie séquentielle classique, visant principalement à découpler l'analyse de fiabilité de la procédure d'optimisation, en séparant l'évaluation des moments de la boucle d'optimisation. Nous avons utilisé une approximation exponentielle locale autour du point de conception courant pour construire des objectifs déterministes équivalents ainsi que des contraintes stochastiques. De manière à obtenir les différents coefficients pour notre approximation, nous avons développé une analyse de sensibilité de la robustesse basée sur une distribution auxiliaire ainsi qu'une analyse de sensibilité des moments basée sur l'approche PCE. Nous montrons la pertinence ainsi que l'efficacité des approches proposées au travers de différents exemples numériques. Nous appliquons ensuite notre approche de RBRDO pour la conception d'un amortisseur dans le domaine du contrôle passif vibratoire d'une structure présentant des grandeurs aléatoires. Les résultats obtenus par notre approche permettent non seulement de réduire la variabilité de la réponse, mais aussi de mieux contrôler l'amplitude de la réponse au travers d'un seuil choisi par avance. / Deterministic design optimization is widely used to design products or systems. However, due to the inherent uncertainties involved in different model parameters or operation processes, deterministic design optimization without considering uncertainties may result in unreliable designs. In this case, it is necessary to develop and implement optimization under uncertainties. One way to deal with this problem is reliability-based robust design optimization (RBRDO), in which additional uncertainty analysis (UA, including both of reliability analysis and moment evaluations) is required. For most practical applications however, UA is realized by Monte Carlo Simulation (MCS) combined with structural analyses that renders RBRDO computationally prohibitive. Therefore, this work focuses on development of efficient and robust methodologies for RBRDO in the context of MCS. We presented a polynomial chaos expansion (PCE) based MCS method for UA, in which the random response is approximated with the PCE. The efficiency is mainly improved by avoiding repeated structural analyses. Unfortunately, this method is not well suited for high dimensional problems, such as dynamic problems. To tackle this issue, we applied the convolution form to compute the dynamic response, in which the PCE is used to approximate the modal properties (i.e. to solve random eigenvalue problem) so that the dimension of uncertainties is reduced since only structural random parameters are considered in the PCE model. Moreover, to avoid the modal intermixing problem when using MCS to solve the random eigenvalue problem, we adopted the MAC factor to quantify the intermixing, and developed a univariable method to check which variable results in such a problem and thereafter to remove or reduce this issue. We proposed a sequential RBRDO to improve efficiency and to overcome the nonconvergence problem encountered in the framework of nested MCS based RBRDO. In this sequential RBRDO, we extended the conventional sequential strategy, which mainly aims to decouple the reliability analysis from the optimization procedure, to make the moment evaluations independent from the optimization procedure. Locally "first-torder" exponential approximation around the current design was utilized to construct the equivalently deterministic objective functions and probabilistic constraints. In order to efficiently calculate the coefficients, we developed the auxiliary distribution based reliability sensitivity analysis and the PCE based moment sensitivity analysis. We investigated and demonstrated the effectiveness of the proposed methods for UA as well as RBRDO by several numerical examples. At last, RBRDO was applied to design the tuned mass damper (TMD) in the context of passive vibration control, for both deterministic and uncertain structures. The associated optimal designs obtained by RBRDO cannot only reduce the variability of the response, but also control the amplitude by the prescribed threshold.

In der vorliegenden Arbeit werden zwei Modelle zur strukturellen Analyse von Stoffwechselsystemen vorgestellt. Die Untersuchung basiert auf der Hypothese, dass heutzutage vorzufindende Stoffwechselsysteme als Ergebnis einer evolutionären Entwicklung, bestimmt durch Mutationsmechanismen und natürlicher Selektion, angesehen werden können. Es kann daher angenommen werden, dass kinetische Parameter sowie strukturelle Eigenschaften im Laufe der Evolution solche Werte angenommen haben, die eine gewisse Optimalität bezüglich ihrer biologischen Funktion darstellen. Das erste Modell untersucht das strukturelle Design ATP und NADH produzierender Systeme, so wie die Glykolyse und der Zitratzyklus. Eine Methode wird präsentiert, die die Beschreibung hypothetischer, chemisch denkbarer, alternativer Stoffwechselwege ermöglicht. Diese Wege werden bezüglich ihrer Effizienz, ATP zu produzieren, untersucht. Es stellt sich heraus, dass die meisten möglichen Wege eine niedrige ATP-Produktionsrate aufweisen und dass die effizientesten Wege einige strukturelle Gemeinsamkeiten besitzen. Die Optimierung bezüglich der ATP-Produktionsrate wird mit einem evolutionären Algorithmus durchgeführt. Folgende Resultate stehen mit dem tatsächlichen Design der Glykolyse und des Zitratzyklus in Einklang: (i) In allen effizienten Wegen befinden sich die ATP-verbrauchenden Reaktionen am Anfang. (ii) In allen effizienten Wegen befinden sich die sowohl die NADH- als auch die ATP-produzierenden Reaktionen am Ende. (iii) Die Anzahl der NADH-Moleküle, die aus einem energiereichen Molekül (Glukose) produziert werden, beläuft sich in allen effizienten Wegen auf vier. Im zweiten Modell werden vollständige Mengen metabolischer Netzwerke konstruiert, wobei von Reaktionen ausgegangen wird, die Änderungen des Kohlenstoffskeletts der beteiligten Metabolite beschreiben. Elementare Netzwerke werden dadurch definiert, dass eine bestimmte chemische Umwandlung durchgeführt werden kann und dass diese Fähigkeit verloren geht, wenn eine der beteiligten Reaktionen ausgeschlossen wird. Übergänge zwischen Netzwerken und Mutationen werden durch den Austausch einer einzigen Reaktion definiert. Es existieren verschiedene Mutationen, solche bei denen Funktionen verloren gehen, welche dazugewonnen werden, und neutrale Mutationen. Mutationen definieren Nachbarschaftsrelationen, die graphentheoretisch beschrieben werden. Eigenschaften wie Durchmesser, Konnektivität und die Abstandsverteilung der Vertizes werden berechnet. Ein Konzept zur Quantifizierung der Robustheit von Netzwerken gegenüber stöchiometrischen Veränderungen wird entwickelt, wobei zwischen starker und schwacher Robustheit unterschieden wird. Evolutionäre Algorithmen werden angewandt, um die Entwicklung von Netzwerkpopulationen unter konstanten und zeitlich veränderlichen Umweltbedingungen zu untersuchen. Es wird gezeigt, dass Populationen sich zu Gruppierungen von Netzwerken hinentwickeln, die gemeinsame Funktionen besitzen und nah benachbart sind. Unter zeitlich veränderlichen Umweltbedingungen zeigt sich, dass multifunktionelle Netzwerke optimal sind und sich im Selektionsprozess durchsetzen. / In the present thesis two models are presented which study the structural design of metabolic systems. The investigation is based on the hypothesis that present day metabolic systems are the result of an evolutionary development governed by mutation mechanisms and natural selection principles. Therefore, it can be assumed that these parameters have reached, during the course of their evolution, values which imply certain optimal properties with respect to their biological function. The first model concerns the structural design of ATP and NADH producing systems such as glycolysis and the citric acid cycle. A method is presented to describe hypothetical, chemically feasible, alternative pathways. We analyse these pathways with respect to their capability to efficiently produce ATP. It is shown that most of the possible pathways result in a very low ATP production rate and that the very efficient pathways share common structural properties. Optimisation with respect to the ATP production rate is performed by an evolutionary algorithm. The following results of our analysis are in close correspondence to the real design of glycolysis and the TCA cycle: (i) In all efficient pathways the ATP consuming reactions are located near the beginning. (ii) In all efficient pathways NADH producing reactions as well as ATP producing reactions are located near the end. (iii) The number of NADH molecules produced by the consumption of one energy-rich molecule (glucose) amounts to four in all efficient pathways. In the second model complete sets of metabolic networks are constructed starting from a limited set of reactions describing changes in the carbon skeleton of biochemical compounds. Elementary networks are defined by the condition that a specific chemical conversion can be performed by a set of given reactions and that this ability will be lost by elimination of any of these reactions. Transitions between networks and mutations of networks are defined by exchanges of single reactions. Different mutations exist such as gain or loss of function mutations and neutral mutations. Based on these mutations neighbourhood relations between networks are established which are described in a graph theoretical way. Basic properties of these graphs are determined such as diameter, connectedness, distance distribution of pairs of vertices. A concept is developed to quantify the robustness of networks against changes in their stoichiometry where we distinguish between strong and weak robustness. Evolutionary algorithms are applied to study the development of network populations under constant and time dependent environmental conditions. It is shown that the populations evolve toward clusters of networks performing a common function and which are closely neighboured. Under changing environmental conditions multifunctional networks prove to be optimal and will be selected.

Esta tese de doutorado aborda os projetos robustos de controle e estimativa de estados para Sistemas Lineares sujeitos a Saltos Markovianos (SLSM) de tempo discreto sob a influência de incertezas paramétricas. Esses projetos são desenvolvidos por meio de extensões dos critérios quadráticos clássicos para SLSM nominais. Os critérios de custo quadrático para os SLSM incertos são formulados na forma de problemas de otimização min-max que permitem encontrar a melhor solução para o pior caso de incerteza (máxima influência de incerteza). Os projetos robustos correspondem às soluções ótimas obtidas por meio da combinação dos métodos de funções penalidade e mínimos quadrados regularizados robustos. Duas situações são investigadas: regular e estimar os estados quando os modos de operações são observados; e estimar os estados sob a hipótese de desconhecimento da cadeia de Markov. Estruturalmente, o regulador e as estimativas de estados assemelham-se às respectivas versões nominais. A recursividade é estabelecida em termos de equações de Riccati sem a necessidade de ajuste de parâmetros auxiliares e dependente apenas das matrizes de parâmetros e ponderações conhecidas. / This thesis deals with recursive robust designs of control and state estimates for discretetime Markovian Jump Linear Systems (MJLS) subject to parametric uncertainties. The designs are developed considering extensions of the standard quadratic cost criteria for MJLS without uncertainties. The quadratic cost criteria for uncertain MJLS are formulated in the form of min-max optimization problems to get the best solution for the worst uncertainty case. The optimal robust schemes correspond to the optimal solution obtained by the combination of penalty function and robust regularized least-squares methods. Two cases are investigated: to control and estimate the states when the operation modes are observed; and, to estimate the states when the Markov chain is unobserved. The optimal robust LQR and Kalman-type state estimates resemble the respective nominal versions. The recursiveness is established by Riccati equations in terms of parameter and weighting matrices previously known and without extra offline computations.

Submitted by Eunice Novais (enovais@uepg.br) on 2018-04-18T18:15:46Z No. of bitstreams: 2 license_rdf: 811 bytes, checksum: e39d27027a6cc9cb039ad269a5db8e34 (MD5) Santos, Moises Souza.pdf: 3654988 bytes, checksum: 6bcf0d0867a3d40291ae626664114a49 (MD5) / Made available in DSpace on 2018-04-18T18:15:46Z (GMT). No. of bitstreams: 2 license_rdf: 811 bytes, checksum: e39d27027a6cc9cb039ad269a5db8e34 (MD5) Santos, Moises Souza.pdf: 3654988 bytes, checksum: 6bcf0d0867a3d40291ae626664114a49 (MD5) Previous issue date: 2018-03-01 / Conselho Nacional de Desenvolvimento Científico e Tecnológico / A importância do estudo de estados quimera em redes neuronais reflete no fato destes possuírem uma forte ligação com alguns tipos de anomalias diagnosticadas no cérebro, por exemplo, epilepsia,mal de Parkinson e Alzheimer. Neste trabalho, realizamos um estudo da aplicação dos gráficos de recorrência, em sua versão espacial, na caracterização da coexistência de estados coerentes e incoerentes em sistemas dinâmicos acoplados. Utilizamos um modelo constituído de osciladores de fase para uma primeira abordagem da matriz de recorrência. Nossos resultados indicam que os diagnósticos quantitativos, baseados no gráfico de recorrência, detectam não só a existência de estados quimera como também estados sincronizados e dessincronizados que por ventura emergem na rede. Sobre esse aspecto, mostramos que a análise via recorrência espacial é superior à utilização do parâmetro de ordem global pós-colapso de estados quimera. Este cenário ocorre em algumas situações onde a rede exibe quimera e esta subitamente desaparece do sistema, que passa a exibir um comportamento sincronizado na frequência de seus osciladores. Neste caso, o parâmetro de ordem global não identifica esta mudança de comportamento sendo necessário uma interpretação via outro diagnóstico. Também utilizamos a rede de conectividade cortical de gato, como forma de acoplamento, e distinguimos duas formas de estados híbridos: uma quimera com a região incoerente em dinâmica temporal de spikes e outra quimera onde a região incoerente apresenta dinâmica temporal de bursts. Identificamos que este último caso de estados quimera possui maior robustez com relação à perturbação externa aplicada no sistema.Concluímos que a região de quimera com bursts necessita de uma intensidade de perturbação duas vezes superior que a região de spikes para deixar de existir na rede de neurônios. / A importância do estudo de estados quimera em redes neuronais reflete no fato destes possuírem uma forte ligação com alguns tipos de anomalias diagnosticadas no cérebro, por exemplo, epilepsia, mal de Parkinson e Alzheimer. Neste trabalho, realizamos um estudo da aplicação dos gráficos de recorrência, em sua versão espacial, na caracterização da coexistência de estados coerentes e incoerentes em sistemas dinâmicos acoplados. Utilizamos um modelo constituído de osciladores de fase para uma primeira abordagem da matriz de recorrência. Nossos resultados indicam que os diagnósticos quantitativos, baseados no gráfico de recorrência, detectam não só a existência de estados quimera como também estados sincronizados e dessincronizados que por ventura emergem na rede. Sobre esse aspecto, mostramos que a análise via recorrência espacial é superior à utilização do parâmetro de ordem global pós-colapso de estados quimera. Este cenário ocorre em algumas situações onde a rede exibe quimera e esta subitamente desaparece do sistema, que passa a exibir um comportamento sincronizado na frequência de seus osciladores. Neste caso, o parâmetro de ordem global não identifica esta mudança de comportamento sendo necessário uma interpretação via outro diagnóstico. Também utilizamos a rede de conectividade cortical de gato, como forma de acoplamento, e distinguimos duas formas de estados híbridos: uma quimera com a região incoerente em dinâmica temporal de spikes e outra quimera onde a região incoerente apresenta dinâmica temporal de bursts. Identificamos que este último caso de estados quimera possui maior robustez com relação à perturbação externa aplicada no sistema.Concluímos que a região de quimera com bursts necessita de uma intensidade de perturbação duas vezes superior que a região de spikes para deixar de existir na rede de neurônios

Les systèmes d'Equations aux Dérivées Partielles Hyperboliques Linéaires du Premier Ordre (EDPs HLPO) permettent de modéliser de nombreux systèmes de lois de conservation. Ils apparaissent, par exemple, lors de la modélisation de problèmes de trafic routier, d'échangeurs de chaleurs, ou de problèmes multiphasiques. Différentes approches ont été proposées pour stabiliser ou observer de tels systèmes. Parmi elles, la méthode de backstepping consiste à transformer le système originel en un système découplé pour lequel la synthèse de la loi de commande est plus simple. Les contrôleur obtenus par cette méthode sont explicites.Dans la première partie de cette thèse, nous présentons des résultats généraux de théorie des systèmes. Plus précisément, nous résolvons les problèmes de stabilisation en temps fini pour une classe générale d'EDPs HLPO. Le temps de convergence minimal atteignable dépend du nombre d'actionneurs disponibles. Les observateurs associés à ces contrôleurs (nécessaires pour envisager une utilisation industrielle de tels contrôleurs) sont obtenus via une approche duale. Un des avantages importants de l'approche considérée dans cette thèse est de montrer que l'espace généré par les solutions de l'EDPs HLPO considérée est isomorphe à l'espace généré par les solutions d'une système neutre à retards distribués.Dans la seconde partie de cette thèse, nous montrons la nécessité d'un changement de stratégie pour résoudre les problèmes de contrôle robuste. Ces questions surviennent nécessairement lorsque sont considérées des applications industrielles pour lesquelles les différents paramètres du système peuvent être mal connus, pour lesquelles des dynamiques peuvent avoir été négligées, de même que des retards agissant sur la commande ou sur la mesure, ou encore pour lesquels les mesures sont bruitées. Nous proposons ainsi quelques modifications sur les lois de commande précédemment développées en y incorporant plusieurs degrés de liberté permettant d'effectuer un compromis entre performance et robustesse. L'analyse de stabilité et de robustesse sous-jacente est rendue possible en utilisant l'isomorphisme précédemment introduit. / Linear First-Order Hyperbolic Partial Differential Equations (LFOH PDEs) represent systems of conservation and balance law and are predominant in modeling of traffic flow, heat exchanger, open channel flow or multiphase flow. Different control approaches have been tackled for the stabilization or observation of such systems. Among them, the backstepping method consists to map the original system to a simpler system for which the control design is easier. The resulting controllers are explicit.In the first part of this thesis, we develop some general results in control theory. More precisely, we solve the problem of finite-time stabilization of a general class of LFOH PDEs using the backstepping methodology. The minimum stabilization time reachable may change depending on the number of available actuators. The corresponding boundary observers (crucial to envision industrial applications) are obtained through a dual approach. An important by-product of the proposed approach is to derive an explicit mapping from the space generated by the solutions of the considered LFOH PDEs to the space generated by the solutions of a general class of neutral systems with distributed delays. This mapping opens new prospects in terms of stability analysis for LFOH PDEs, extending the stability analysis methods developed for neutral systems.In the second part of the thesis, we prove the necessity of a change of strategy for robust control while considering industrial applications, for which the major limitation is known to be the robustness of the resulting control law to uncertainties in the parameters, delays in the loop, neglected dynamics or disturbances and noise acting on the system. In some situations, one may have to renounce to finite-time stabilization to ensure the existence of robustness margins. We propose some adjustments in the previously designed control laws by means of several degrees of freedom enabling trade-offs between performance and robustness. The robustness analysis is fulfilled using the explicit mapping between LFOH PDEs and neutral systems previously introduced.

A crescente capacidade do poder computacional e a proliferação de dispositivos ao nosso redor vem permitindo o desenvolvimento de novas e sofisticadas interfaces para interação humano-computador que reagem à presença e ao estado de seus usuários. Como o olhar tem a capacidade de transmitir muitas informações sobre o usuário, rastreadores de olhar, dispositivos que estimam a direção para onde uma pessoa olha, tem papel importante no desenvolvimento de tais interfaces. Entre suas aplicações temos o auxílio a pessoas com dificuldades motoras, que podem utilizar um rastreador de olhar como substituto ao mouse, aplicações de diagnóstico, que estudam evidências do comportamento humano, ou ainda o desenvolvimento de interfaces que utilizem a informação sobre o olhar como um canal a mais de comunicação com o usuário para perceber suas intenções. Muitas técnicas para atingir tal objetivo foram desenvolvidas mas as tradicionais ainda oferecem certas dificuldades de uso para seus usuários como a intolerância a movimentos de cabeça e a necessidade de calibração por sessão de uso. Neste trabalho fizemos um levantamento de uma série de técnicas de rastreamento de olhar, indo das mais tradicionais até algumas mais recentes que visam melhorar a facilidade de uso destes sistemas. Uma das técnicas mais promissoras utiliza múltiplas fontes de luz fixadas nos cantos do monitor do computador. Através da análise da posição dos reflexos gerados por essas fontes de luz sobre a córnea, juntamente com a informação da posição da pupila, presentes em imagens capturadas do olho, é possível estimar o ponto observado no monitor. Devido às suas vantagens ela foi escolhida para estudo mais detalhado e implementação. Extensos testes utilizando simulações foram realizados para avaliar seu desempenho. Foi também desenvolvida uma extensão dessa técnica, utilizando um modelo mais preciso do olho, visando melhorar sua precisão. Ao final apresentamos nossa implementação, baseada nessa extensão da técnica original, que é tolerante a movimentação da face e mostramos os resultados obtidos em testes realizados com um grupo de usuários. / Recent advances in computing power and the proliferation of computing devices around us allowed the development of new computer interfaces which can react to the presence and state of its users. Since gaze can transmit a lot of information about the user, gaze trackers, devices that can estimate the direction which a person is looking at, have an important role in the development of such interfaces. Among gaze tracking applications, we have aid for people with limited motor skills, human behavior studies, and the development of interfaces that can take gaze information as an additional communication channel with the user. Lots of techniques have been developed to reach this goal, but they have some problems that make them hard to be widely used. These problems are the need of calibration for each use session and the need to keep the user\'s head still. In this work we studied some of the existing gaze tracking techniques, from the more traditional ones to more recent ones. One of the most interesting techniques makes use of multiple light sources fixed at the monitor\'s corners. By analyzing the positions of corneal reflections and the pupil present in captured images of the eye, it is possible to estimate the gaze point on the monitor screen. Due to its advantages this technique was chosen for a deeper study and implementation. Lots of experiments using simulated data have been carried out to validate the technique. Using a more accurate model of the eye, an extension for this technique was also developed to increase its precision. Finally, we present our implementation, that allows for large head movement, as well as test results obtained from real users.

Submitted by Aline Batista (alinehb.ufpel@gmail.com) on 2017-05-05T17:33:32Z No. of bitstreams: 2 license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) Diferença Fuzzi intuicionista - robustez, dualidade e conjugação.pdf: 1119948 bytes, checksum: b732f8ac3cf0e1c348e50484c7cc1a05 (MD5) / Approved for entry into archive by Aline Batista (alinehb.ufpel@gmail.com) on 2017-05-05T22:17:27Z (GMT) No. of bitstreams: 2 license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) Diferença Fuzzi intuicionista - robustez, dualidade e conjugação.pdf: 1119948 bytes, checksum: b732f8ac3cf0e1c348e50484c7cc1a05 (MD5) / Made available in DSpace on 2017-05-05T22:17:38Z (GMT). No. of bitstreams: 2 license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) Diferença Fuzzi intuicionista - robustez, dualidade e conjugação.pdf: 1119948 bytes, checksum: b732f8ac3cf0e1c348e50484c7cc1a05 (MD5) Previous issue date: 2016-08-22 / Sem bolsa / Esta dissertação foca, sobretudo, nos conceitos fundamentais relativos ao estudo da robustez, dualidade e conjugação na Lógica Fuzzy (FL) e sua extensão intuicionista proposta por Atanassov (A-IFL). A metodologia de avaliação da sensibilidade ponto-a-ponto é aplicada a conectivos fuzzy e conectivos fuzzy intuicionistas, considerando a ação de negações fortes e automorfismos. O objetivo principal neste trabalho consiste na avaliação da robustez de operadores de diferença, representáveis por composição de negações e agregações da LF e da A-IFL. O operador de diferença tem aplicação direta em conceitos da FL e da A-IFL, quando do uso de conceitos de distância, medidas de similaridade e entropia. O trabalho colabora com a investigação da robustez na construção dual da classe de operadores de diferença em LF e A-IFL, incluindo possíveis construções conjugadas obtidas por automorfismos representáveis. / This dissertation focuses mainly on fundamental concepts relating to the study of robustness, duality and in conjunction Fuzzy Logic (FL) and its intuitionistic extension proposed by Atanassov (A-IFL). The methodology for assessing the sensitivity point-to-point is applied to fuzzy connectives and fuzzy connective intuitionists considering the action of strong denials and automorphisms. The main objective of this study is to assess the robustness of di?erence operators, representable by composition of denials and aggregations of LF and A-IFL. The di?erence operator has direct application of the concepts and the FL-IFL, when using distance concepts of similarity and entropy measures. The research work cooperates with the robustness of the dual construction di?erence operator class LF and A-IFL, including possible constructions conjugate obtained by automorphisms representable.

Made available in DSpace on 2014-06-11T19:30:54Z (GMT). No. of bitstreams: 0 Previous issue date: 2008-04-28Bitstream added on 2014-06-13T18:41:06Z : No. of bitstreams: 1 oliveira_lc_dr_sjrp.pdf: 1552331 bytes, checksum: ff9be69b536d540081c28a83bf038868 (MD5) / Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) / A topologia do estado nativo de uma proteína desempenha um papel crucial no processo de enovelamento. Neste trabalho uma nova aproximação utilizando aspectos topol ogicos para investigar a evolução protéica e apresentada. O modelo utiliza uma rede c ubica 3 3 3 de 27 monômeros e um mapa de conexões entre diferentes conformações em espa co de fase estrutural e de sequência. Desenhamos a melhor sequência não frustrada para cada uma das 103346 conformações maximamente compactas usando um algorítimo que maximiza o número de tipos de monômeros na sequência. Isto significa que cada sequência não pode possuir contatos desfavor aveis. O n umero m aximo de tipos de monômeros e 5. A sequência-conformação e considerada \protein-like se ela tem uma unica conformação de mais baixa energia, alem de acessibilidade e robustez. De todas as conformações maximamente compactas, somente 4; 75% geraram sequências \protein-like', o qual são o alvo neste estudo. Com esses dados realizamos simulações de Monte Carlo (MC) no qual examinamos as melhores sequêencias estruturas baseando-se no ZScore. A simulação e iniciada com uma sequência aleatória no qual e testada em todas as conformações, seguindo as regras estipuladas por MC. Se o ZScore aumenta, assumimos que a nova conformação e mais estável que a anterior. Esse processo e repetido até que as sequências otimamente desenhadas (com mais alto ZScore) são alcançadas. Mantendo as trajetórias originadas via MC, um mapa de conectividade sequências-estruturas e obtido. Os resultados mostram trajetórias conectadas com estruturas com baixos valores de ZScore. O aumento do ZScore ao longo da simulação conduz a um pequeno grupo de conformações preferenciais. O modelo sugere um funil de estruturas para a evolução de proteínas no qual as estruturas do fundo estão associadas com o ii \motif de uma proteína... / The topology of a protein native state plays a crucial role in the folding process. In this work a new approach using topological aspects to investigate the protein evolutions is presented. The model uses the 27-mer in a cubic lattice of 3 3 3, and a conection map between di erent conformations is found in the sequence and structural phase space. We designed the best unfrustrated sequence for each of the 103346 maximally compact conformation, using an algorithm that maximizes the number for monomers types in the sequence. This means that each sequence cannot have unfavorable contacts. The maximum number of types of monomer is 5. The sequence-conformation is considered protein-like if it has a unique lowest energy conformation, accessible and robust . Out of all maximally compact conformations, only 4,75% generated protein-like sequence, with are targeted in this study. With this data we performed a Monte Carlo simulations in which we probe for better sequence-structure based on Zscore. The simulation start which a random sequence and it is tested all conformations, nding its conformations according to the Monte Carlo rules. If the Zscore increases, we assume that the new conformation is more stable than the previous. This process is repeated until the optimally designed sequence (with the highest Zscore) is reached. Keeping track of all the Monte Carlo trajectories, a map of conectivity of sequence-structures is obtained. The results shows trajectories connected with structures of low Zscore values. The increase of Zscore along of the simulation leads to a small group of preferred conformations. The model suggest a funnel like structure for folding evolution, in which the structures at the bottom of the funnel are associated with the motif of a protein. This result can be a possible iv explanation for the restricted number of conformations compared to the large number of sequences... (Complete abstract click electronic access below)