L’imagerie hyperspectrale est très largement employée en télédétection pour diverses applications, dans le domaine civil comme dans le domaine militaire. Une image hyperspectrale est le résultat de l’acquisition d’une seule scène observée dans plusieurs longueurs d’ondes. Par conséquent, chacun des pixels constituant cette image est représenté par un vecteur de mesures (généralement des réflectances) appelé spectre. Une étape majeure dans l’analyse des données hyperspectrales consiste à identifier les composants macroscopiques (signatures) présents dans la région observée et leurs proportions correspondantes (abondances). Les dernières techniques développées pour ces analyses ne modélisent pas correctement ces images. En effet, habituellement ces techniques supposent l’existence de pixels purs dans l’image, c’est-à-dire des pixels constitué d’un seul matériau pur. Or, un pixel est rarement constitué d’éléments purs distincts l’un de l’autre. Ainsi, les estimations basées sur ces modèles peuvent tout à fait s’avérer bien loin de la réalité. Le but de cette étude est de proposer de nouveaux algorithmes d’estimation à l’aide d’un modèle plus adapté aux propriétés intrinsèques des images hyperspectrales. Les paramètres inconnus du modèle sont ainsi déduits dans un cadre Bayésien. L’utilisation de méthodes de Monte Carlo par Chaînes de Markov (MCMC) permet de surmonter les difficultés liées aux calculs complexes de ces méthodes d’estimation. / Hyperspectral imagery has been widely used in remote sensing for various civilian and military applications. A hyperspectral image is acquired when a same scene is observed at different wavelengths. Consequently, each pixel of such image is represented as a vector of measurements (reflectances) called spectrum. One major step in the analysis of hyperspectral data consists of identifying the macroscopic components (signatures) that are present in the sensored scene and the corresponding proportions (concentrations). The latest techniques developed for this analysis do not properly model these images. Indeed, these techniques usually assume the existence of pure pixels in the image, i.e. pixels containing a single pure material. However, a pixel is rarely composed of pure spectrally elements, distinct from each other. Thus, such models could lead to weak estimation performance. The aim of this thesis is to propose new estimation algorithms with the help of a model that is better suited to the intrinsic properties of hyperspectral images. The unknown model parameters are then infered within a Bayesian framework. The use of Markov Chain Monte Carlo (MCMC) methods allows one to overcome the difficulties related to the computational complexity of these inference methods.

This thesis focuses on estimating parameters of appropriate model for daily returns using the Markov Chain Monte Carlo method (MCMC) and Bayesian statistics. We describe MCMC methods, such as Gibbs sampling and Metropolis- Hastings algorithm and their basic properties. After that, we introduce different financial models. Particularly we focus on the lognormal autoregressive model. Later we theoretically apply Gibbs sampling to lognormal autoregressive model using principles of Bayesian statistics. Afterwards, we analyze procedu- res, that we used in simulations of posterior distribution using Gibbs sampling. Finally, we present processed output of both simulated and real data analysis.

Neste trabalho é introduzido um novo modelo estatístico para modelar dados limitados no intervalo continuo (0;1). O modelo proposto é construído sob uma transformação de variáveis, onde a variável transformada é resultado da combinação de uma variável com distribuição normal padrão e a função de distribuição acumulada da distribuição valor extremo generalizado. Para o novo modelo são estudadas suas propriedades estruturais. A nova família é estendida para modelos de regressão, onde o modelo é reparametrizado na mediana da variável resposta e este conjuntamente com o parâmetro de dispersão são relacionados com covariáveis através de uma função de ligação. Procedimentos inferênciais são desenvolvidos desde uma perspectiva clássica e bayesiana. A inferência clássica baseia-se na teoria de máxima verossimilhança e a inferência bayesiana no método de Monte Carlo via cadeias de Markov. Além disso estudos de simulação foram realizados para avaliar o desempenho das estimativas clássicas e bayesianas dos parâmetros do modelo. Finalmente um conjunto de dados de câncer colorretal é considerado para mostrar a aplicabilidade do modelo. / In this research a new statistical model is introduced to model data restricted in the continuous interval (0;1). The proposed model is constructed under a transformation of variables, in which the transformed variable is the result of the combination of a variable with standard normal distribution and the cumulative distribution function of the generalized extreme value distribution. For the new model its structural properties are studied. The new family is extended to regression models, in which the model is reparametrized in the median of the response variable and together with the dispersion parameter are related to covariables through a link function. Inferential procedures are developed from a classical and Bayesian perspective. The classical inference is based on the theory of maximum likelihood, and the Bayesian inference is based on the Markov chain Monte Carlo method. In addition, simulation studies were performed to evaluate the performance of the classical and Bayesian estimates of the model parameters. Finally a set of colorectal cancer data is considered to show the applicability of the model

La tomographie des temps de première arrivée vise à retrouver un modèle de vitesse de propagation des ondes sismiques à partir des temps de première arrivée mesurés. Cette technique nécessite la résolution d’un problème inverse afin d’obtenir un modèle sismique cohérent avec les données observées. Il s'agit d'un problème mal posé pour lequel il n'y a aucune garantie quant à l'unicité de la solution. L’approche bayésienne permet d’estimer la distribution spatiale de la vitesse de propagation des ondes sismiques. Il en résulte une meilleure quantification des incertitudes associées. Cependant l’approche reste relativement coûteuse en temps de calcul, les algorithmes de Monte Carlo par chaînes de Markov (MCMC) classiquement utilisés pour échantillonner la loi a posteriori des paramètres n'étant efficaces que pour un nombre raisonnable de paramètres. Elle demande, de ce fait, une réflexion à la fois sur la paramétrisation du modèle de vitesse afin de réduire la dimension du problème et sur la définition de la loi a priori des paramètres. Le sujet de cette thèse porte essentiellement sur cette problématique.Le premier modèle que nous considérons est basé sur un modèle de mosaïque aléatoire, le modèle de Jonhson-Mehl, dérivé des mosaïques de Voronoï déjà proposées en tomographie bayésienne. Contrairement à la mosaïque de Voronoï, les cellules de Johsnon-mehl ne sont pas forcément convexes et sont bornées par des portions d’hyperboloïdes, offrant ainsi des frontières lisses entre les cellules. Le deuxième modèle est, quant à lui, décrit par une combinaison linéaire de fonctions gaussiennes, centrées sur la réalisation d'un processus ponctuel de Poisson. Pour chaque modèle, nous présentons un exemple de validation sur des champs de vitesse simulés. Nous appliquons ensuite notre méthodologie à un modèle synthétique plus complexe qui sert de benchmark dans l'industrie pétrolière. Nous proposons enfin, un modèle de vitesse basé sur la théorie du compressive sensing pour reconstruire le champ de vitesse. Ce modèle, encore imparfait, ouvre plusieurs pistes de recherches futures.Dans ce travail, nous nous intéressons également à un jeu de données réelles acquises dans le contexte de la fracturation hydraulique. Nous développons dans ce contexte une méthode d'inférence bayésienne trans-dimensionnelle et hiérarchique afin de traiter efficacement la complexité du modèle à couches. / First arrival time tomography aims at inferring the seismic wave propagation velocity using experimental first arrival times. In our study, we rely on a Bayesian approach to estimate the wave velocity and the associated uncertainties. This approach incorporates the information provided by the data and the prior knowledge of the velocity model. Bayesian tomography allows for a better estimation of wave velocity as well asassociated uncertainties. However, this approach remains fairly expensive, and MCMC algorithms that are used to sample the posterior distribution are efficient only as long as the number of parameters remains within reason. Hence, their use requires a careful reflection both on the parameterization of the velocity model, in order to reduce the problem's dimension, and on the definition of the prior distribution of the parameters. In this thesis, we introduce new parsimonious parameterizations enabling to accurately reproduce the wave velocity field with the associated uncertainties.The first parametric model that we propose uses a random Johnson-Mehl tessellation, a variation of the Voronoï tessellation. The second one uses Gaussian kernels as basis functions. It is especially adapted to the detection of seismic wave velocity anomalies. Each anomaly isconsidered to be a linear combination of these basis functions localized at the realization of a Poisson point process. We first illustrate the tomography results with a synthetic velocity model, which contains two small anomalies. We then apply our methodology to a more advanced and more realistic synthetic model that serves as a benchmark in the oil industry. The tomography results reveal the ability of our algorithm to map the velocity heterogeneitieswith precision using few parameters. Finally, we propose a new parametric model based on the compressed sensing techniques. The first results are encouraging. However, the model still has some weakness related to the uncertainties estimation.In addition, we analyse real data in the context of induced microseismicity. In this context, we develop a trans-dimensional and hierarchical approach in order to deal with the full complexity of the layered model.

Les algorithmes d'inférence ou d'optimisation possèdent généralement des hyperparamètres qu'il est nécessaire d'ajuster. Nous nous intéressons ici à l'automatisation de cette étape d'ajustement et considérons différentes méthodes qui y parviennent en apprenant en ligne la structure du problème considéré.La première moitié de cette thèse explore l'ajustement des hyperparamètres en apprentissage artificiel. Après avoir présenté et amélioré le cadre générique de l'optimisation séquentielle à base de modèles (SMBO), nous montrons que SMBO s'applique avec succès à l'ajustement des hyperparamètres de réseaux de neurones profonds. Nous proposons ensuite un algorithme collaboratif d'ajustement qui mime la mémoire qu'ont les humains d'expériences passées avec le même algorithme sur d'autres données.La seconde moitié de cette thèse porte sur les algorithmes MCMC adaptatifs, des algorithmes d'échantillonnage qui explorent des distributions de probabilité souvent complexes en ajustant leurs paramètres internes en ligne. Pour motiver leur étude, nous décrivons d'abord l'observatoire Pierre Auger, une expérience de physique des particules dédiée à l'étude des rayons cosmiques. Nous proposons une première partie du modèle génératif d'Auger et introduisons une procédure d'inférence des paramètres individuels de chaque événement d'Auger qui ne requiert que ce premier modèle. Ensuite, nous remarquons que ce modèle est sujet à un problème connu sous le nom de label switching. Après avoir présenté les solutions existantes, nous proposons AMOR, le premier algorithme MCMC adaptatif doté d'un réétiquetage en ligne qui résout le label switching. Nous présentons une étude empirique et des résultats théoriques de consistance d'AMOR, qui mettent en lumière des liens entre le réétiquetage et la quantification vectorielle / Inference and optimization algorithms usually have hyperparameters that require to be tuned in order to achieve efficiency. We consider here different approaches to efficiently automatize the hyperparameter tuning step by learning online the structure of the addressed problem. The first half of this thesis is devoted to hyperparameter tuning in machine learning. After presenting and improving the generic sequential model-based optimization (SMBO) framework, we show that SMBO successfully applies to the task of tuning the numerous hyperparameters of deep belief networks. We then propose an algorithm that performs tuning across datasets, mimicking the memory that humans have of past experiments with the same algorithm on different datasets. The second half of this thesis deals with adaptive Markov chain Monte Carlo (MCMC) algorithms, sampling-based algorithms that explore complex probability distributions while self-tuning their internal parameters on the fly. We start by describing the Pierre Auger observatory, a large-scale particle physics experiment dedicated to the observation of atmospheric showers triggered by cosmic rays. The models involved in the analysis of Auger data motivated our study of adaptive MCMC. We derive the first part of the Auger generative model and introduce a procedure to perform inference on shower parameters that requires only this bottom part. Our model inherently suffers from label switching, a common difficulty in MCMC inference, which makes marginal inference useless because of redundant modes of the target distribution. After reviewing existing solutions to label switching, we propose AMOR, the first adaptive MCMC algorithm with online relabeling. We give both an empirical and theoretical study of AMOR, unveiling interesting links between relabeling algorithms and vector quantization.

Nous présentons une nouvelle méthode pour étudier les relations entre la structure génétique des populations et l'environnement. Cette méthode repose sur des modèles hiérarchiques bayésiens qui utilisent conjointement des données génétiques multi-locus et des données spatiales, environnementales et/ou culturelles. Elle permet d'estimer la structure génétique des populations, d'évaluer ses liens avec des covariables non génétiques, et de projeter la structure génétique des populations en fonction de ces covariables. Dans un premier temps, nous avons appliqué notre approche à des données de génétique humaine pour évaluer le rôle de la géographie et des langages dans la structure génétique des populations amérindiennes. Dans un deuxième temps, nous avons étudié la structure génétique des populations pour 20 espèces de plantes alpines et nous avons projeté les modifications intra spécifiques qui pourront être causées par le réchauffement climatique. / We introduce a new method to study the relationships between population genetic structure and environment. This method is based on Bayesian hierarchical models which use both multi-loci genetic data, and spatial, environmental, and/or cultural data. Our method provides the inference of population genetic structure, the evaluation of the relationships between the structure and non-genetic covariates, and the prediction of population genetic structure based on these covariates. We present two applications of our Bayesian method. First, we used human genetic data to evaluate the role of geography and languages in shaping Native American population structure. Second, we studied the population genetic structure of 20 Alpine plant species and we forecasted intra-specific changes in response to global warming. STAR

L'objectif de ce travail est de développer de nouvelles méthodes statistiques de reconstruction d'image spatiale (3D) et spatio-temporelle (3D+t) en Tomographie par Émission de Positons (TEP). Le but est de proposer des méthodes efficaces, capables de reconstruire des images dans un contexte de faibles doses injectées tout en préservant la qualité de l'interprétation. Ainsi, nous avons abordé la reconstruction sous la forme d'un problème inverse spatial et spatio-temporel (à observations ponctuelles) dans un cadre bayésien non paramétrique. La modélisation bayésienne fournit un cadre pour la régularisation du problème inverse mal posé au travers de l'introduction d'une information dite a priori. De plus, elle caractérise les grandeurs à estimer par leur distribution a posteriori, ce qui rend accessible la distribution de l'incertitude associée à la reconstruction. L'approche non paramétrique quant à elle pourvoit la modélisation d'une grande robustesse et d'une grande flexibilité. Notre méthodologie consiste à considérer l'image comme une densité de probabilité dans (pour une reconstruction en k dimensions) et à chercher la solution parmi l'ensemble des densités de probabilité de . La grande dimensionalité des données à manipuler conduit à des estimateurs n'ayant pas de forme explicite. Cela implique l'utilisation de techniques d'approximation pour l'inférence. La plupart de ces techniques sont basées sur les méthodes de Monte-Carlo par chaînes de Markov (MCMC). Dans l'approche bayésienne non paramétrique, nous sommes confrontés à la difficulté majeure de générer aléatoirement des objets de dimension infinie sur un calculateur. Nous avons donc développé une nouvelle méthode d'échantillonnage qui allie à la fois bonnes capacités de mélange et possibilité d'être parallélisé afin de traiter de gros volumes de données. L'approche adoptée nous a permis d'obtenir des reconstructions spatiales 3D sans nécessiter de voxellisation de l'espace, et des reconstructions spatio-temporelles 4D sans discrétisation en amont ni dans l'espace ni dans le temps. De plus, on peut quantifier l'erreur associée à l'estimation statistique au travers des intervalles de crédibilité. / The aim of this work is to develop new statistical methods for spatial (3D) and space-time (3D+t) Positron Emission Tomography (PET) reconstruction. The objective is to propose efficient reconstruction methods in a context of low injected doses while maintaining the quality of the interpretation. We tackle the reconstruction problem as a spatial or a space-time inverse problem for point observations in a \Bayesian nonparametric framework. The Bayesian modeling allows to regularize the ill-posed inverse problem via the introduction of a prior information. Furthermore, by characterizing the unknowns with their posterior distributions, the Bayesian context allows to handle the uncertainty associated to the reconstruction process. Being nonparametric offers a framework for robustness and flexibility to perform the modeling. In the proposed methodology, we view the image to reconstruct as a probability density in(for reconstruction in k dimensions) and seek the solution in the space of whole probability densities in . However, due to the size of the data, posterior estimators are intractable and approximation techniques are needed for posterior inference. Most of these techniques are based on Markov Chain Monte-Carlo methods (MCMC). In the Bayesian nonparametric approach, a major difficulty raises in randomly sampling infinite dimensional objects in a computer. We have developed a new sampling method which combines both good mixing properties and the possibility to be implemented on a parallel computer in order to deal with large data sets. Thanks to the taken approach, we obtain 3D spatial reconstructions without any ad hoc space voxellization and 4D space-time reconstructions without any discretization, neither in space nor in time. Furthermore, one can quantify the error associated to the statistical estimation using the credibility intervals.

Neste trabalho desenvolvemos um estudo sobre modelos auto-regressivos com heterocedasticidade (ARCH) e modelos auto-regressivos com erros ARCH (AR-ARCH). Apresentamos os procedimentos para a estimação dos modelos e para a seleção da ordem dos mesmos. As estimativas dos parâmetros dos modelos são obtidas utilizando duas técnicas distintas: a inferência Clássica e a inferência Bayesiana. Na abordagem de Máxima Verossimilhança obtivemos intervalos de confiança usando a técnica Bootstrap e, na abordagem Bayesiana, adotamos uma distribuição a priori informativa e uma distribuição a priori não-informativa, considerando uma reparametrização dos modelos para mapear o espaço dos parâmetros no espaço real. Este procedimento nos permite adotar distribuição a priori normal para os parâmetros transformados. As distribuições a posteriori são obtidas através dos métodos de simulação de Monte Carlo em Cadeias de Markov (MCMC). A metodologia é exemplificada considerando séries simuladas e séries do mercado financeiro brasileiro / In this work we present a study of autoregressive conditional heteroskedasticity models (ARCH) and autoregressive models with autoregressive conditional heteroskedasticity errors (AR-ARCH). We also present procedures for the estimation and the selection of these models. The estimates of the parameters of those models are obtained using both Maximum Likelihood estimation and Bayesian estimation. In the Maximum Likelihood approach we get confidence intervals using Bootstrap resampling method and in the Bayesian approach we present informative prior and non-informative prior distributions, considering a reparametrization of those models in order to map the space of the parameters into real space. This procedure permits to choose prior normal distributions for the transformed parameters. The posterior distributions are obtained using Monte Carlo Markov Chain methods (MCMC). The methodology is exemplified considering simulated and Brazilian financial series

Neste trabalho apresentamos uma análise Bayesiana para confiabilidade de sistemas de redes usando métodos de simulação de Monte Carlo via Cadeias de Markov. Assumimos diferentes densidades a priori para as confiabilidades dos componentes individuais, com o objetivo de obtermos sumários de interesse. A metodologia é exemplificada condiderando um sistema de rede com sete componentes e um caso especial de sistema complexo composto por nove componentes. Consideramos ainda confiabilidade de redes tipo k-out--of-m com alguns exemplos numéricos / In this work we present a Bayesian approach for network reliability systems using Marov Chain Monte Carlo methods. We assume different prior densities for the individual component reliabilities th to get the posterior summaries of interest. The methodology is exemplified considering a network system with seven components and a special case of complex system with nine components. We also consider k-out-of-m system reliabiility with some numerical examples

Os modelos normais de regressão têm sido utilizados durante muitos anos para a análise de dados. Mesmo nos casos em que a normalidade não podia ser suposta, tentava-se algum tipo de transformação com o intuito de alcançar a normalidade procurada. No entanto, na prática, essas suposições sobre normalidade e linearidade nem sempre são satisfeitas. Como alternativas à técnica clássica, foram desenvolvidas novas classes de modelos de regressão. Nesse contexto, focamos a classe de modelos em que a distribuição assumida para a variável resposta pertence à classe de distribuições simétricas. O objetivo geral desse trabalho é a modelagem desta classe no contexto bayesiano, em particular a modelagem da classe de modelos não-lineares heterocedásticos simétricos. Vale ressaltar que esse trabalho tem ligação com duas linhas de pesquisa, a saber: a inferência estatística abordando aspectos da teoria assintótica e a inferência bayesiana considerando aspectos de modelagem e critérios de seleção de modelos baseados em métodos de simulação de Monte Carlo em Cadeia de Markov (MCMC). Uma primeira etapa consiste em apresentar a classe dos modelos não-lineares heterocedásticos simétricos bem como a inferência clássica dos parâmetros desses modelos. Posteriormente, propomos uma abordagem bayesiana para esses modelos, cujo objetivo é mostrar sua viabilidade e comparar a inferência bayesiana dos parâmetros estimados via métodos MCMC com a inferência clássica das estimativas obtidas por meio da ferramenta GAMLSS. Além disso, utilizamos o método bayesiano de análise de influência caso a caso baseado na divergência de Kullback-Leibler para detectar observações influentes nos dados. A implementação computacional foi desenvolvida no software R e para detalhes dos programas pode ser consultado aos autores do trabalho / The normal regression models have been used for many years for data analysis. Even in cases where normality could not be assumed, was trying to be some kind of transformation in order to achieve the normality sought. However, in practice, these assumptions about normality and linearity are not always satisfied. As alternatives to classical technique new classes of regression models were developed. In this context, we focus on the class of models in which the distribution assumed for the response variable belongs to the symmetric distributions class. The aim of this work is the modeling of this class in the bayesian context, in particular the modeling of the nonlinear models heteroscedastic symmetric class. Note that this work is connected with two research lines, the statistical inference addressing aspects of asymptotic theory and the bayesian inference considering aspects of modeling and criteria for models selection based on simulation methods Monte Carlo Markov Chain (MCMC). A first step is to present the nonlinear models heteroscedastic symmetric class as well as the classic inference of parameters of these models. Subsequently, we propose a bayesian approach to these models, whose objective is to show their feasibility and compare the estimated parameters bayesian inference by MCMC methods with the classical inference of the estimates obtained by GAMLSS tool. In addition, we use the bayesian method of influence analysis on a case based on the Kullback-Leibler divergence for detecting influential observations in the data. The computational implementation was developed in the software R and programs details can be found at the studys authors