Made available in DSpace on 2016-06-02T20:06:11Z (GMT). No. of bitstreams: 1 DissMP.pdf: 748309 bytes, checksum: b6a638a5f9ec09f6622480b42f13d699 (MD5) Previous issue date: 2006-02-22 / Financiadora de Estudos e Projetos / This work has as main objective to insert covariates in the capture probability of the multiple capture-recapture method for closed animal population. Factors like climate, seasons of the year, animal size, could a¤ect the animal capture probability. We revise the methodology concepts, we make a study about the posteriori parameters sensibility, we present new parameters for the capture probability in specific situations and we insert covariates in the model used by Castledine (1981) through bayesian methods. The bayesian analysis was made through several studies of stochastic simulation through MCMC (Monte Carlo Markov Chain) with simulated and real data to obtain the population size posteriori results. / Este trabalho tem como objetivo principal a inserção de covariáveis nas probabilidades de captura do método de captura-recaptura múltipla para população fechada. No caso de população animal, por exemplo, fatores como clima, época do ano, tamanho do animal, podem afetar a probabilidade de captura do animal. Revisamos os conceitos da metodologia, fazemos um breve estudo sobre a sensibilidade das estimativas a posteriori em relação a escolha dos hiperparâmetros, apresentamos uma reparametrização para a probabilidade de captura em situações específicas e, motivados nessa reparametrização, inserimos covariáveis no modelo proposto por Castledine (1981) por meio de métodos bayesianos. A análise bayesiana foi feita através de vários estudos de simulação estocástica via MCMC (Monte Carlo Markov Chain) com dados simulados e reais para obter os resultados a posteriori do tamanho populacional.

Habenaria é um dos maiores gêneros da família Orchidaceae, e estimativas atuais pressupoem a existência de aproximadamente 835 espécies. Habenaria seção Pentadactylae com 34 espécies é a maior entre as 14 seções do gênero existente no novo mundo e compreende um conjunto de espécies morfologicamente bastante heterogênea. A fim de investigar a monofilia da seção e sua relação com outras seções do gênero, foram executadas análise Bayesiana e de Máxima Parcimônia com o emprego de um marcador nuclear (ITS) e três marcadores plastidiais (matK, intron trnK, rps16-trnk). Os resultados demonstraram que a seção Pentadactylae é altamente polifilética. Baseado nas análises filogenéticas e reavaliação de caracteres morfológicos, a seção Pentadactylae foi recircunscrita neste trabalho e sete espécies são aceitas: H. dutraei, H. ekmaniana, H. exaltata, H. henscheniana, H. megapotamensis, H. montevidensis e H. pentadactyla, enquanto outras 32 espécies foram excluídas. Habenaria crassipes é reconhecida como um sinônimo de H. exaltata. Lectótipos são designados para H. crassipes e H. recta. Todas as espécies da seção habitam pântanos ou locais bastante úmidos; com área de distribuição passando pelo norte da Argentina, Uruguai, Paraguai, sul, sudeste e centro do Brasil. O estado do Rio Grande do Sul (sul do Brasil), possivelmente, constitui um centro de diversidade da seção onde todas as espécies podem ser encontradas. A biologia reprodutiva de duas espécies da seção Pentadactylae, H. megapotamensis e H. montevidensis; e duas espécies da seção Macroceratitae, H. johannensis e H. macronectar, foram estudas. Todas as espécies estudadas oferecem néctar como recompensa floral aos polinizadores, produzido no interior de um prolongamento do labelo denominado esporão. Habenaria montevidensis é polinizada por borboletas da família Hesperiidae, enquanto as demais espécies são polinizadas por mariposas da família Sphingidae. Todas as espécies estudadas são auto-compatíveis mas dependentes de agentes polinizadores para a produção de frutos. O sucesso reprodutivo é alto (69,48 - 93%). Na área de estudo, todas as quatro espécies estudadas são reprodutivamente isoladas devido a um conjunto de fatores tais como diferenças na morfologia floral e diferentes polinizadores. / Habenaria is one of the largest genus of Orchidaceae family and current stimates accounts to the existence of 835 species. Habenaria section Pentadactylae with 34 species is the largest among the 14 New World sections of the genus and comprises a morphologically heterogeneous group of species. To investigate the monophyly of the section and the relation with other sections of the genus, Bayesian and parsimony analyses using one nuclear marker (ITS) and three plastid markers (matK, trnK intron, rps16-trnK) were performed. The results demonstrated that sect. Pentadactylae is highly polyphyletic. Based on the phylogenetic analyses and re-evaluation of morphological characters, Habenaria sect. Pentadactylae is re-circumscribed and seven species are accepted for the section: H. dutraei, H. ekmaniana, H. exaltata, H. henscheniana, H. megapotamensis, H. montevidensis and H. pentadactyla, while other 32 species were excluded. Habenaria crassipes is included under the synonym of H. exaltata. Lectotypes are designated for H. crassipes and H. recta. All species in the section are from marshes or wet grasslands and range from Northern Argentina, Uruguai, Paraguai and south, southeast and center of Brazil. The Rio Grande do Sul state (south Brazil), possibly constitute a diversity center of the section where every species can be founded. Most are rare, known by few populations, and threatened due to loss of habitat and population decline. The reproductive biology of two species from the section Pentadactylae, H. megapotamensis and H. montevidensis; and two species from the section Macroceratitae, H. johannensis and H. macronectar, were studied. All studied species offer nectar as floral reward concealed in a labellar process termed spur. Habenaria montevidensis is pollinated by Hesperiidae butterflies, while the remaining species are pollinated by Sphingidae moths. All studied species are self-compatible, but pollinator-dependent. The reproductive success is high (69.48 - 93%). At the study site, every four studied species are reproductively isolated by a set of factors that includes differing floral morphologies and different pollinators.

The Beta distribution is a bounded domain distribution which has dominated the modeling the distribution of random variable that assume value between 0 and 1. Bounded domain distributions arising in various situations such as rates, proportions and index. Motivated by an analysis of electoral votes percentages (where a distribution with support on the positive real numbers was used, although a distribution with limited support could be more suitable) we focus on alternative distributions to Beta distribution with emphasis in regression models. In this work, initially we present the Simplex mixture model as a flexible model to modeling the distribution of bounded random variable then we extend the model to the context of regression models with the inclusion of covariates. The parameters estimation is discussed for both models considering Bayesian inference. We apply these models to simulated data sets in order to investigate the performance of the estimators. The results obtained were satisfactory for all the cases investigated. Finally, we introduce a parameterization of the L-Logistic distribution to be used in the context of regression models and we extend it to a mixture of mixed models. / A distribuição beta é uma distribuição com suporte limitado que tem dominado a modelagem de variáveis aleatórias que assumem valores entre 0 e 1. Distribuições com suporte limitado surgem em várias situações como em taxas, proporções e índices. Motivados por uma análise de porcentagens de votos eleitorais, em que foi assumida uma distribuição com suporte nos números reais positivos quando uma distribuição com suporte limitado seira mais apropriada, focamos em modelos alternativos a distribuição beta com enfase em modelos de regressão. Neste trabalho, apresentamos, inicialmente, um modelo de mistura de distribuições Simplex como um modelo flexível para modelar a distribuição de variáveis aleatórias que assumem valores em um intervalo limitado, em seguida estendemos o modelo para o contexto de modelos de regressão com a inclusão de covariáveis. A estimação dos parâmetros foi discutida para ambos os modelos, considerando o método bayesiano. Aplicamos os dois modelos a dados simulados para investigarmos a performance dos estimadores usados. Os resultados obtidos foram satisfatórios para todos os casos investigados. Finalmente, introduzimos a distribuição L-Logistica no contexto de modelos de regressão e posteriormente estendemos este modelo para o contexto de misturas de modelos de regressão mista.

Habenaria é um dos maiores gêneros da família Orchidaceae, e estimativas atuais pressupoem a existência de aproximadamente 835 espécies. Habenaria seção Pentadactylae com 34 espécies é a maior entre as 14 seções do gênero existente no novo mundo e compreende um conjunto de espécies morfologicamente bastante heterogênea. A fim de investigar a monofilia da seção e sua relação com outras seções do gênero, foram executadas análise Bayesiana e de Máxima Parcimônia com o emprego de um marcador nuclear (ITS) e três marcadores plastidiais (matK, intron trnK, rps16-trnk). Os resultados demonstraram que a seção Pentadactylae é altamente polifilética. Baseado nas análises filogenéticas e reavaliação de caracteres morfológicos, a seção Pentadactylae foi recircunscrita neste trabalho e sete espécies são aceitas: H. dutraei, H. ekmaniana, H. exaltata, H. henscheniana, H. megapotamensis, H. montevidensis e H. pentadactyla, enquanto outras 32 espécies foram excluídas. Habenaria crassipes é reconhecida como um sinônimo de H. exaltata. Lectótipos são designados para H. crassipes e H. recta. Todas as espécies da seção habitam pântanos ou locais bastante úmidos; com área de distribuição passando pelo norte da Argentina, Uruguai, Paraguai, sul, sudeste e centro do Brasil. O estado do Rio Grande do Sul (sul do Brasil), possivelmente, constitui um centro de diversidade da seção onde todas as espécies podem ser encontradas. A biologia reprodutiva de duas espécies da seção Pentadactylae, H. megapotamensis e H. montevidensis; e duas espécies da seção Macroceratitae, H. johannensis e H. macronectar, foram estudas. Todas as espécies estudadas oferecem néctar como recompensa floral aos polinizadores, produzido no interior de um prolongamento do labelo denominado esporão. Habenaria montevidensis é polinizada por borboletas da família Hesperiidae, enquanto as demais espécies são polinizadas por mariposas da família Sphingidae. Todas as espécies estudadas são auto-compatíveis mas dependentes de agentes polinizadores para a produção de frutos. O sucesso reprodutivo é alto (69,48 - 93%). Na área de estudo, todas as quatro espécies estudadas são reprodutivamente isoladas devido a um conjunto de fatores tais como diferenças na morfologia floral e diferentes polinizadores. / Habenaria is one of the largest genus of Orchidaceae family and current stimates accounts to the existence of 835 species. Habenaria section Pentadactylae with 34 species is the largest among the 14 New World sections of the genus and comprises a morphologically heterogeneous group of species. To investigate the monophyly of the section and the relation with other sections of the genus, Bayesian and parsimony analyses using one nuclear marker (ITS) and three plastid markers (matK, trnK intron, rps16-trnK) were performed. The results demonstrated that sect. Pentadactylae is highly polyphyletic. Based on the phylogenetic analyses and re-evaluation of morphological characters, Habenaria sect. Pentadactylae is re-circumscribed and seven species are accepted for the section: H. dutraei, H. ekmaniana, H. exaltata, H. henscheniana, H. megapotamensis, H. montevidensis and H. pentadactyla, while other 32 species were excluded. Habenaria crassipes is included under the synonym of H. exaltata. Lectotypes are designated for H. crassipes and H. recta. All species in the section are from marshes or wet grasslands and range from Northern Argentina, Uruguai, Paraguai and south, southeast and center of Brazil. The Rio Grande do Sul state (south Brazil), possibly constitute a diversity center of the section where every species can be founded. Most are rare, known by few populations, and threatened due to loss of habitat and population decline. The reproductive biology of two species from the section Pentadactylae, H. megapotamensis and H. montevidensis; and two species from the section Macroceratitae, H. johannensis and H. macronectar, were studied. All studied species offer nectar as floral reward concealed in a labellar process termed spur. Habenaria montevidensis is pollinated by Hesperiidae butterflies, while the remaining species are pollinated by Sphingidae moths. All studied species are self-compatible, but pollinator-dependent. The reproductive success is high (69.48 - 93%). At the study site, every four studied species are reproductively isolated by a set of factors that includes differing floral morphologies and different pollinators.

Submitted by (ana.araujo@ufrpe.br) on 2016-08-04T13:26:23Z No. of bitstreams: 1 Magali Teresopolis Reis Amaral.pdf: 5438608 bytes, checksum: a3ca949533ae94adaf7883fd465a627a (MD5) / Made available in DSpace on 2016-08-04T13:26:23Z (GMT). No. of bitstreams: 1 Magali Teresopolis Reis Amaral.pdf: 5438608 bytes, checksum: a3ca949533ae94adaf7883fd465a627a (MD5) Previous issue date: 2008-08-18 / The adjustment of the weight-age growth curves for animals plays an important role in animal production planning. These adjusted growth curves must be coherent with the biological interpretation of animal growth, which often demands imposition of constraints on model parameters.The inference of the parameters of nonlinear models with constraints, using classical techniques, presents various difficulties. In order to bypass those difficulties, a bayesian approach for adjustment of the growing curves is proposed. In this respect the bayesian proposed approach introduces restrictions on model parameters through choice of the prior density. Due to the nonlinearity, the posterior density of those parameters does not have a kernel that can be identified among the traditional distributions, and their moments can only be obtained using numerical techniques. In this work the MCMC simulation (Monte Carlo chain Markov) was implemented to obtain a summary of the posterior density. Besides, selection model criteria were used for the observed data, based on generated samples of the posterior density.The main purpose of this work is to show that the bayesian approach can be of practical use, and to compare the bayesian inference of the estimated parameters considering noninformative prior density (from Jeffreys), with the classical inference obtained by the Gauss-Newton method. Therefore it was possible to observe that the calculation of the confidence intervals based on the asymptotic theory fails, indicating non significance of certain parameters of some models, while in the bayesian approach the intervals of credibility do not present this problem. The programs in this work were implemented in R language,and to illustrate the utility of the proposed method, analysis of real data was performed, from an experiment of evaluation of system of crossing among cows from different herds, implemented by Embrapa Pecuária Sudeste. The data correspond to 12 measurements of weight of animals between 8 and 19 months old, from the genetic groups of the races Nelore and Canchim, belonging to the genotype AALLAB (Paz 2002). The results reveal excellent applicability of the bayesian method, where the model of Richard presented difficulties of convergence both in the classical and in the bayesian approach (with non informative prior). On the other hand the logistic model provided the best adjustment of the data for both methodologies when opting for non informative and informative prior density. / O ajuste de curva de crescimento peso-idade para animais tem um papel importante no planejamento da produção animal. No entanto, as curvas de crescimento ajustadas devem ser coerentes com as interpretações biológicas do crescimento do animal, o que exige muitas vezes que sejam impostas restrições aos parâmetros desse modelo.A inferência de parâmetros de modelos não lineares sujeito a restrições, utilizando técnicas clássicas apresenta diversas dificuldades. Para contornar estas dificuldades, foi proposta uma abordagem bayesiana para ajuste de curvas de crescimento. Neste sentido,a abordagem bayesiana proposta introduz as restrições nos parâmetros dos modelos através das densidades de probabilidade a priori adotadas. Devido à não linearidade, as densidades a posteriori destes parâmetros não têm um núcleo que possa ser identificado entre as distribuições tradicionalmente conhecidas e os seus momentos só podem ser obtidos numericamente. Neste trabalho, as técnicas de simulação de Monte Carlo Cadeia de Markov (MCMC) foram implementadas para obtenção de um sumário das densidades a posteriori. Além disso, foram utilizados critérios de seleção do melhor modelo para um determinado conjunto de dados baseados nas amostras geradas das densidades a posteriori.O objetivo principal deste trabalho é mostrar a viabilidade da abordagem bayesiana e comparar a inferência bayesiana dos parâmetros estimados, considerando-se densidades a priori não informativas (de Jeffreys), com a inferência clássica das estimativas obtidas pelo método de Gauss-Newton. Assim, observou-se que o cálculo de intervalos de confiança, baseado na teoria assintótica, falha, levando a não significância de certos parâmetros de alguns modelos. Enquanto na abordagem bayesiana os intervalos de credibilidade não apresentam este problema. Os programas utilizados foram implementados no R e para ilustração da aplicabilidade do método proposto, foram realizadas análises de dados reais oriundos de um experimento de avaliação de sistema de cruzamento entre raças bovinas de corte, executado na Embrapa Pecuária Sudeste. Os dados correspondem a 12 mensurações de peso dos 8 aos 19 meses de idade do grupo genético das raças Nelore e Canchim, pertencente ao grupo de genotípico AALLAB, ver (Paz 2002). Os resultados revelaram excelente aplicabilidade do método bayesiano, destacando que o modelo de Richard apresentou dificuldades de convergência tanto na abordagem clássica como bayesiana (com priori não informativa). Por outro lado o modelo Logístico foi quem melhor se ajustou aos dados em ambas metodologias quando se optou por densidades a priori não informativa e informativa.

La croissance du marché des pompes à chaleur contribue à l'augmentation de la thermosensibilité de la demande électrique. Il devient nécessaire de mieux comprendre l'impact des usages thermosensibles de l'électricité, notamment concernant ceux qui sont corrélés de manière non linéaire à la température extérieure. Dans cette optique, cette thèse vise à construire un cadre de modélisation qui permette i) d'analyser les facteurs d'influence de la thermosensibilité à partir d'une description physique des usages thermosensibles, et ii) de réaliser des diagnostics de ces paramètres d'influence tout en tenant compte des incertitudes associées. Une approche de modélisation hybride qui bénéficie des avantages de modèles statistiques et de modèles physiques est principalement employée pour répondre à ces questions.La première étape consiste à estimer la part thermosensible de la demande réelle par un modèle prédictif top-down. On développe ensuite un modèle d'analyse physique de la thermosensibilité à l'échelle régionale à partir de la thermique du bâtiment. On s'appuie notamment sur des modèles pseudo-physiques de performance de pompes à chaleur qui sont régressés sur des données constructeur ou des mesures de performances réelles. Un COP régional est déterminé pour l'ensemble des PAC installées. Enfin, les paramètres d'influence du modèle de thermosensibilité ainsi développé sont estimés à l'aide de l'approche bayésienne, qui offre un cadre pour le traitement de l'incertitude sous la forme de probabilités. Des coefficients équivalents de déperditions thermiques, une température intérieure équivalente ainsi que les parts du chauffage Joule et par PAC pour le parc de bâtiments régional ont été obtenus. / The growing heat pump market contributes to the increase in temperature sensitivity of electricity demand. It becomes necessary to understand the impact of temperature sensitive end-uses of electricity, including those which are correlated non-linearly to the outside temperature. In this context, this thesis aims to build a modeling framework to i) analyze the influencing factors of the temperature sensitivity of electricity demand from a physical description of temperature-sensitive equipment, and ii) to perform diagnoses of these parameters of influence by taking into account the associated uncertainties. A hybrid modeling approach that benefits the advantages of statistical models and physical models is used to answer these questions.Firstly, the temperature-sensitive part of electricity demand is estimated by a predictive top-down model. Then a physical model to analyze the temperature sensitivity at regional level is developed based on building thermal energy needs. A regional coefficient of performance (COP) is determined for the whole installed heat pumps by using pseudo-physical models which are regressed on manufacturer data or actual performance measures. Finally, the parameters of influence of the developed temperature sensitivity model are estimated using the Bayesian approach which provides a framework for the treatment of uncertainty in the form of probabilities. Equivalent coefficients of heat loss, an equivalent internal temperature, as well as the share of Joule heating and the share of heat pumps for the regional building stock are obtained.

Des approches parcimonieuses sont aujourd'hui développées pour la modélisation de la qualité des rejets urbains par temps de pluie, e adéquation avec la quantité de données disponibles. De plus, l'analyse des incertitudes apparaît comme un outil incontournable pour le test des modèles. Parallèlement, le développement des techniques de mesure en continu en réseau, spectrométrie et turbidité, permet l'obtention de données continues de flux de matières en suspension et de demande chimique en oxygène en grand nombre, apportant une information riche. Ce travail constitue une des premières études en hydrologie urbaine basée sur l'exploitation d'une grande base de données acquises par la mesure de la turbidité. Des mesures sur la période 2004-2008 ont été exploitées sur deux sites. Après traitement et validation, 263 et 239 événements pluvieux ont été retenus. L'analyse des données a permis la formulation d'hypothèses sur la génération des flux pour la proposition de modèles adaptés. Le test de l'approche multi-régression a confirmé la nécessité de construire des modèles locaux, basés sur une analyse approfondie des données. Les meilleurs modèles obtenus sont ceux pour la masse événementielle qui parviennent à reproduire en tendance la variabilité des observations. La méthode bayésienne a été utilisée pour le test d'un modèle d'Accumulation-Erosion-Transfert simple à l'échelle du bassin versant. Les premiers résultats mettent e défaut la structure du modèle testé. Cependant ces premiers tests ont démontré l'efficacité de la procédure d'analyse bayésienne, dont l'application du principe d'apprentissage permet d'améliorer de manière significative les structures des modèles. / More and more urban water managers are tackling the issue of water quality modelling. Current research works focus on parsimonious modelling approaches that match the amount of data available for calibration. Moreover uncertainties analysis now appears as an integrated step and a powerful tool in models testing. In parallel, development of in sewer continuous measurements based on spectrometry and turbidimetry techniques, provides large data base of continuous total suspended solids and chemical oxygen demand concentrations, providing much information on fluxes dynamics. This research work is one of the first studies in urban hydrology based on a large turbidity data base. Data from two sites have been treated and validated, with measurements over the period 2004-2008. 263 and 239 stormwater events were selected for the modelling work for the two sites. Data analysis provided insights for making assumptions on the pollutant fluxes generation and proposing adapted models. Test of multi-regression approach that it is necessary to build local approaches based on detailed data analysis. Best models were obtained for event mass, data variability could be reproduced in trend. Formal Bayesian approach was used for testing a simple global Accumulation-Erosion-Transfer model for Chassieu. First results evidenced the difficulties of the model to reproduce the dynamics variability. This may be due to the simple structure. However these first tests have demonstrated the efficiency of the Bayesian analysis procedure. ln particular, the application of the learning principle showed that model structure can be significantly and efficiently improved.

This thesis explores methodologies of modeling and evaluating the operational performance of a typical aeroengine having field experience over two decades. Upon failure, the engine is repaired and restored to flight worthy condition and hence comes under the purview of repairable systems. Operational performance of the engine is being measured in terms of five functions of time, namely, M(t), which is the expected number of system failures in the time interval [0,t]; system failure rate m(t), which is an unconditional quantity and is simply the derivative of M(t); ρ(t), the conditional failure intensity given the history of a system Ht, which is nothing but limdt→1 Prob(System fails in [t,t + dt] |Ht); and M′(t) and m′(t), which are 0 dt conditional entities analogous to M(t) and m(t) defined in the same spirit as that of ρ(t), the details of which are given in the third chapter of the thesis. These functions are being estimated using field failure-repair data of 418 aeroengines, where the observations on time between failures are being measured in number of flying hours logged in between failures, and the corresponding repair duration is being measured in number of calendar days. To start with, using the superimposed renewal process model the above quantities M(t), m(t), m′(t), M′(t) and ρ(t) are estimated both in the frequentist as well as the Bayesian framework. Subsequently repair times have been incorporated into the model and analysed using both frequentist and Bayesian approaches. Next, the model of Lawless and Thiagarajah (1996) which incorporates both renewal and time trend, has been generalized to include repair time as well, and a comprehensive methodology of Bayesian model selection under this model has been developed. After introducing the research problem in the first chapter, the engineering system description leading to the identification of the failure modes, repair practice and the variables of interest is taken up in the following chapter at the outset, as a pre-requisite to the stochastic modeling and the statistical analysis that to follow in the remainder of the thesis. As the first stochastic model, the number of system failures in a given time interval is modeled as a superimposed renewal process with the constituent independent renewal processes running in different component sockets having Weibull inter failure times. This model is first empirically validated using the field failure data and then using this model, the five quantities of interest as mentioned above viz. M(t), m(t), ρ(t), M′(t) and m′(t) are analysed from a frequentist maximum likelihood perspective. A Bayesian analysis of the same follows in the subsequent chapter. Next, the repair effect is incorporated into the superimposed renewal process model by considering the Weibull parameters of inter failure times of the constituent renewal processes running in independent component sockets as a polynomial in the last repair time. The nature of this polynomial relationships are empirically deter-mined and the Weibull assumption is validated through a test of hypothesis. Different polynomial relationships lead to consideration of several models, with the correct ones chosen through a series of likelihood ratio tests. Next based on the appropriate models a maximum likelihood analysis of M(t), ρ(t) and M′(t) has been carried out. Like the simple superimposed renewal process model, Bayesian analysis of this model incorporating repair times is carried out in the following chapter. In the Bayesian setup however, the problem of model selection could be kept unrestricted to non-nested models as well (unlike the previous chapter, where only nested models could be considered), and a comprehensive model selection exercise has been carried out with the aid of intrinsic Bayes factors and training data sets. The last but one chapter presents a generalised model of Lawless and Thiagarajah (1996) for performance evaluation of aeroengines that incorporate renewals, time trends and the repair characteristics. Here also since the primary problem is one of model selection, the entire analysis like in the preceding chapter has been carried out under the Bayesian frame-work. The final chapter concludes the thesis by comparing the empirical results obtained in the previous five chapters, summarising the main contributions of the thesis and providing directions for future research.

En apprentissage supervisé, obtenir les réels labels pour un ensemble de données peut être très fastidieux et long. Aujourd'hui, les récentes avancées d'Internet ont permis le développement de services d'annotations en ligne, faisant appel au crowdsourcing pour collecter facilement des labels. Néanmoins, le principal inconvénient de ces services réside dans le fait que les annotateurs peuvent avoir des niveaux d'expertise très hétérogènes. De telles données ne sont alors pas forcément fiables. Par conséquent, la gestion de l'incertitude des annotateurs est un élément clé pour l'apprentissage à partir de multiples annotateurs non experts. Dans cette thèse, nous proposons des algorithmes probabilistes qui traitent l'incertitude des annotateurs et la qualité des données durant la phase d'apprentissage. Trois modèles sont proposés: IGNORE permet de classer de nouvelles instances tout en évaluant les annotateurs en terme de performance d'annotation qui dépend de leur incertitude. X-IGNORE intègre la qualité des données en plus de l'incertitude des juges. En effet, X-IGNORE suppose que la performance des annotateurs dépend non seulement de leur incertitude mais aussi de la qualité des données qu'ils annotent. Enfin, ExpertS répond au problème de sélection d'annotateurs durant l'apprentissage. ExpertS élimine les annotateurs les moins performants, et se base ainsi uniquement sur les labels des bons annotateurs (experts) lors de l'étape d'apprentissage. De nombreuses expérimentations effectuées sur des données synthétiques et réelles montrent la performance et la stabilité de nos modèles par rapport à différents algorithmes de la littérature. / In supervised learning tasks, obtaining the ground truth label for each instance of the training dataset can be difficult, time-consuming and/or expensive. With the advent of infrastructures such as the Internet, an increasing number of web services propose crowdsourcing as a way to collect a large enough set of labels from internet users. The use of these services provides an exceptional facility to collect labels from anonymous annotators, and thus, it considerably simplifies the process of building labels datasets. Nonetheless, the main drawback of crowdsourcing services is their lack of control over the annotators and their inability to verify and control the accuracy of the labels and the level of expertise for each labeler. Hence, managing the annotators' uncertainty is a clue for learning from imperfect annotations. This thesis provides three algorithms when learning from multiple uncertain annotators. IGNORE generates a classifier that predict the label of a new instance and evaluate the performance of each annotator according to their level of uncertainty. X-Ignore, considers that the performance of the annotators both depends on their uncertainty and on the quality of the initial dataset to be annotated. Finally, ExpertS deals with the problem of annotators' selection when generating the classifier. It identifies experts annotators, and learn the classifier based only on their labels. We conducted in this thesis a large set of experiments in order to evaluate our models, both using experimental and real world medical data. The results prove the performance and accuracy of our models compared to previous state of the art solutions in this context.

Grâce aux combinaison d’antirétroviraux (ART), le taux de transmission mère-enfant (TME) du Virus de l’Immunodéficience Humaine VIH est maintenant réduit à moins de 1%. Des progrès restent néanmoins à faire, en particuliers chez les femmes débutant tardivement leurs consultations prénatales. Mais l’évaluation de nouvelles stratégies ART pour la prévention de la TME (PTME) devient de plus en plus complexe à cause des exigences statistiques liées au fait que les transmissions sont rares. Dés lors, il devient crucial de modéliser l’efficacité des ART sur la charge virale (CV) et la transmission. Dans cette thèse, nous modélisons les mécanismes et le moment de la TME du VIH en utilisant des modèles mixtes. Les données proviennent de 4 études de prévention périnatales du VIH en Thaïlande (PHPT-1, PHPT-2, PHPT-5 1st phase et PHPT-5 2nd phase), une base de données historiques de qualité, collectées de 1996 à 2015, où différentes prophylaxies ART ont été prescrites à plus de 4000 femmes enceintes pendant des durées variables: Zidovudine (ZDV) seule; ZDV + une dose unique de nevirapine périnatale (sdNVP); ZDV+lopinavir/ritonavir (LPV/r) ou ZDV+LPV/r+lamivudine (3TC). Les modèles développés apportent un éclairage sur les mécanismes et le moment de la PTME ainsi que sur le rôle respectif de différents ART sur la réduction de la CV et sur leur contribution à la prophylaxie pre/post exposition. Partant de ces connaissances, une approche bayésienne est appliquée à une étude de cas (essai de supériorité dans le contexte d’un événement rare comme la TME du VIH) afin d’améliorer la puissance statistique tout en limitant la taille de l’échantillon, et comparée à une approche fréquentiste. / Nowadays, with the use of highly efficacious antiretroviral (ART) combination, the rate of mother-to-child transmission (MTCT) of Human Immunodeficiency Virus (HIV) is reduced to less than 1%. However, there are still some gaps to be filled, especially in women who initiate antenatal prophylaxis late. The evaluation of new ART strategies or drug combinations for the prevention of MTCT (PMTCT) of HIV becomes more difficult in view of the statistical requirements in case of rare outcomes. Therefore modeling the impact of ART on maternal viral load (VL) and transmission rate is increasingly important. In this thesis, the mechanisms and timing of MTCT of HIV were investigated using mixed model approaches. Data were derived from four perinatal HIV prevention studies in Thailand (PHPT-1, PHPT-2, PHPT-5 1st phase and PHPT-5 2nd phase), a unique set of quality historical data, consistently collected from 1996 to 2015, in which different ART prophylaxis regimens were provided to more than 4000 HIV-infected pregnant women for varying durations: Zidovudine (ZDV) alone; ZDV plus perinatal single dose nevirapine; ZDV+lopinavir/ritonavir (LPV/r) or ZDV+LPV/r+lamivudine. The developed models provide insights on the mechanisms and timing of PMTCT as well as on the respective role of different ARTs on the maternal VL reduction and in turn on their contribution to pre/post-exposure prophylaxis in MTCT of HIV. With this prior knowledge, a Bayesian design is applied to a case study, a trial aimed at evaluating treatment superiority in the context of rare outcomes such as MTCT of HIV, to improve the statistical power –while limiting sample size— and compared to a frequentist design.