Compte tenu de la vulnérabilité de la population rurale de la région sahélienne aux aléas pluviométriques, et devant les ambitions de certains acteurs d’utiliser le levier de l’usage des terres pour contribuer à l’atténuation du changement climatique, il est important de comprendre les facteurs contribuant à la variabilité de la couverture végétale au Sahel.Une synthèse de la littérature expliquant l’évolution récente de la végétation au Sahel est donc d’abord présentée. Les études s’intéressant au paradigme qui souligne l’impact de l’usage des terres sur les précipitations en Afrique de l’Ouest évaluent principalement ces effets par le couplage de modèles dynamiques globaux de végétation – DGVM – avec des modèles de circulation générale. C’est à l’amélioration d’un tel DGVM, ORCHIDEE, développé à l’Institut Pierre Simon Laplace, que le reste du travail cherche à contribuer.Comme d’autres études ont montré qu’il était possible d’utiliser en première approximation les steppes pâturées et les jachères pour décrire le comportement global de la surface sahélienne, les écarts entre modèle et mesures sont caractérisés pour une jachère située à proximité de Wankama (Niger). Plus précisément, les forces et faiblesses de la paramétrisation et de la structure par défaut du modèle sont diagnostiqués, et l’importance de la réduction d’erreur permise par l’optimisation de certains des paramètres est donnée. En particulier, l’emploi d’une résolution aux différences finies de la diffusion de l’eau dans la colonne de sol est évalué, dans la mesure où cela permet de mieux simuler la réponse rapide du flux évaporatoire aux événements pluvieux que le schéma conceptuel utilisé par défaut dans ORCHIDEE.Le réalisme du modèle est également mesuré à l’échelle régionale, par la comparaison d’observations de NDVI GIMMS_3G à la couverture végétale simulée par le modèle en réponse à différents forçages climatiques . Si les modifications introduites au cours du travail ne permettent pas de mieux décrire les tendances de la végétation au cours des dernières décennies, tirer partie des leçons du présent travail pourra se révéler utile. Il en est de même des conclusions de l’étude de la transitivité des biais conditionnels du modèle réalisée avec Tao Wang et présentée en annexe B. / The evolution of the land-surface conditions is often assessed through the use of “dynamic global vegetation models”, as is shown in a review of the current understanding of the factors of variability and of the recent evolution of the vegetation cover in the Sahel. Such models are also coupled to atmospheric general circulation models to evaluate the land feedback on precipitation in monsoonal climates.Thus, the improvement of the skills of such surface models to simulate the radiative and turbulent fluxes between the land of surface and the atmosphere in the Sahel over a range of scales from hourly to multi-annual has a potential to have significant implications. This is especially true considering the vulnerability of the rural population of the region, which largely relies on rainfed agriculture and the interest on the evolution of the carbon stocks of ecosystems in the context of climate change. Such a work on the ORCHIDEE model is presented here. In complement to croplands, rangelands and fallows represent a large share of the sahelian landscapes and have intermediate characteristics between erosion glacis and acacia bushes. As such, their evolution (in terms of albedo, roughness length,…) may be used to study the Sahel ecosystem behaviour as a first approximation. Differences between model outputs and field observations are quantified for a fallow close to Wankama (Niger). More precisely, some of the drawbacks of the standard parametrisation and structure of the model are diagnosed, and the range of reduction of the model-observation mismatch that results from optimizing some of the parameters are given (plant phenology,…). In particular, the use of a finite difference resolution of the soil water diffusion is considered as it enables to better simulate the fast response of evaporative fluxes to rainfall than the conceptual scheme routinely used in ORCHIDEE. The benefits of the use of such a “physical” hydrological scheme on the different outputs of the surface scheme is evaluated.The realism of the model is also measured at the regional scale, through a comparison with GIMMS_3G NDVI time series over West Africa. If the modifications that have been introduced in the model don’t improve its ability to describe the vegetation cover trends over the last decades in the region, several lessons can be kept from the analysis that has been realised, especially from the work on the transitivity of state-dependant model biases that has been conducted with Tao Wang and which is presented in annex B.

Relevance is an essential concept in information retrieval (IR) and relevance estimation is a fundamental IR task. It involves not only document relevance estimation, but also estimation of user's information need. Relevance-based language model aims to estimate a relevance model (i.e., a relevant query term distribution) from relevance feedback documents. The true relevance model should be generated from truly relevant documents. The ideal estimation of the true relevance model is expected to be not only effective in terms of mean retrieval performance (e.g., Mean Average Precision) over all the queries, but also stable in the sense that the performance is stable across different individual queries. In practice, however, in approximating/estimating the true relevance model, the improvement of retrieval effectiveness often sacrifices the retrieval stability, and vice versa. In this thesis, we propose to explore and analyze such effectiveness-stability tradeoff from a new perspective, i.e., the bias-variance tradeoff that is a fundamental theory in statistical estimation. We first formulate the bias, variance and the trade-off between them for retrieval performance as well as for query model estimation. We then analytically and empirically study a number of factors (e.g., query model complexity, query model combination, document weight smoothness and irrelevant documents removal) that can affect the bias and variance. Our study shows that the proposed bias-variance trade-off analysis can serve as an analytical framework for query model estimation. We then investigate in depth on two particular key factors: document weight smoothness and removal of irrelevant documents, in query model estimation, by proposing novel methods for document weight smoothing and irrelevance distribution separation, respectively. Systematic experimental evaluation on TREC collections shows that the proposed methods can improve both retrieval effectiveness and retrieval stability of query model estimation. In addition to the above main contributions, we also carry out initial exploration on two further directions: the formulation of bias-variance in personalization and looking at the query model estimation via a novel theoretical angle (i.e., Quantum theory) that has partially inspired our research.

Airbus réalise pour chaque avion et pour chaque équipement de nombreux essais, au sol ou en vol et doit garantir qu'en tout point de vol possible, la température de chacun des équipements reste inférieure à la température limite correspondante. Pour pouvoir valider la température de chaque équipement dans l'enveloppe de vol, il faudrait disposer d'essais réalisés aux frontières. Or, tous les essais en vol sont confrontés aux contraintes climatiques et opérationnelles qui ne permettent pas d'explorer tout le domaine. C'est pourquoi Airbus a besoin d'élaborer des méthodes d'extrapolation de température, de manière à prédire le comportement thermique des matériaux et des équipements dans les pires conditions. Les techniques proposées sont basées sur la théorie de l'identification de systèmes qui consiste à déterminer des modèles de comportement d'un point de vue heuristique à partir de mesures et considérations physiques. Plus précisément, le présent document valide les modèles ARX comme un outil pour l'identification de la température du système. Les modèles et les techniques sont étudiés, tout d'abord, d'un point de vue de la simulation numérique et après, confrontés face à des tests représentatifs au laboratoire. Les techniques proposées permettent prédire la température des composants avion pour des conditions différentes / Airbus must ensure that for all flight conditions that a given aircraft could face, the temperature of each powerplant system must be less than the corresponding critical temperature. In order to validate the temperature of each device in the flight envelope, tests at the border should be done. Airbus produces for each aircraft component many trials, either in flight or ground. However, all flight tests are faced with climatic and operational constraints which do not permit exploring the whole area. That's why Airbus needs to develop methods of extrapolation of temperature in order to predict the thermal behavior of materials and equipments in the worst conditions. The proposed techniques are based on the system identification theory which consists on heuristically determining an analytical model using physical insights and measurements. More precisely, this paper validates ARX models as a tool for the identification of the system's temperature. The models and techniques are studied, first, from a numerical simulation point of view and second, based on laboratory representative tests. The proposed techniques allow predicting the temperature of aircraft components at different conditions

Conventional reliability analysis methods assume that a simulation model is able to represent the real physics accurately. However, this assumption may not always hold as the simulation model could be biased due to simplifications and idealizations. Simulation models are approximate mathematical representations of real-world systems and thus cannot exactly imitate the real-world systems. The accuracy of a simulation model is especially critical when it is used for the reliability calculation. Therefore, a simulation model should be validated using prototype testing results for reliability analysis. However, in practical engineering situation, experimental output data for the purpose of model validation is limited due to the significant cost of a large number of physical testing. Thus, the model validation needs to be carried out to account for the uncertainty induced by insufficient experimental output data as well as the inherent variability existing in the physical system and hence in the experimental test results. Therefore, in this study, a confidence-based model validation method that captures the variability and the uncertainty, and that corrects model bias at a user-specified target confidence level, has been developed. Reliability assessment using the confidence-based model validation can provide conservative estimation of the reliability of a system with confidence when only insufficient experimental output data are available. Without confidence-based model validation, the designed product obtained using the conventional reliability-based design optimization (RBDO) optimum could either not satisfy the target reliability or be overly conservative. Therefore, simulation model validation is necessary to obtain a reliable optimum product using the RBDO process. In this study, the developed confidence-based model validation is integrated in the RBDO process to provide truly confident RBDO optimum design. The developed confidence-based model validation will provide a conservative RBDO optimum design at the target confidence level. However, it is challenging to obtain steady convergence in the RBDO process with confidence-based model validation because the feasible domain changes as the design moves (i.e., a moving-target problem). To resolve this issue, a practical optimization procedure, which terminates the RBDO process once the target reliability is satisfied, is proposed. In addition, the efficiency is achieved by carrying out deterministic design optimization (DDO) and RBDO without model validation, followed by RBDO with the confidence-based model validation. Numerical examples are presented to demonstrate that the proposed RBDO approach obtains a conservative and practical optimum design that satisfies the target reliability of designed product given a limited number of experimental output data. Thus far, while the simulation model might be biased, it is assumed that we have correct distribution models for input variables and parameters. However, in real practical applications, only limited numbers of test data are available (parameter uncertainty) for modeling input distributions of material properties, manufacturing tolerances, operational loads, etc. Also, as before, only a limited number of output test data is used. Therefore, a reliability needs to be estimated by considering parameter uncertainty as well as biased simulation model. Computational methods and a process are developed to obtain confidence-based reliability assessment. The insufficient input and output test data induce uncertainties in input distribution models and output distributions, respectively. These uncertainties, which arise from lack of knowledge – the insufficient test data, are different from the inherent input distributions and corresponding output variabilities, which are natural randomness of the physical system.