Analyse du comportement hétérogène des usagers dans un réseau

Le nombre important de véhicules sur le réseau routier peut entraîner des problèmes d'encombrement et de sécurité. Les usagers des réseaux routiers qui nous intéressent sont les camionneurs qui transportent des marchandises, pouvant rouler avec des véhicules non conformes ou emprunter des routes interdites pour gagner du temps. Le transport de matières dangereuses est réglementé et certains lieux, surtout les ponts et les tunnels, leur sont interdits d'accès. Pour aider à faire appliquer les lois en vigueur, il existe un système de contrôles routiers composé de structures fixes et de patrouilles mobiles. Le déploiement stratégique de ces ressources de contrôle mise sur la connaissance du comportement des camionneurs que nous allons étudier à travers l'analyse de leurs choix de routes. Un problème de choix de routes peut se modéliser en utilisant la théorie des choix discrets, elle-même fondée sur la théorie de l'utilité aléatoire. Traiter ce type de problème avec cette théorie est complexe. Les modèles que nous utiliserons sont tels, que nous serons amenés à faire face à des problèmes de corrélation, puisque plusieurs routes partagent probablement des arcs. De plus, puisque nous travaillons sur le réseau routier du Québec, le choix de routes peut se faire parmi un ensemble de routes dont le nombre est potentiellement infini si on considère celles ayant des boucles. Enfin, l'étude des choix faits par un humain n'est pas triviale. Avec l'aide du modèle de choix de routes retenu, nous pourrons calculer une expression de la probabilité qu'une route soit prise par le camionneur. Nous avons abordé cette étude du comportement en commençant par un travail de description des données collectées. Le questionnaire utilisé par les contrôleurs permet de collecter des données concernant les camionneurs, leurs véhicules et le lieu du contrôle. La description des données observées est une étape essentielle, car elle permet de présenter clairement à un analyste potentiel ce qui est accessible pour étudier les comportements des camionneurs. Les données observées lors d'un contrôle constitueront ce que nous appellerons une observation. Avec les attributs du réseau, il sera possible de modéliser le réseau routier du Québec. Une sélection de certains attributs permettra de spécifier la fonction d'utilité et par conséquent la fonction permettant de calculer les probabilités de choix de routes par un camionneur. Il devient alors possible d'étudier un comportement en se basant sur des observations. Celles provenant du terrain ne nous donnent pas suffisamment d'information actuellement et même en spécifiant bien un modèle, l'estimation des paramètres n'est pas possible. Cette dernière est basée sur la méthode du maximum de vraisemblance. Nous avons l'outil, mais il nous manque la matière première que sont les observations, pour continuer l'étude. L'idée est de poursuivre avec des observations de synthèse. Nous ferons des estimations avec des observations complètes puis, pour se rapprocher des conditions réelles, nous continuerons avec des observations partielles. Ceci constitue d'ailleurs un défi majeur. Nous proposons pour ces dernières, de nous servir des résultats des travaux de (Bierlaire et Frejinger, 2008) en les combinant avec ceux de (Fosgerau, Frejinger et Karlström, 2013). Bien qu'elles soient de nature synthétiques, les observations que nous utilisons nous mèneront à des résultats tels, que nous serons en mesure de fournir une proposition concrète qui pourrait aider à optimiser les décisions des responsables des contrôles routiers. En effet, nous avons réussi à estimer, sur le réseau réel du Québec, avec un seuil de signification de 0,05 les valeurs des paramètres d'un modèle de choix de routes discrets, même lorsque les observations sont partielles. Ces résultats donneront lieu à des recommandations sur les changements à faire dans le questionnaire permettant de collecter des données.

Advances in scaling deep learning algorithms

Les algorithmes d'apprentissage profond forment un nouvel ensemble de méthodes puissantes pour l'apprentissage automatique. L'idée est de combiner des couches de facteurs latents en hierarchies. Cela requiert souvent un coût computationel plus elevé et augmente aussi le nombre de paramètres du modèle. Ainsi, l'utilisation de ces méthodes sur des problèmes à plus grande échelle demande de réduire leur coût et aussi d'améliorer leur régularisation et leur optimization. Cette thèse adresse cette question sur ces trois perspectives. Nous étudions tout d'abord le problème de réduire le coût de certains algorithmes profonds. Nous proposons deux méthodes pour entrainer des machines de Boltzmann restreintes et des auto-encodeurs débruitants sur des distributions sparses à haute dimension. Ceci est important pour l'application de ces algorithmes pour le traitement de langues naturelles. Ces deux méthodes (Dauphin et al., 2011; Dauphin and Bengio, 2013) utilisent l'échantillonage par importance pour échantilloner l'objectif de ces modèles. Nous observons que cela réduit significativement le temps d'entrainement. L'accéleration atteint 2 ordres de magnitude sur plusieurs bancs d'essai. Deuxièmement, nous introduisont un puissant régularisateur pour les méthodes profondes. Les résultats expérimentaux démontrent qu'un bon régularisateur est crucial pour obtenir de bonnes performances avec des gros réseaux (Hinton et al., 2012). Dans Rifai et al. (2011), nous proposons un nouveau régularisateur qui combine l'apprentissage non-supervisé et la propagation de tangente (Simard et al., 1992). Cette méthode exploite des principes géometriques et permit au moment de la publication d'atteindre des résultats à l'état de l'art. Finalement, nous considérons le problème d'optimiser des surfaces non-convexes à haute dimensionalité comme celle des réseaux de neurones. Tradionellement, l'abondance de minimum locaux était considéré comme la principale difficulté dans ces problèmes. Dans Dauphin et al. (2014a) nous argumentons à partir de résultats en statistique physique, de la théorie des matrices aléatoires, de la théorie des réseaux de neurones et à partir de résultats expérimentaux qu'une difficulté plus profonde provient de la prolifération de points-selle. Dans ce papier nous proposons aussi une nouvelle méthode pour l'optimisation non-convexe.

Evidence of extended orientational order in amorphous Fe/Sm thin films

Amorphous thin films of Fe/Sm, prepared by evaporation methods, have been magnetically characterized and the results were interpreted in terms of the random magnets theory. The samples behave as 2D and 3D random magnets depending on the total thickness of the film. From our data the existence of orientational order, which greatly influences the magnetic behavior of the films, is also clear.

Modelling climate change: the role of unresolved processes

Our understanding of the climate system has been revolutionized recently, by the development of sophisticated computer models. The predictions of such models are used to formulate international protocols, intended to mitigate the severity of global warming and its impacts. Yet, these models are not perfect representations of reality, because they remove from explicit consideration many physical processes which are known to be key aspects of the climate system, but which are too small or fast to be modelled. The purpose of this paper is to give a personal perspective of the current state of knowledge regarding the problem of unresolved scales in climate models. A recent novel solution to the problem is discussed, in which it is proposed, somewhat counter-intuitively, that the performance of models may be improved by adding random noise to represent the unresolved processes.

A closer look at chaotic advection in the stratosphere: part II: statistical diagnostics

Statistical diagnostics of mixing and transport are computed for a numerical model of forced shallow-water flow on the sphere and a middle-atmosphere general circulation model. In particular, particle dispersion statistics, transport fluxes, Liapunov exponents (probability density functions and ensemble averages), and tracer concentration statistics are considered. It is shown that the behavior of the diagnostics is in accord with that of kinematic chaotic advection models so long as stochasticity is sufficiently weak. Comparisons with random-strain theory are made.

Ensemble downscaling in coupled solar wind-magnetosphere modeling for space weather forecasting

Advanced forecasting of space weather requires simulation of the whole Sun-to-Earth system, which necessitates driving magnetospheric models with the outputs from solar wind models. This presents a fundamental difficulty, as the magnetosphere is sensitive to both large-scale solar wind structures, which can be captured by solar wind models, and small-scale solar wind “noise,” which is far below typical solar wind model resolution and results primarily from stochastic processes. Following similar approaches in terrestrial climate modeling, we propose statistical “downscaling” of solar wind model results prior to their use as input to a magnetospheric model. As magnetospheric response can be highly nonlinear, this is preferable to downscaling the results of magnetospheric modeling. To demonstrate the benefit of this approach, we first approximate solar wind model output by smoothing solar wind observations with an 8 h filter, then add small-scale structure back in through the addition of random noise with the observed spectral characteristics. Here we use a very simple parameterization of noise based upon the observed probability distribution functions of solar wind parameters, but more sophisticated methods will be developed in the future. An ensemble of results from the simple downscaling scheme are tested using a model-independent method and shown to add value to the magnetospheric forecast, both improving the best estimate and quantifying the uncertainty. We suggest a number of features desirable in an operational solar wind downscaling scheme.

Preference reversals for ambiguity aversion

This paper finds preference reversals in measurements of ambiguity aversion, even if psychological and informational circumstances are kept constant. The reversals are of a fundamentally different nature than the reversals found before because they cannot be explained by context-dependent weightings of attributes. We offer an explanation based on Sugden's random-reference theory, with different elicitation methods generating different random reference points. Then measurements of ambiguity aversion that use willingness to pay are confounded by loss aversion and hence overestimate ambiguity aversion.

We're Chained: an analysis of systemic risk in finance

This dissertation presents two papers on how to deal with simple systemic risk measures to assess portfolio risk characteristics. The first paper deals with the Granger-causation of systemic risk indicators based in correlation matrices in stock returns. Special focus is devoted to the Eigenvalue Entropy as some previous literature indicated strong re- sults, but not considering different macroeconomic scenarios; the Index Cohesion Force and the Absorption Ratio are also considered. Considering the S&P500, there is not ev- idence of Granger-causation from Eigenvalue Entropies and the Index Cohesion Force. The Absorption Ratio Granger-caused both the S&P500 and the VIX index, being the only simple measure that passed this test. The second paper develops this measure to capture the regimes underlying the American stock market. New indicators are built using filtering and random matrix theory. The returns of the S&P500 is modelled as a mixture of normal distributions. The activation of each normal distribution is governed by a Markov chain with the transition probabilities being a function of the indicators. The model shows that using a Herfindahl-Hirschman Index of the normalized eigenval- ues exhibits best fit to the returns from 1998-2013.

Interpretação sísmica e análise de atributos - área do levantamento sísmico 3D de siririzinho (Bacia Sergipe-Alagoas, NE do Brasil)

The main objective of the present thesis was the seismic interpretation and seismic attribute analysis of the 3D seismic data from the Siririzinho high, located in the Sergipe Sub-basin (southern portion of Sergipe-Alagoas Basin). This study has enabled a better understanding of the stratigraphy and structure that the Siririzinho high experienced during its development. In a first analysis, we used two types of filters: the dip-steered median filter, was used to remove random noise and increase the lateral continuity of reflections, and fault-enhancement filter was applied to enhance the reflection discontinuities. After this filtering step similarity and curvature attributes were applied in order to identify and enhance the distribution of faults and fractures. The use of attributes and filtering greatly contributed to the identification and enhancement of continuity of faults. Besides the application of typical attributes (similarity and curvature) neural network and fingerprint techniques were also used, which generate meta-attributes, also aiming to highlight the faults; however, the results were not satisfactory. In a subsequent step, well log and seismic data analysis were performed, which allowed the understanding of the distribution and arrangement of sequences that occur in the Siririzinho high, as well as an understanding of how these units are affected by main structures in the region. The Siririzinho high comprises an elongated structure elongated in the NS direction, capped by four seismo-sequences (informally named, from bottom to top, the sequences I to IV, plus the top of the basement). It was possible to recognize the main NS-oriented faults, which especially affect the sequences I and II, and faults oriented NE-SW, that reach the younger sequences, III and IV. Finally, with the interpretation of seismic horizons corresponding to each of these sequences, it was possible to define a better understanding of geometry, deposition and structural relations in the area.

THERMAL RADIATION FROM A FLUCTUATING EVENT HORIZON

Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq)

The spectrum of D-slash in QCD via replicas

Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP)

Fluctuating commutative geometry

We use the framework of noncommutative geometry to define a discrete model for fluctuating geometry. Instead of considering ordinary geometry and its metric fluctuations, we consider generalized geometries where topology and dimension can also fluctuate. The model describes the geometry of spaces with a countable number n of points. The spectral principle of Connes and Chamseddine is used to define dynamics. We show that this simple model has two phases. The expectation value , the average number of points in the universe, is finite in one phase and diverges in the other. Moreover, the dimension delta is a dynamical observable in our model, and plays the role of an order parameter. The computation of is discussed and an upper bound is found, < 2. We also address another discrete model defined on a fixed d = 1 dimension, where topology fluctuates. We comment on a possible spontaneous localization of topology.

Universalidade em matrizes aleatórias via problemas de Riemann-Hilbert

Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP)

Análise de ambiguidade em métodos de campo potenciais através de análise fatorial, Q-Modal

Este trabalho apresenta uma metodologia para o estudo da ambiguidade na interpretação de dados geofísicos. Várias soluções alternativas, representativas da região de maior ambiguidade no espaço de parâmetros são obtidas, sendo posteriormente grupadas e ordenadas pela análise fatorial modo Q. Esta metodologia foi aplicada a dados sintéticos de campo potencial simulando-se causas de ambiguidade como discretização e truncamento da anomalia e a presença de ruídos aleatório e geológico. Um único prisma foi usado como modelo interpretativo, sendo a espessura a principal causa de ambiguidade tanto na gravimetria como na magnetometria. Segue-se a profundidade do topo sempre associada à espessura, quando o sinal da anomalia é alto. Quando a anomalia tem sinal baixo, a largura torna-se o segundo parâmetro mais importante, também associada à espessura. Ao contrário da presença de interferências geológicas, a presença de ruído aleatório nos campos, não é fator importante na ambiguidade. A aplicação da metodologia a dados reais ilustra o papel desta análise na caracterização de soluções alternativas e a importância da informação a priori na caracterização das causas de ambiguidade. A metodologia apresentada pode ser empregada em diversos estágios de um programa de prospecção fornecendo em cada estágio uma análise dos principais fatores causadores da ambiguidade, que poderá ser util no planejamento dos estágios seguintes. Comparada a outros métodos de análise de ambiguidade, como por exemplo regiões de confiança, a metodologia estudada destaca-se por não precisar satisfazer premissas estatísticas sobre a distribuição dos erros.

Análise comparativa de três métodos de inversão aplicados a dados gravimétricos e magnéticos em perfil

Tradicionalmente, o método dos mínimos quadrados tem sido empregado na inversão não linear de dados de campo potencial. No caso em que as observações dos campos gravimétrico ou magnético contém apenas ruído Gaussiano. O método dos mínimos quadrados não apresenta problemas. Entretanto, quando as observações são perturbadas por ruído não Gaussiano, ou mesmo por ruído não aleatório, como é o caso de muitos ruídos geológicos, o método dos mínimos quadrados torna-se bastante ineficiente, e métodos alternativos devem ser empregados a fim de produzir interpretações realísticas. Neste trabalho, uma comparação é feita entre os métodos dos mínimos quadrados, dos mínimos absolutos e do ajuste-M, aplicados à inversão não linear de dados de campo potencial. A comparação é efetuada usando-se dados teóricos, onde diversas situações geológicas são simuladas. Os resultados mostram que na presença de ruído geológico, caracterizado por pequeno corpo raso acima do corpo principal, ou por corpo grande, adjacente ao corpo principal, o ajuste-M apresenta desempenho muito superior ao dos mínimos quadrados e dos mínimos absolutos. Na presença de ruído Gaussiano, entretanto, o ajuste-M tem um desempenho inferior aos outros dois métodos. Como o ruído Gaussiano é um ruído branco, parte dele pode ser removido por um filtro passa baixa adequado, sem muita perda do sinal, o que não ocorre com o ruído geológico que contém componentes importantes de baixo número de onda. Desse modo o ajuste-M se torna uma ferramenta importante na interpretação de áreas geologicamente complexas, onde é comum a contaminação das anomalias por ruído geológico. Os três métodos em estudo são aplicados a uma anomalia magnética real causada por uma intrusão de diabásio em forma de dique, em sedimentos arenosos da formação Piauí na Bacia do Parnaíba. Os três métodos apresentaram resultados semelhantes indicando que tanto o nível de ruído Gaussiano como geológico são baixos nesta anomalia.