Improving sampling, optimization and feature extraction in Boltzmann machines

L’apprentissage supervisé de réseaux hiérarchiques à grande échelle connaît présentement un succès fulgurant. Malgré cette effervescence, l’apprentissage non-supervisé représente toujours, selon plusieurs chercheurs, un élément clé de l’Intelligence Artificielle, où les agents doivent apprendre à partir d’un nombre potentiellement limité de données. Cette thèse s’inscrit dans cette pensée et aborde divers sujets de recherche liés au problème d’estimation de densité par l’entremise des machines de Boltzmann (BM), modèles graphiques probabilistes au coeur de l’apprentissage profond. Nos contributions touchent les domaines de l’échantillonnage, l’estimation de fonctions de partition, l’optimisation ainsi que l’apprentissage de représentations invariantes. Cette thèse débute par l’exposition d’un nouvel algorithme d'échantillonnage adaptatif, qui ajuste (de fa ̧con automatique) la température des chaînes de Markov sous simulation, afin de maintenir une vitesse de convergence élevée tout au long de l’apprentissage. Lorsqu’utilisé dans le contexte de l’apprentissage par maximum de vraisemblance stochastique (SML), notre algorithme engendre une robustesse accrue face à la sélection du taux d’apprentissage, ainsi qu’une meilleure vitesse de convergence. Nos résultats sont présent ́es dans le domaine des BMs, mais la méthode est générale et applicable à l’apprentissage de tout modèle probabiliste exploitant l’échantillonnage par chaînes de Markov. Tandis que le gradient du maximum de vraisemblance peut-être approximé par échantillonnage, l’évaluation de la log-vraisemblance nécessite un estimé de la fonction de partition. Contrairement aux approches traditionnelles qui considèrent un modèle donné comme une boîte noire, nous proposons plutôt d’exploiter la dynamique de l’apprentissage en estimant les changements successifs de log-partition encourus à chaque mise à jour des paramètres. Le problème d’estimation est reformulé comme un problème d’inférence similaire au filtre de Kalman, mais sur un graphe bi-dimensionnel, où les dimensions correspondent aux axes du temps et au paramètre de température. Sur le thème de l’optimisation, nous présentons également un algorithme permettant d’appliquer, de manière efficace, le gradient naturel à des machines de Boltzmann comportant des milliers d’unités. Jusqu’à présent, son adoption était limitée par son haut coût computationel ainsi que sa demande en mémoire. Notre algorithme, Metric-Free Natural Gradient (MFNG), permet d’éviter le calcul explicite de la matrice d’information de Fisher (et son inverse) en exploitant un solveur linéaire combiné à un produit matrice-vecteur efficace. L’algorithme est prometteur: en terme du nombre d’évaluations de fonctions, MFNG converge plus rapidement que SML. Son implémentation demeure malheureusement inefficace en temps de calcul. Ces travaux explorent également les mécanismes sous-jacents à l’apprentissage de représentations invariantes. À cette fin, nous utilisons la famille de machines de Boltzmann restreintes “spike & slab” (ssRBM), que nous modifions afin de pouvoir modéliser des distributions binaires et parcimonieuses. Les variables latentes binaires de la ssRBM peuvent être rendues invariantes à un sous-espace vectoriel, en associant à chacune d’elles, un vecteur de variables latentes continues (dénommées “slabs”). Ceci se traduit par une invariance accrue au niveau de la représentation et un meilleur taux de classification lorsque peu de données étiquetées sont disponibles. Nous terminons cette thèse sur un sujet ambitieux: l’apprentissage de représentations pouvant séparer les facteurs de variations présents dans le signal d’entrée. Nous proposons une solution à base de ssRBM bilinéaire (avec deux groupes de facteurs latents) et formulons le problème comme l’un de “pooling” dans des sous-espaces vectoriels complémentaires.

Domain/Multi-Domain Protection and Provisioning in Optical Networks

L’évolution récente des commutateurs de sélection de longueurs d’onde (WSS -Wavelength Selective Switch) favorise le développement du multiplexeur optique d’insertionextraction reconfigurable (ROADM - Reconfigurable Optical Add/Drop Multiplexers) à plusieurs degrés sans orientation ni coloration, considéré comme un équipement fort prometteur pour les réseaux maillés du futur relativement au multiplexage en longueur d’onde (WDM -Wavelength Division Multiplexing ). Cependant, leur propriété de commutation asymétrique complique la question de l’acheminement et de l’attribution des longueur d’ondes (RWA - Routing andWavelength Assignment). Or la plupart des algorithmes de RWA existants ne tiennent pas compte de cette propriété d’asymétrie. L’interruption des services causée par des défauts d’équipements sur les chemins optiques (résultat provenant de la résolution du problème RWA) a pour conséquence la perte d’une grande quantité de données. Les recherches deviennent ainsi incontournables afin d’assurer la survie fonctionnelle des réseaux optiques, à savoir, le maintien des services, en particulier en cas de pannes d’équipement. La plupart des publications antérieures portaient particulièrement sur l’utilisation d’un système de protection permettant de garantir le reroutage du trafic en cas d’un défaut d’un lien. Cependant, la conception de la protection contre le défaut d’un lien ne s’avère pas toujours suffisante en termes de survie des réseaux WDM à partir de nombreux cas des autres types de pannes devenant courant de nos jours, tels que les bris d’équipements, les pannes de deux ou trois liens, etc. En outre, il y a des défis considérables pour protéger les grands réseaux optiques multidomaines composés de réseaux associés à un domaine simple, interconnectés par des liens interdomaines, où les détails topologiques internes d’un domaine ne sont généralement pas partagés à l’extérieur. La présente thèse a pour objectif de proposer des modèles d’optimisation de grande taille et des solutions aux problèmes mentionnés ci-dessus. Ces modèles-ci permettent de générer des solutions optimales ou quasi-optimales avec des écarts d’optimalité mathématiquement prouvée. Pour ce faire, nous avons recours à la technique de génération de colonnes afin de résoudre les problèmes inhérents à la programmation linéaire de grande envergure. Concernant la question de l’approvisionnement dans les réseaux optiques, nous proposons un nouveau modèle de programmation linéaire en nombres entiers (ILP - Integer Linear Programming) au problème RWA afin de maximiser le nombre de requêtes acceptées (GoS - Grade of Service). Le modèle résultant constitue celui de l’optimisation d’un ILP de grande taille, ce qui permet d’obtenir la solution exacte des instances RWA assez grandes, en supposant que tous les noeuds soient asymétriques et accompagnés d’une matrice de connectivité de commutation donnée. Ensuite, nous modifions le modèle et proposons une solution au problème RWA afin de trouver la meilleure matrice de commutation pour un nombre donné de ports et de connexions de commutation, tout en satisfaisant/maximisant la qualité d’écoulement du trafic GoS. Relativement à la protection des réseaux d’un domaine simple, nous proposons des solutions favorisant la protection contre les pannes multiples. En effet, nous développons la protection d’un réseau d’un domaine simple contre des pannes multiples, en utilisant les p-cycles de protection avec un chemin indépendant des pannes (FIPP - Failure Independent Path Protecting) et de la protection avec un chemin dépendant des pannes (FDPP - Failure Dependent Path-Protecting). Nous proposons ensuite une nouvelle formulation en termes de modèles de flots pour les p-cycles FDPP soumis à des pannes multiples. Le nouveau modèle soulève un problème de taille, qui a un nombre exponentiel de contraintes en raison de certaines contraintes d’élimination de sous-tour. Par conséquent, afin de résoudre efficacement ce problème, on examine : (i) une décomposition hiérarchique du problème auxiliaire dans le modèle de décomposition, (ii) des heuristiques pour gérer efficacement le grand nombre de contraintes. À propos de la protection dans les réseaux multidomaines, nous proposons des systèmes de protection contre les pannes d’un lien. Tout d’abord, un modèle d’optimisation est proposé pour un système de protection centralisée, en supposant que la gestion du réseau soit au courant de tous les détails des topologies physiques des domaines. Nous proposons ensuite un modèle distribué de l’optimisation de la protection dans les réseaux optiques multidomaines, une formulation beaucoup plus réaliste car elle est basée sur l’hypothèse d’une gestion de réseau distribué. Ensuite, nous ajoutons une bande pasiv sante partagée afin de réduire le coût de la protection. Plus précisément, la bande passante de chaque lien intra-domaine est partagée entre les p-cycles FIPP et les p-cycles dans une première étude, puis entre les chemins pour lien/chemin de protection dans une deuxième étude. Enfin, nous recommandons des stratégies parallèles aux solutions de grands réseaux optiques multidomaines. Les résultats de l’étude permettent d’élaborer une conception efficace d’un système de protection pour un très large réseau multidomaine (45 domaines), le plus large examiné dans la littérature, avec un système à la fois centralisé et distribué.

Numerical Feasibility Study for Treated Wastewater Recharge as a Tool to Impede Saltwater Intrusion in the Coastal Aquifer of Gaza – Palestine

The ongoing depletion of the coastal aquifer in the Gaza strip due to groundwater overexploitation has led to the process of seawater intrusion, which is continually becoming a serious problem in Gaza, as the seawater has further invaded into many sections along the coastal shoreline. As a first step to get a hold on the problem, the artificial neural network (ANN)-model has been applied as a new approach and an attractive tool to study and predict groundwater levels without applying physically based hydrologic parameters, and also for the purpose to improve the understanding of complex groundwater systems and which is able to show the effects of hydrologic, meteorological and anthropogenic impacts on the groundwater conditions. Prediction of the future behaviour of the seawater intrusion process in the Gaza aquifer is thus of crucial importance to safeguard the already scarce groundwater resources in the region. In this study the coupled three-dimensional groundwater flow and density-dependent solute transport model SEAWAT, as implemented in Visual MODFLOW, is applied to the Gaza coastal aquifer system to simulate the location and the dynamics of the saltwater–freshwater interface in the aquifer in the time period 2000-2010. A very good agreement between simulated and observed TDS salinities with a correlation coefficient of 0.902 and 0.883 for both steady-state and transient calibration is obtained. After successful calibration of the solute transport model, simulation of future management scenarios for the Gaza aquifer have been carried out, in order to get a more comprehensive view of the effects of the artificial recharge planned in the Gaza strip for some time on forestall, or even to remedy, the presently existing adverse aquifer conditions, namely, low groundwater heads and high salinity by the end of the target simulation period, year 2040. To that avail, numerous management scenarios schemes are examined to maintain the ground water system and to control the salinity distributions within the target period 2011-2040. In the first, pessimistic scenario, it is assumed that pumping from the aquifer continues to increase in the near future to meet the rising water demand, and that there is not further recharge to the aquifer than what is provided by natural precipitation. The second, optimistic scenario assumes that treated surficial wastewater can be used as a source of additional artificial recharge to the aquifer which, in principle, should not only lead to an increased sustainable yield of the latter, but could, in the best of all cases, revert even some of the adverse present-day conditions in the aquifer, i.e., seawater intrusion. This scenario has been done with three different cases which differ by the locations and the extensions of the injection-fields for the treated wastewater. The results obtained with the first (do-nothing) scenario indicate that there will be ongoing negative impacts on the aquifer, such as a higher propensity for strong seawater intrusion into the Gaza aquifer. This scenario illustrates that, compared with 2010 situation of the baseline model, at the end of simulation period, year 2040, the amount of saltwater intrusion into the coastal aquifer will be increased by about 35 %, whereas the salinity will be increased by 34 %. In contrast, all three cases of the second (artificial recharge) scenario group can partly revert the present seawater intrusion. From the water budget point of view, compared with the first (do nothing) scenario, for year 2040, the water added to the aquifer by artificial recharge will reduces the amount of water entering the aquifer by seawater intrusion by 81, 77and 72 %, for the three recharge cases, respectively. Meanwhile, the salinity in the Gaza aquifer will be decreased by 15, 32 and 26% for the three cases, respectively.

Implementación del modelo hidrológico de Témez para la evaluación de recursos hídricos con GRASS GIS: fase superficial y subterránea

Se presenta la implementación del modelo hidrológico distribuido de Témez sobre GRASS GIS. Este modelo se utiliza para la evaluación de recursos hídricos en régimen natural con paso mensual y para la totalidad del territorio español, tal como aparece en el Libro Blanco del Agua en España. A partir de las variables de entrada, precipitación y evapotranspiración potencial y los parámetros hidrológicos, el modelo obtiene los mapas de los distintos almacenamientos, humedad en el suelo y volumen de acuífero, y de las variables de salida del ciclo hidrológico, evapotranspiración y escorrentía total, obtenida esta última como suma de la escorrentía superficial y subterránea. El objetivo final del trabajo es la implementación de los componentes superficiales y subterráneos en el modelo hidrológico, desarrollando para ello un programa que hace funcional en GRASS GIS el modelo matemático en que se basa la evaluación de recursos hídricos

First-order molecular descriptors for molecular steric similarity

En aquest treball s'analitza la contribució estèrica de les molècules a les seves propietats químiques i físiques, mitjançant l'avaluació del seu volum i de la seva mesura de semblança, a partir d'ara definits com a descriptors moleculars de primer ordre. La difeèsncia entre aquests dos conceptes ha estat aclarida: mentre que el volum és la magnitud de l'espai que ocupa la molècula com a entitat global, la mesura de semblança ens dóna una idea de com està distribuïda la densitat electrònica al llarg d'aquest volum, i reflecteix més les diferències locals existents. L'ús de diverses aproximacions per a l'obtenció d'ambdós valors ha estat analitzat sobre diferents classes d'isòmers

Evaluating a new LISFLOOD-FP formulation with data from the summer 2007 floods in Tewkesbury, UK

Two-dimensional flood inundation modelling is a widely used tool to aid flood risk management. In urban areas, where asset value and population density are greatest, the model spatial resolution required to represent flows through a typical street network (i.e. < 10m) often results in impractical computational cost at the whole city scale. Explicit diffusive storage cell models become very inefficient at such high resolutions, relative to shallow water models, because the stable time step in such schemes scales as a quadratic of resolution. This paper presents the calibration and evaluation of a recently developed new formulation of the LISFLOOD-FP model, where stability is controlled by the Courant–Freidrichs–Levy condition for the shallow water equations, such that, the stable time step instead scales linearly with resolution. The case study used is based on observations during the summer 2007 floods in Tewkesbury, UK. Aerial photography is available for model evaluation on three separate days from the 24th to the 31st of July. The model covered a 3.6 km by 2 km domain and was calibrated using gauge data from high flows during the previous month. The new formulation was benchmarked against the original version of the model at 20 m and 40 m resolutions, demonstrating equally accurate performance given the available validation data but at 67x faster computation time. The July event was then simulated at the 2 m resolution of the available airborne LiDAR DEM. This resulted in a significantly more accurate simulation of the drying dynamics compared to that simulated by the coarse resolution models, although estimates of peak inundation depth were similar.

Metallo-organic domino reactions: C-H versus C-C bond breaking

HL and MeL are prepared by condensing benzil dihydrazone with 2-formylpyridine and 2-acetylpyridine, respectively, in 1:2 molar proportions. While in a reaction with [Ru-(C6H6)Cl-2](2), HL yields the cation [Ru(C6H6){5,6-diphenyl-3-(pyridin-2-yl)- 1,2,4-triazine}Cl](+), MeL gives the cation [Ru(C6H6)(MeL)Cl](+). Both the cations are isolated as their hexafluorophosphate salts and characterised by X-ray crystallography. In the case of HL, double domino electrocyclic/elimination reactions are found to occur. The electrocyclic reaction occurs in a C=N-N=C-C=N fragment of HL and the elimination reaction involves breaking of a C-H bond of HL. Density functional calculations on model complexes indicate that the identified electrocyclic reaction is thermochemically as well as kinetically feasible for both HL and MeL in the gas phase. For a double domino reaction, similar to that operative in HL, to occur for MeL, breaking of a C-C bond would be required in the elimination step. Our model calculations show the energy barrier for this elimination step to be much higher (329.1 kJ mol(-1)) for MeL than that for HL (96.3 kJ mol(-1)). Thus, the domino reaction takes place for HL and not for MeL. This accounts for the observed stability of [Ru(C6H6)-(MeL)Cl](+) under the reaction conditions employed.

Non-testability of equal weights spatial dependence

We show that any invariant test for spatial autocorrelation in a spatial error or spatial lag model with equal weights matrix has power equal to size. This result holds under the assumption of an elliptical distribution. Under Gaussianity, we also show that any test whose power is larger than its size for at least one point in the parameter space must be biased.

Optimal observers for ARMA models

A polynomial-based ARMA model, when posed in a state-space framework can be regarded in many different ways. In this paper two particular state-space forms of the ARMA model are considered, and although both are canonical in structure they differ in respect of the mode in which disturbances are fed into the state and output equations. For both forms a solution is found to the optimal discrete-time observer problem and algebraic connections between the two optimal observers are shown. The purpose of the paper is to highlight the fact that the optimal observer obtained from the first state-space form, commonly known as the innovations form, is not that employed in an optimal controller, in the minimum-output variance sense, whereas the optimal observer obtained from the second form is. Hence the second form is a much more appropriate state-space description to use for controller design, particularly when employed in self-tuning control schemes.

A comparative analysis of projected impacts of climate change on river runoff from global and catchment-scale hydrological models

We present a comparative analysis of projected impacts of climate change on river runoff from two types of distributed hydrological model, a global hydrological model (GHM) and catchment-scale hydrological models (CHM). Analyses are conducted for six catchments that are global in coverage and feature strong contrasts in spatial scale as well as climatic and development conditions. These include the Liard (Canada), Mekong (SE Asia), Okavango (SW Africa), Rio Grande (Brazil), Xiangu (China) and Harper's Brook (UK). A single GHM (Mac-PDM.09) is applied to all catchments whilst different CHMs are applied for each catchment. The CHMs typically simulate water resources impacts based on a more explicit representation of catchment water resources than that available from the GHM, and the CHMs include river routing. Simulations of average annual runoff, mean monthly runoff and high (Q5) and low (Q95) monthly runoff under baseline (1961-1990) and climate change scenarios are presented. We compare the simulated runoff response of each hydrological model to (1) prescribed increases in global mean temperature from the HadCM3 climate model and (2)a prescribed increase in global-mean temperature of 2oC for seven GCMs to explore response to climate model and structural uncertainty. We find that differences in projected changes of mean annual runoff between the two types of hydrological model can be substantial for a given GCM, and they are generally larger for indicators of high and low flow. However, they are relatively small in comparison to the range of projections across the seven GCMs. Hence, for the six catchments and seven GCMs we considered, climate model structural uncertainty is greater than the uncertainty associated with the type of hydrological model applied. Moreover, shifts in the seasonal cycle of runoff with climate change are presented similarly by both hydrological models, although for some catchments the monthly timing of high and low flows differs.This implies that for studies that seek to quantify and assess the role of climate model uncertainty on catchment-scale runoff, it may be equally as feasible to apply a GHM as it is to apply a CHM, especially when climate modelling uncertainty across the range of available GCMs is as large as it currently is. Whilst the GHM is able to represent the broad climate change signal that is represented by the CHMs, we find, however, that for some catchments there are differences between GHMs and CHMs in mean annual runoff due to differences in potential evaporation estimation methods, in the representation of the seasonality of runoff, and in the magnitude of changes in extreme monthly runoff, all of which have implications for future water management issues.

Groundwater residence time distributions in peatlands: implications for peat decomposition and accumulation

Peat soils consist of poorly decomposed plant detritus, preserved by low decay rates, and deep peat deposits are globally significant stores in the carbon cycle. High water tables and low soil temperatures are commonly held to be the primary reasons for low peat decay rates. However, recent studies suggest a thermodynamic limit to peat decay, whereby the slow turnover of peat soil pore water may lead to high concentrations of phenols and dissolved inorganic carbon. In sufficient concentrations, these chemicals may slow or even halt microbial respiration, providing a negative feedback to peat decay. We document the analysis of a simple, one-dimensional theoretical model of peatland pore water residence time distributions (RTDs). The model suggests that broader, thicker peatlands may be more resilient to rapid decay caused by climate change because of slow pore water turnover in deep layers. Even shallow peat deposits may also be resilient to rapid decay if rainfall rates are low. However, the model suggests that even thick peatlands may be vulnerable to rapid decay under prolonged high rainfall rates, which may act to flush pore water with fresh rainwater. We also used the model to illustrate a particular limitation of the diplotelmic (i.e., acrotelm and catotelm) model of peatland structure. Model peatlands of contrasting hydraulic structure exhibited identical water tables but contrasting RTDs. These scenarios would be treated identically by diplotelmic models, although the thermodynamic limit suggests contrasting decay regimes. We therefore conclude that the diplotelmic model be discarded in favor of model schemes that consider continuous variation in peat properties and processes.

Focusing the kaleidoscope: investigating the newly formed role of 'Academic Lead’ at a research-led University

Now, more than ever, higher education institutions are reflecting on the need for flexible leadership models to help adapt to the fast changing academic environment. Rapid shifts in the sector are contributing to a kaleidoscopic ‘supercomplexity’ of challenges, structures, processes and value frameworks for academics who lead and for those who are led. How are institutions’ leadership structures and roles developing in response to these changes? And how do these responses affect academic staff in relation to their identity, status and career trajectory? This paper reports on a Leadership Foundation funded research project exploring the ways in which one UK institution has implemented a new ‘distributed’ leadership model. Crucially, the project examines the impact of the model on both those who are leaders and those being led.

The Potential Of Kohonen Method For Classifying Rivers Focused On Regional Environmental Flow Determination With ELOHA

This article highlights the potential benefits that the Kohonen method has for the classification of rivers with similar characteristics by determining regional ecological flows using the ELOHA (Ecological Limits of Hydrologic Alteration) methodology. Currently, there are many methodologies for the classification of rivers, however none of them include the characteristics found in Kohonen method such as (i) providing the number of groups that actually underlie the information presented, (ii) used to make variable importance analysis, (iii) which in any case can display two-dimensional classification process, and (iv) that regardless of the parameters used in the model the clustering structure remains. In order to evaluate the potential benefits of the Kohonen method, 174 flow stations distributed along the great river basin “Magdalena-Cauca” (Colombia) were analyzed. 73 variables were obtained for the classification process in each case. Six trials were done using different combinations of variables and the results were validated against reference classification obtained by Ingfocol in 2010, whose results were also framed using ELOHA guidelines. In the process of validation it was found that two of the tested models reproduced a level higher than 80% of the reference classification with the first trial, meaning that more than 80% of the flow stations analyzed in both models formed invariant groups of streams.

On testing the Phillips curves, the IS Curves, and the interaction between fiscal and monetary policies

Esta tese é composta por três ensaios sobre testes empíricos de curvas de Phillips, curvas IS e a interação entre as políticas fiscal e monetária. O primeiro ensaio ("Curvas de Phillips: um Teste Abrangente") testa curvas de Phillips usando uma especificação autoregressiva de defasagem distribuída (ADL) que abrange a curva de Phillips Aceleracionista (APC), a curva de Phillips Novo Keynesiana (NKPC), a curva de Phillips Híbrida (HPC) e a curva de Phillips de Informação Rígida (SIPC). Utilizamos dados dos Estados Unidos (1985Q1--2007Q4) e do Brasil (1996Q1--2012Q2), usando o hiato do produto e alternativamente o custo marginal real como medida de pressão inflacionária. A evidência empírica rejeita as restrições decorrentes da NKPC, da HPC e da SIPC, mas não rejeita aquelas da APC. O segundo ensaio ("Curvas IS: um Teste Abrangente") testa curvas IS usando uma especificação ADL que abrange a curva IS Keynesiana tradicional (KISC), a curva IS Novo Keynesiana (NKISC) e a curva IS Híbrida (HISC). Utilizamos dados dos Estados Unidos (1985Q1--2007Q4) e do Brasil (1996Q1--2012Q2). A evidência empírica rejeita as restrições decorrentes da NKISC e da HISC, mas não rejeita aquelas da KISC. O terceiro ensaio ("Os Efeitos da Política Fiscal e suas Interações com a Política Monetária") analisa os efeitos de choques na política fiscal sobre a dinâmica da economia e a interação entre as políticas fiscal e monetária usando modelos SVARs. Testamos a Teoria Fiscal do Nível de Preços para o Brasil analisando a resposta do passivo do setor público a choques no superávit primário. Para a identificação híbrida, encontramos que não é possível distinguir empiricamente entre os regimes Ricardiano (Dominância Monetária) e não-Ricardiano (Dominância Fiscal). Entretanto, utilizando a identificação de restrições de sinais, existe evidência que o governo seguiu um regime Ricardiano (Dominância Monetária) de janeiro de 2000 a junho de 2008.

Ensaios sobre a estrutura a termo da taxa de juros

Esta tese é composta de três artigos que analisam a estrutura a termo das taxas de juros usando diferentes bases de dados e modelos. O capítulo 1 propõe um modelo paramétrico de taxas de juros que permite a segmentação e choques locais na estrutura a termo. Adotando dados do tesouro americano, duas versões desse modelo segmentado são implementadas. Baseado em uma sequência de 142 experimentos de previsão, os modelos propostos são comparados à benchmarks e concluí-se que eles performam melhor nos resultados das previsões fora da amostra, especialmente para as maturidades curtas e para o horizonte de previsão de 12 meses. O capítulo 2 acrescenta restrições de não arbitragem ao estimar um modelo polinomial gaussiano dinâmico de estrutura a termo para o mercado de taxas de juros brasileiro. Esse artigo propõe uma importante aproximação para a série temporal dos fatores de risco da estrutura a termo, que permite a extração do prêmio de risco das taxas de juros sem a necessidade de otimização de um modelo dinâmico completo. Essa metodologia tem a vantagem de ser facilmente implementada e obtém uma boa aproximação para o prêmio de risco da estrutura a termo, que pode ser usada em diferentes aplicações. O capítulo 3 modela a dinâmica conjunta das taxas nominais e reais usando um modelo afim de não arbitagem com variáveis macroeconômicas para a estrutura a termo, afim de decompor a diferença entre as taxas nominais e reais em prêmio de risco de inflação e expectativa de inflação no mercado americano. Uma versão sem variáveis macroeconômicas e uma versão com essas variáveis são implementadas e os prêmios de risco de inflação obtidos são pequenos e estáveis no período analisado, porém possuem diferenças na comparação dos dois modelos analisados.