The Levenberg-Marquardt algorithm is a classical method for solving nonlinear systems of equations that can come from various applications in engineering and economics. Recently, Levenberg-Marquardt methods turned out to be a valuable principle for obtaining fast convergence to a solution of the nonlinear system if the classical nonsingularity assumption is replaced by a weaker error bound condition. In this way also problems with nonisolated solutions can be treated successfully. Such problems increasingly arise in engineering applications and in mathematical programming. In this thesis we use Levenberg-Marquardt algorithms to deal with nonlinear equations, multi-objective optimization and complementarity problems. We develop new algorithms for solving these problems and investigate their convergence properties. For sufficiently smooth nonlinear equations we provide convergence results for inexact Levenberg-Marquardt type algorithms. In particular, a sharp bound on the maximal level of inexactness that is sufficient for a quadratic (or a superlinear) rate of convergence is derived. Moreover, the theory developed is used to show quadratic convergence of a robust projected Levenberg-Marquardt algorithm. The use of Levenberg-Marquardt type algorithms for unconstrained multi-objective optimization problems is investigated in detail. In particular, two globally and locally quadratically convergent algorithms for these problems are developed. Moreover, assumptions under which the error bound condition for a Pareto-critical system is fulfilled are derived. We also treat nonsmooth equations arising from reformulating complementarity problems by means of NCP functions. For these reformulations, we show that existing smoothness conditions are not satisfied at degenerate solutions. Moreover, we derive new results for positively homogeneous functions. The latter results are used to show that appropriate weaker smoothness conditions (enabling a local Q-quadratic rate of convergence) hold for certain reformulations. / Der Levenberg-Marquardt-Algorithmus ist ein klassisches Verfahren zur Lösung von nichtlinearen Gleichungssystemen, welches in verschiedenen Anwendungen der Ingenieur-und Wirtschaftswissenschaften vorkommen kann. Kürzlich, erwies sich das Verfahren als ein wertvolles Instrument für die Gewährleistung einer schnelleren Konvergenz für eine Lösung des nichtlinearen Systems, wenn die klassische nichtsinguläre Annahme durch eine schwächere Fehlerschranke der eingebundenen Bedingung ersetzt wird. Auf diese Weise, lassen sich ebenfalls Probleme mit nicht isolierten Lösungen erfolgreich behandeln. Solche Probleme ergeben sich zunehmend in den praktischen, ingenieurwissenschaftlichen Anwendungen und in der mathematischen Programmierung. In dieser Arbeit verwenden wir Levenberg-Marquardt- Algorithmus für nichtlinearere Gleichungen, multikriterielle Optimierung - und nichtlineare Komplementaritätsprobleme. Wir entwickeln neue Algorithmen zur Lösung dieser Probleme und untersuchen ihre Konvergenzeigenschaften. Für ausreichend differenzierbare nichtlineare Gleichungen, analysieren und bieten wir Konvergenzergebnisse für ungenaue Levenberg-Marquardt-Algorithmen Typen. Insbesondere, bieten wir eine strenge Schranke für die maximale Höhe der Ungenauigkeit, die ausreichend ist für eine quadratische (oder eine superlineare) Rate der Konvergenz. Darüber hinaus, die entwickelte Theorie wird verwendet, um quadratische Konvergenz eines robusten projizierten Levenberg-Marquardt-Algorithmus zu zeigen. Die Verwendung von Levenberg-Marquardt-Algorithmen Typen für unbeschränkte multikriterielle Optimierungsprobleme im Detail zu untersucht. Insbesondere sind zwei globale und lokale quadratische konvergente Algorithmen für multikriterielle Optimierungsprobleme entwickelt worden. Die Annahmen wurden hergeleitet, unter welche die Fehlerschranke der eingebundenen Bedingung für ein Pareto-kritisches System erfüllt ist. Wir behandeln auch nicht differenzierbare nichtlineare Gleichungen aus Umformulierung der nichtlinearen Komplementaritätsprobleme durch NCP-Funktionen. Wir zeigen für diese Umformulierungen, dass die bestehenden differenzierbaren Bedingungen nicht zufrieden mit degenerierten Lösungen sind. Außerdem, leiten wir neue Ergebnisse für positiv homogene NCP-Funktionen. Letztere Ergebnisse werden verwendet um zu zeigen, dass geeignete schwächeren differenzierbare Bedingungen (so dass eine lokale Q-quadratische Konvergenzgeschwindigkeit ermöglichen) für bestimmte Umformulierungen gelten.

Massive shales and fractures are the main cause of seismic anisotropy in the upper-most part of the crust, caused either by sedimentary or tectonic processes. Neglecting the effect of seismic anisotropy in seismic processing algorithms may incorrectly image the seismic reflectors. This will also influence the quantitative amplitude analysis such as the acoustic or elastic impedance inversion and amplitude versus offsets analysis. Therefore it is important to obtain anisotropy parameters from seismic data. Conventional layer stripping inversion schemes and reflector based reflectivity inversion methods are solely dependent upon a specific reflector, without considering the effect of the other layers. This, on one hand, does not take the effect of transmission in reflectivity inversion into the account, and on the other hand, ignores the information from the waves travelling toward the lower layers. I provide a framework to integrate the information for each specific layer from all the rays which have travelled across this layer. To estimate anisotropy parameters I have implemented unconstrained minimization algorithms such as nonlinear conjugate gradients and variable metric methods, I also provide a nonlinear least square method, based on the Levenberg-Marquardt algorithm. In a stack of horizontal transversely isotropic layers with vertical axis of symmetry, where the layer properties are laterally invariant, we provide two different inversion schemes; traveltime and waveform inversion. / Both inversion schemes utilize compressional and joint compressional and converted shear waves. A new exact traveltime equation has been formulated for a dipping transversely isotropic system of layers. These traveltimes are also parametrized by the ray parameters for each ray element. I use the Newton method of minimization to estimate the ray parameter using a random prior model from a uniform distribution. Numerical results show that with the assumption of weak anisotropy, Thomsen’s anisotropy parameters can be estimated with a high accuracy. The inversion algorithms have been implemented as a software package in a C++ object oriented environment.

Identificação e predição do comportamento aeroelástico representa um grande desafio para a análise e controle de fenômenos aeroelásticos adversos. A modelagem aeroelástica requer informações tanto sobre a dinâmica estrutural quanto sobre o comportamento aerodinâmico não estacionário. No entanto, a maioria das metodologias disponíveis atualmente são baseadas no desacoplamento entre o modelo estrutural e o modelo aerodinâmico não estacionário. Conseqüentemente, métodos alternativos são bem vindos na área de pesquisa aerolástica. Entre os métodos alternativos está o funcional multicamada, que fornece uma rigorosa representação matemática apropriada para modelagem aeroelástica e pode ser obtido através de redes neurais artificiais. Esse trabalho apresenta uma aplicação desse método, consistindo de um procedimento de identificação baseado em redes neurais artificiais que representam o funcional da resposta aeroelástica. O modelo neural foi treinado usando o algoritmo de Levenberg-Marquardt, o qual tem sido considerado um método de otimização muito eficiente. Ele combina a garantia de convergência do método do gradiente e o alto desempenho do método de Newton, sem a necessidade de calcular as derivadas de segunda ordem. Um modelo de asa ensaiado em túnel de vento foi usado para fornecer a resposta aeroelástica. A asa foi fixada a uma mesa giratória e um motor elétrico lhe fornecia o movimento de incidência. Essa representação aeroelástica funcional foi testada para diversas condições operacionais do túnel de vento. Os resultados mostraram que o uso de redes neurais na identificação da resposta aeroelástica é um método alternativo promissor, o qual permite uma rápida avaliação da resposta aerolástica do modelo. / Identification and prediction of aeroelastic behavior presents a significant challenge for the analysis and control of adverse aeroelastic phenomena. Aeroelastic modeling requires information from both structural dynamics and unsteady aerodynamic behavior. However, the majority of methodologies available today are based on the decoupling of structural model from the unsteady aerodynamic model. Therefore, alternative methods are mostly welcome in the aeroelastic research field. Among the alternative methods there is the multi-layer functional (MLF), that allows a rigorous mathematical framework appropriate for aeroelastic modeling and can be realized by means of artificial neural networks. This work presents an identification procedure based on artificial neural networks to represent the motion-induced aeroelastic response functional. The neural network model has been trained using the Levenberg-Marquardt algorithm that has been considered a very efficient optimization method. It combines the guaranteed convergence of steepest descent and the higher performance of the Newton\'s method, without the necessity of second derivatives calculation. A wind tunnel aeroelastic wing model has been used to provide motion-induced aeroelastic responses. The wing has been fixed to a turntable, and an electrical motor provides the incidence motion to the wing. This aeroelastic functional representation is then tested for a range of the wind tunnel model operational boundaries. The results showed that the use of neural networks in the aeroelastic response identification is a promising alternative method, which allows fast evaluation of aeroelastic response model.

Identificação e predição do comportamento aeroelástico representa um grande desafio para a análise e controle de fenômenos aeroelásticos adversos. A modelagem aeroelástica requer informações tanto sobre a dinâmica estrutural quanto sobre o comportamento aerodinâmico não estacionário. No entanto, a maioria das metodologias disponíveis atualmente são baseadas no desacoplamento entre o modelo estrutural e o modelo aerodinâmico não estacionário. Conseqüentemente, métodos alternativos são bem vindos na área de pesquisa aerolástica. Entre os métodos alternativos está o funcional multicamada, que fornece uma rigorosa representação matemática apropriada para modelagem aeroelástica e pode ser obtido através de redes neurais artificiais. Esse trabalho apresenta uma aplicação desse método, consistindo de um procedimento de identificação baseado em redes neurais artificiais que representam o funcional da resposta aeroelástica. O modelo neural foi treinado usando o algoritmo de Levenberg-Marquardt, o qual tem sido considerado um método de otimização muito eficiente. Ele combina a garantia de convergência do método do gradiente e o alto desempenho do método de Newton, sem a necessidade de calcular as derivadas de segunda ordem. Um modelo de asa ensaiado em túnel de vento foi usado para fornecer a resposta aeroelástica. A asa foi fixada a uma mesa giratória e um motor elétrico lhe fornecia o movimento de incidência. Essa representação aeroelástica funcional foi testada para diversas condições operacionais do túnel de vento. Os resultados mostraram que o uso de redes neurais na identificação da resposta aeroelástica é um método alternativo promissor, o qual permite uma rápida avaliação da resposta aerolástica do modelo. / Identification and prediction of aeroelastic behavior presents a significant challenge for the analysis and control of adverse aeroelastic phenomena. Aeroelastic modeling requires information from both structural dynamics and unsteady aerodynamic behavior. However, the majority of methodologies available today are based on the decoupling of structural model from the unsteady aerodynamic model. Therefore, alternative methods are mostly welcome in the aeroelastic research field. Among the alternative methods there is the multi-layer functional (MLF), that allows a rigorous mathematical framework appropriate for aeroelastic modeling and can be realized by means of artificial neural networks. This work presents an identification procedure based on artificial neural networks to represent the motion-induced aeroelastic response functional. The neural network model has been trained using the Levenberg-Marquardt algorithm that has been considered a very efficient optimization method. It combines the guaranteed convergence of steepest descent and the higher performance of the Newton\'s method, without the necessity of second derivatives calculation. A wind tunnel aeroelastic wing model has been used to provide motion-induced aeroelastic responses. The wing has been fixed to a turntable, and an electrical motor provides the incidence motion to the wing. This aeroelastic functional representation is then tested for a range of the wind tunnel model operational boundaries. The results showed that the use of neural networks in the aeroelastic response identification is a promising alternative method, which allows fast evaluation of aeroelastic response model.

Orientadores: Francisco de Assis Magalhães Gomes Neto, Jose Mario Martinez / Tese (doutorado) - Universidade Estadual de Campinas, Instituto de Matematica, Estatistica e Computação Cientifica / Made available in DSpace on 2018-08-12T08:17:16Z (GMT). No. of bitstreams: 1 Medeiros_LuizAntoniodaSilva_D.pdf: 2528214 bytes, checksum: 42e1946a32b63c9fc5cd56b10d24d5cb (MD5) Previous issue date: 2008 / Resumo: Neste trabalho, apresentamos um algoritmo iterativo para a minimização de somas de quadrados de funções suaves, com restrições de caixa. O algoritmo é fortemente inspirado no trabalho de Birgin e Martínez [4]. A diferença principal está na escolha da direção de busca e na introdução de uma nova técnica de aceleração, usada para atualizar o passo. A cada iteração, definimos uma face ativa e resolvemos, nessa face, um subproblema quadrático irrestrito através do método evenberg-Marquardt (ver [26], [28] e [33]), obtendo uma direção de descida e uma aproximação x+ para a solução do problema. Ainda usando apenas as variáveis livres, tentamos acelerar o método definindo uma nova aproximaçaoo xa como combinação linear das últimas p - 1 aproximações da solução e do vetor x+. Os coeficientes desta combinação linear são calculados convenientemente através da resolução de um problema de Quadrados Mínimos com uma restrição de igualdade. O subproblema que determina o passo acelerado leva em conta as informações sobre a função objetivo nessas p soluções aproximadas. Como em [4], executamos uma busca linear ao longo da direção e usamos técnicas de projeção para adicionar novas restrições. Para deixar a face ativa, usamos a direção do gradiente espectral projetado [5]. Experimentos númericos são apresentados para confirmar a eficiência e robustez do novo algoritmo. / Abstract: In this work, we present an active set algorithm for minimizing the sum of squares of smooth functions, with box constraints. The algorithm is highly inspired in the work of Birgin and Mart'inez [4]. The differences are concentrated on the chosen search direction and on the use of an acceleration technique to update the step. At each iteration, we define an active face and solve an unconstrained quadratic subproblem using the Levenberg-Marquardt method (see [26], [28] and [33]), obtaining a descent direction and an approximate solution x+. Using only the free variables, we try to accelerate the method defining a new solution xa as a linear combination of the last p-1 approximate solutions together with x+. The coefficients of this linear combination are conveniently computed solving a constrained least squares problem that takes into account the objective function values of these p approximate solutions. Like in [4], we compute a line search and use projection techniques to add new constraints to the active set. The spectral projected gradient [5] is used to leave the current active face. Numerical experiments confirm that the algorithm is both efficient and robust. / Doutorado / Matematica Aplicada / Doutor em Matemática Aplicada

The Levenberg-Marquardt algorithm is a classical method for solving nonlinear systems of equations that can come from various applications in engineering and economics. Recently, Levenberg-Marquardt methods turned out to be a valuable principle for obtaining fast convergence to a solution of the nonlinear system if the classical nonsingularity assumption is replaced by a weaker error bound condition. In this way also problems with nonisolated solutions can be treated successfully. Such problems increasingly arise in engineering applications and in mathematical programming. In this thesis we use Levenberg-Marquardt algorithms to deal with nonlinear equations, multi-objective optimization and complementarity problems. We develop new algorithms for solving these problems and investigate their convergence properties. For sufficiently smooth nonlinear equations we provide convergence results for inexact Levenberg-Marquardt type algorithms. In particular, a sharp bound on the maximal level of inexactness that is sufficient for a quadratic (or a superlinear) rate of convergence is derived. Moreover, the theory developed is used to show quadratic convergence of a robust projected Levenberg-Marquardt algorithm. The use of Levenberg-Marquardt type algorithms for unconstrained multi-objective optimization problems is investigated in detail. In particular, two globally and locally quadratically convergent algorithms for these problems are developed. Moreover, assumptions under which the error bound condition for a Pareto-critical system is fulfilled are derived. We also treat nonsmooth equations arising from reformulating complementarity problems by means of NCP functions. For these reformulations, we show that existing smoothness conditions are not satisfied at degenerate solutions. Moreover, we derive new results for positively homogeneous functions. The latter results are used to show that appropriate weaker smoothness conditions (enabling a local Q-quadratic rate of convergence) hold for certain reformulations. / Der Levenberg-Marquardt-Algorithmus ist ein klassisches Verfahren zur Lösung von nichtlinearen Gleichungssystemen, welches in verschiedenen Anwendungen der Ingenieur-und Wirtschaftswissenschaften vorkommen kann. Kürzlich, erwies sich das Verfahren als ein wertvolles Instrument für die Gewährleistung einer schnelleren Konvergenz für eine Lösung des nichtlinearen Systems, wenn die klassische nichtsinguläre Annahme durch eine schwächere Fehlerschranke der eingebundenen Bedingung ersetzt wird. Auf diese Weise, lassen sich ebenfalls Probleme mit nicht isolierten Lösungen erfolgreich behandeln. Solche Probleme ergeben sich zunehmend in den praktischen, ingenieurwissenschaftlichen Anwendungen und in der mathematischen Programmierung. In dieser Arbeit verwenden wir Levenberg-Marquardt- Algorithmus für nichtlinearere Gleichungen, multikriterielle Optimierung - und nichtlineare Komplementaritätsprobleme. Wir entwickeln neue Algorithmen zur Lösung dieser Probleme und untersuchen ihre Konvergenzeigenschaften. Für ausreichend differenzierbare nichtlineare Gleichungen, analysieren und bieten wir Konvergenzergebnisse für ungenaue Levenberg-Marquardt-Algorithmen Typen. Insbesondere, bieten wir eine strenge Schranke für die maximale Höhe der Ungenauigkeit, die ausreichend ist für eine quadratische (oder eine superlineare) Rate der Konvergenz. Darüber hinaus, die entwickelte Theorie wird verwendet, um quadratische Konvergenz eines robusten projizierten Levenberg-Marquardt-Algorithmus zu zeigen. Die Verwendung von Levenberg-Marquardt-Algorithmen Typen für unbeschränkte multikriterielle Optimierungsprobleme im Detail zu untersucht. Insbesondere sind zwei globale und lokale quadratische konvergente Algorithmen für multikriterielle Optimierungsprobleme entwickelt worden. Die Annahmen wurden hergeleitet, unter welche die Fehlerschranke der eingebundenen Bedingung für ein Pareto-kritisches System erfüllt ist. Wir behandeln auch nicht differenzierbare nichtlineare Gleichungen aus Umformulierung der nichtlinearen Komplementaritätsprobleme durch NCP-Funktionen. Wir zeigen für diese Umformulierungen, dass die bestehenden differenzierbaren Bedingungen nicht zufrieden mit degenerierten Lösungen sind. Außerdem, leiten wir neue Ergebnisse für positiv homogene NCP-Funktionen. Letztere Ergebnisse werden verwendet um zu zeigen, dass geeignete schwächeren differenzierbare Bedingungen (so dass eine lokale Q-quadratische Konvergenzgeschwindigkeit ermöglichen) für bestimmte Umformulierungen gelten.

Dans la fabrication d’un produit, la maîtrise des propriétés physiques des matériaux utilisés est indispensable. Il est donc nécessaire de déterminer leurs propriétés comportementales. On exploite en général des propriétés physiques intermédiaires telles que les propriétés électromagnétiques. Nous avons mis au point une technique de caractérisation sans contact de matériaux non transparents en transposant les concepts de base de l’ellipsométrie optique en hyperfréquence. La caractérisation se fait par résolution d’un problème inverse par deux méthodes numériques : une méthode d’optimisation classique utilisant l’algorithme itératif de Levenberg Marquardt et une méthode de régression par l’uti1isation de réseaux de neurones du type perceptron multicouches. Avec la première méthode, on détermine deux paramètres du matériau sous test à savoir l’indice de réfraction et l’indice d’extinction. Avec la deuxième, on détermine les deux indices ainsi que l’épaisseur de 1’échantillon. Pour la validation, nous avons monté un banc expérimental en espace libre à 30 GHz en transmission et en incidence oblique avec lequel nous avons effectué des mesures sur des échantillons de téflon et d’époxy de différentes épaisseurs (1 à 30 mm). Nous avons obtenu une caractérisation satisfaisante de l’indice et de l’épaisseur. Nous avons ensuite fait des mesures de trois types de papier dont la caractérisation de l’indice était satisfaisante sans toutefois les discriminer. Ces travaux ont montré la possibilité de caractériser des matériaux épais et non transparents par une technique ellipsométrique / The control of the physical properties of materials used to manufacture a product is essential. Therefore, it is necessary to determine their behavioral properties. Generally, we get them through intermediate physical properties such as electromagnetic properties. We have developed a technique for contactless characterization of non-transparent materials by applying the basic concepts of optical ellipsometry in the microwave domain. The characterization is done by solving an inverse problem through two different numerical methods: a classical optimization method using the iterative algorithm of Levenberg Marquardt and a regression method by using neural networks particularly the multilayer perceptron. With the first method, we can determine two parameters which are the refractive index and the extinction index of the sample under test. With the second method, we can determine the indices and the thickness of the sample. As a validation, we set up a 30 GHz - experimental free space bench configurated for oblique transmission incidence measurement that we have used to carry out measurements on PTFE and epoxy samples having different thicknesses (1 to 30 mm). We obtained a satisfactory characterization of the index and thickness. Then, we have carried out measurements on three types of paper. The index was satisfactorily characterized but they could not be distinguished. These studies have shown that it is possible to characterize thick and non-transparent materials using ellipsometric technics

A passive radar systems opportunistic ability to exploit ambient radio signal reflections makes it ideal for covert target tracking. This strategy, referred to as passive covert radar (PCR) or passive coherent location (PCL), typically exploits FM radio or television signals from powerful local transmitters. In addition to covertness, the absence of a dedicated transmitter helps reduce costs and overall system complexity. While a variety of measurements can be used to estimate a targets position and velocity, such as time difference of arrival (TDOA) and direction of arrival (DOA), this thesis focuses on using only Doppler shift measurements to estimate a targets state. The work presented in this thesis examines the use of Doppler shift measurements from multiple receivers to solve the target tracking and association problem. A nonlinear least squares error (NLSE) estimation technique, called the Levenberg-Marquardt (L-M) algorithm, is used to determine a targets state (position, velocity) from these Doppler shift measurements. More than one target state can potentially produce identical Doppler shift profiles. In a single-receiver, single-target scenario, it is shown that three additional ghost targets caused by symmetry produce the same Doppler shift response. These ghosts may make state estimation impossible if receive antennas are not physically positioned to block out ghost targets. While the NLSE technique tends to give an accurate solution in one quadrant, three other solutions will symmetrically exist in each of the remaining three quadrants. The addition of either another receiver or another measurement (such as DOA) is needed to break this quadrant ambiguity. This thesis considers adding multiple receivers to accurately associate and track multiple targets. Two target association methods (sequential and simultaneous) are developed, and their computational requirements and accuracy are compared. A grid-aided L-M search technique is investigated in an attempt to provide a better initial target state guess to these association and tracking algorithms. The analysis and simulation results suggest it is feasible to perform multi-target association and tracking using Doppler shift as the sole measurement. Both of the proposed methods gave optimal target association and converged to reasonably accurate state estimates in most of the Monte Carlo runs.

The purpose of this work was to develop methods to estimate the diffusion coefficient and elimination constant for dexamethasone in subcutaneous tissue. Solutions to the diffusion equation were found for different conditions relevant to implantation and injection. These solutions were then used as models for measured autoradiography data where the unknown model parameters were the diffusion coefficient and the elimination constant. The diffusion coefficient and elimination constant were then estimated by curve fitting the measured data to these models. Having these estimates would be of practical importance since inflammation surrounding implantable glucose sensors may be controlled through local release of dexamethasone at the site of implantation. Derivation of the appropriate model, how the model was used to estimate D and k, and various specific profile examples were investigated in detail. Osmotic pumps containing [3H]- dexamethasone were implanted into the subcutaneous tissue of rats. Digital autoradiography was used to measure the distribution of the [3H]-dexamethasone within the subcutaneous tissue at 6, 24, and 60 hours after implantation. Measured concentration profiles, near the catheter tip through which the agent was released, were compared to solutions of the diffusion equation in order to characterize drug diffusion coefficients and elimination constants. There was good agreement between the experimental data and the mathematical model used for estimation. The diffusion coefficient for dexamethasone in subcutaneous tissue was found to be D = 4.11+-1.77x10E-10 m2/s, and the elimination rate constant was found to be k = 3.65+-2.24x10E-5/s. Additionally, [3H]-dexamethasone was injected into the subcutaneous tissue of rats. Digital autoradiography was again used to measure the distribution of the [3H]- dexamethasone within the subcutaneous tissue at 2.5 and 20 minutes after injection. Measured concentration profiles were again compared to a mathematical model of drug diffusion for injection. There was good agreement between the experimental data and the mathematical model. The diffusion coefficient found using this simple injection method was 4.01+-2.01x10E-10 m2/s. The simple method given here for the determination of the diffusion coefficient is general enough to be applied to other substances and tissues as well.

A reconstruction algorithm in computerized tomography is a procedure for reconstructing the attenuation coefficientscient, a real-valued function associated with the object of interest, from the measured projection data. Generally speaking, reconstruction algorithms in CT fall into two categories: direct, e.g., filtered back-projection (FBP), or iterative. In this thesis, we discuss a new fast matrix-free iterative reconstruction method based on a polyenergetic model. While most modern x-ray CT scanners rely on the well-known filtered back-projection algorithm, the corresponding reconstructions can be corrupted by beam hardening artifacts. These artifacts arise from the unrealistic physical assumption of monoenergetic x-ray beams. In this thesis, to compensate, we use an alternative model that accounts for differential absorption of polyenergetic x-ray photons and discretize it directly. We do not assume any prior knowledge about the physical properties of the scanned object. We study and implement different solvers and nonlinear unconstrained optimization methods, such as a Newton-like method and an extension of the Levenberg-Marquardt-Fletcher algorithm. We explain how we can use the structure of the Radon matrix and the properties of FBP to make our method matrix-free and fast. Finally, we discuss how we regularize our problem by applying different regularization methods, such as Tikhonov and regularization in the 1-norm. We present numerical reconstructions based on the associated nonlinear discrete formulation incorporating various iterative optimization methods.