Optimization is a scientific discipline that lies at the boundary<br />between pure and applied mathematics. Indeed, while on the one hand<br />some of its developments involve rather theoretical concepts, its<br />most successful algorithms are on the other hand heavily used by<br />numerous companies to solve scheduling and design problems on a<br />daily basis.<br /><br />Our research started with the study of the conic formulation for<br />convex optimization problems. This approach was already studied in<br />the seventies but has recently gained a lot of interest due to<br />development of a new class of algorithms called interior-point<br />methods. This setting is able to exploit the two most important<br />characteristics of convexity:<br /><br />- a very rich duality theory (existence of a dual problem that is<br />strongly related to the primal problem, with a very symmetric<br />formulation),<br />- the ability to solve these problems efficiently,<br />both from the theoretical (polynomial algorithmic complexity) and<br />practical (implementations allowing the resolution of large-scale<br />problems) points of view.<br /><br />Most of the research in this area involved so-called self-dual<br />cones, where the dual problem has exactly the same structure as the<br />primal: the most famous classes of convex optimization problems<br />(linear optimization, convex quadratic optimization and semidefinite<br />optimization) belong to this category. We brought some contributions <br />in this field:<br />- a survey of interior-point methods for linear optimization, with <br />an emphasis on the fundamental principles that lie behind the design <br />of these algorithms,<br />- a computational study of a method of linear approximation of convex <br />quadratic optimization (more precisely, the second-order cone that <br />can be used in the formulation of quadratic problems is replaced by a <br />polyhedral approximation whose accuracy can be guaranteed a priori),<br />- an application of semidefinite optimization to classification, <br />whose principle consists in separating different classes of patterns <br />using ellipsoids defined in the feature space (this approach was <br />successfully applied to the prediction of student grades).<br /><br />However, our research focussed on a much less studied category of<br />convex problems which does not rely on self-dual cones, i.e.<br />structured problems whose dual is formulated very differently from<br />the primal. We studied in particular<br />- geometric optimization, developed in the late sixties, which<br />possesses numerous application in the field of engineering<br />(entropy optimization, used in information theory, also belongs to<br />this class of problems)<br />- l_p-norm optimization, a generalization of linear and convex<br />quadratic optimization, which allows the formulation of constraints<br />built around expressions of the form |ax+b|^p (where p is a fixed<br />exponent strictly greater than 1).<br /><br />For each of these classes of problems, we introduced a new type of<br />convex cone that made their formulation as standard conic problems<br />possible. This allowed us to derive very simplified proofs of the<br />classical duality results pertaining to these problems, notably weak<br />duality (a mere consequence of convexity) and the absence of a<br />duality gap (strong duality property without any constraint<br />qualification, which does not hold in the general convex case). We<br />also uncovered a very surprising result that stipulates that<br />geometric optimization can be viewed as a limit case of l_p-norm<br />optimization. Encouraged by the similarities we observed, we<br />developed a general framework that encompasses these two classes of<br />problems and unifies all the previously obtained conic formulations.<br /><br />We also brought our attention to the design of interior-point<br />methods to solve these problems. The theory of polynomial algorithms<br />for convex optimization developed by Nesterov and Nemirovski asserts<br />that the main ingredient for these methods is a computable<br />self-concordant barrier function for the corresponding cones. We<br />were able to define such a barrier function in the case of<br />l_p-norm optimization (whose parameter, which is the main<br />determining factor in the algorithmic complexity of the method, is<br />proportional to the number of variables in the formulation and<br />independent from p) as well as in the case of the general<br />framework mentioned above.<br /><br />Finally, we contributed a survey of the self-concordancy property,<br />improving some useful results about the value of the complexity<br />parameter for certain categories of barrier functions and providing<br />some insight on the reason why the most commonly adopted definition<br />for self-concordant functions is the best possible.

In many everyday tasks it is necessary to compare the performance of the individuals in a population described by two or more criteria, for example comparing products in order to decide which is the best to purchase in terms of price and quality. Other examples are the comparison of universities, countries, the infrastructure in a telecommunications network, and the candidate solutions to a multi- or many-objective problem. In all of these cases, visualising the individuals better allows a decision maker to interpret their relative performance. This thesis explores methods for understanding and visualising multi- and many-criterion populations. Since people cannot generally comprehend more than three spatial dimensions the visualisation of many-criterion populations is a non-trivial task. We address this by generating visualisations based on the dominance relation which defines a structure in the population and we introduce two novel visualisation methods. The first method explicitly illustrates the dominance relationships between individuals as a graph in which individuals are sorted into Pareto shells, and is enhanced using many-criterion ranking methods to produce a finer ordering of individuals. We extend the power index, a method for ranking according to a single criterion, into the many-criterion domain by defining individual quality in terms of tournaments. The second visualisation method uses a new dominance-based distance in conjunction with multi-dimensional scaling, and we show that dominance can be used to identify an intuitive low-dimensional mapping of individuals, placing similar individuals close together. We demonstrate that this method can visualise a population comprising a large number of criteria. Heatmaps are another common method for presenting high-dimensional data, however they suffer from a drawback of being difficult to interpret if dissimilar individuals are placed close to each other. We apply spectral seriation to produce an ordering of individuals and criteria by which the heatmap is arranged, placing similar individuals and criteria close together. A basic version, computing similarity with the Euclidean distance, is demonstrated, before rank-based alternatives are investigated. The procedure is extended to seriate both the parameter and objective spaces of a multi-objective population in two stages. Since this process describes a trade-off, favouring the ordering of individuals in one space or the other, we demonstrate methods that enhance the visualisation by using an evolutionary optimiser to tune the orderings. One way of revealing the structure of a population is by highlighting which individuals are extreme. To this end, we provide three definitions of the “edge” of a multi-criterion mutually non-dominating population. All three of the definitions are in terms of dominance, and we show that one of them can be extended to cope with many-criterion populations. Because they can be difficult to visualise, it is often difficult for a decision maker to comprehend a population consisting of a large number of criteria. We therefore consider criterion selection methods to reduce the dimensionality with a view to preserving the structure of the population as quantified by its rank order. We investigate the efficacy of greedy, hill-climber and evolutionary algorithms and cast the dimension reduction as a multi-objective problem.

La modélisation et l’optimisation des procédés de polymérisation d’oléfines en prenant en compte les caractéristiques de la polymérisation et du procédé mis en jeu peuvent aider à améliorer les procédés de polymérisation industriels. Dans cette étude, un modèle a été développé et ses paramètres déterminés afin de calculer des propriétés thermodynamiques et physiques du système de polymérisation de l’éthylène. Des modèles en régime permanent et transitoire sont élaborés pour les procédés de polymérisation industriels en suspension ou en phase gazeuse, basés sur les cinétiques de polymérisation en présence de catalyseur Ziegler-Natta. La conduite des réacteurs a été analysée et le changement de grade simulé / Modeling and optimization of olefin polymerization processes based on polymerization and process characteristics provide guidance to plants and improve industrial processes. In this work, a model is proposed and its parameters determined to calculate thermodynamic and physical properties of the ethylene polymerization system. Based on industrial Zigler-Natta catalyzed multi-active sites ethylene polymerization kinetics, both steady and transient state plant-scale models are developed for industrial slurry and gas phase ethylene polymerization processes. The operating conditions of the reactors are analyzed; the grade transition and process optimization simulated

Ces dernières années, la phase de roulage au sol des avions a été mise en avant dans l'étude des retards aériens sur les grands aéroports. Cependant, le lien entre cette phase et l'optimisation des séquences d'avions sur les pistes reste encore peu étudié. L'objectif de réaliser des séquences optimales sur les pistes doit pourtant permettre de mieux gérer le trafic au sol, pour respecter les créneaux de décollage imposés tout en réduisant les retards des avions : dans cette thèse, un algorithme de calcul de séquences optimales est mis en place et intégré à la gestion du trafic au sol, modélisée comme un problème de résolution de conflits entre avions. Deux méthodes d'optimisation sont alors comparées : une méthode déterministe (utilisant un algorithme de type branch and bound) et une méthode stochastique (utilisant un algorithme génétique). Chacune des deux méthodes pouvant fonctionner avec et sans considération des séquences optimales sur les pistes. Les simulations effectuées montrent qu'une réduction significative des retards peut être espérée lorsque les séquences sont optimisées et anticipées. La méthode stochastique trouve de meilleures solutions, notamment en ce qui concerne la gestion des arrivées, mais la méthode déterministe reste intéressante, grâce à son temps de calcul bien plus rapide. / In the last few years, many studies concerning air traffic delays have focused on ground traffic management at busy airports. However, the link between the aircraft taxiing stage and runway scheduling optimisation is still rarely considered. Performing optimal aircraft sequences on runways should allow us to enhance the taxiing stage, while applying calculated take-off slots and reducing globally the aircraft mean delay. In this thesis, an algorithm is first defined to compute optimal aircraft schedules on runways. It is then integrated into the ground traffic management process, modeled as a conflict resolution problem between aircraft. A deterministic method (using a branch and bound algorithm) and a stochastic method (using a genetic algorithm) are both used to try and solve this problem. Each of these methods can work with and without the consideration of optimal runway scheduling. The simulations carried out show that the anticipation of the optimal runway schedules can yield a significant delay reduction for airport ground traffic. The stochastic method provides the best solutions, especially for arriving aircraft, while the deterministic method remains a considerable option because of its very fast running time.

L’importance des enjeux attachés à l’attractivité fiscale du territoire peut conduire certains Etats membres à exercer des discriminations ou à adopter un comportement protectionniste. Toutefois, l’absence d’harmonisation en matière de fiscalité directe au niveau européen ne doit pas être subie par le contribuable mais utilisée dans un but d’optimisation fiscale. En effet, ce phénomène résulte de la concurrence des législations fiscales et c’est la raison pour laquelle le droit communautaire interdit aux Etats membres de contrecarrer leurs effets en l’absence de fraude ou d’évasion fiscale. Le droit communautaire ne remet pas en cause la compétence des Etats membres pour délimiter leurs pouvoirs de taxation, mais l’exercice de ce pouvoir se heurte à la nécessité de respecter les libertés communautaires consacrées par le traité de Rome. Ainsi, l’obligation de conformité des dispositifs nationaux au droit communautaire est une garantie capitale pour le contribuable. / .

Fragment assembly methods represent the state of the art in computational protein structure prediction. However, one limitation of these methods, particularly for larger protein structures, is inadequate conformational sampling. This thesis describes studies aimed at uncovering potential causes of ineffective sampling, and the development of methods to try and address these problems. To identify behaviours that might lead to poor conformational sampling, we developed measures to study fragment-based sampling trajectories. Applying these measures to the Rosetta Abinitio and EdaFold methods showed similarities and differences in the ways that these methods make predictions, and pointed to common limitations. In both protocols, structural features such as alpha-helices were more frequently altered during the search, as compared with regions such as loops. Analyses of the fragment libraries used by these methods showed that fragments covering loop regions were less likely to possess native-like structural features, and this likely exacerbated the problems of inadequate sampling in these regions. Inadequate loop sampling leads to poor fold-level exploration within individual runs of methods such as Rosetta, and this necessitates the use of many independent runs. Guided by these findings, we developed new heuristic-based search algorithms. These algorithms were designed to facilitate the exploration of multiple energy basins within runs. Over many runs, the enhanced exploration in our protocols produced decoy sets with larger fractions of native-like solutions as compared to runs of Rosetta. Experiments with different fragment sets indicated that our methods could better translate increased fragment set quality into improvements in predictive accuracy distributions. These improvements depend most strongly on the ability of search algorithms to reliably generate native-like structures using a fragment set. In contrast, inadequate retention of native-like decoys when associated with unfavourable score values appears to be less of an issue. This thesis shows that targeted developments in conformational sampling strategies can improve the accuracy and reliability of predictions. With effective conformational sampling methods, developments in methods for fragment set construction and other areas may more reliably enhance predictive ability.

Evolutionary Structural Optimisation (ESO) is a relatively new design tool used to improve and optimise the design of structures. It is a heuristic method where a few elements of an initial design domain of finite elements are iteratively removed. Such a process is carried out repeatedly until an optimum design is achieved, or until a desired given area or volume is reached. There have been many contributions to the ESO procedure since its conception back in 1992. For example, a provision known as Bi-Directional ESO (BESO) has now been incorporated where elements may not only be removed, but added. Also, rather than deal with elements where they are either present or not, the designer now has the option to change the element's properties in a progressive fashion. This includes the modulus of elasticity, the density of the material and the thickness of plate elements, and is known as Morphing ESO. In addition to the algorithmic aspects of ESO, a large preference exists to optimise a structure based on a selection of criteria for various physical processes. Such examples include stress minimisation, buckling and electromagnetic problems. In a changing world that demands the enhancement of design tools and methods that incorporate optimisation, the development of methods like ESO to accommodate this demand is called for. It is this demand that this thesis seeks to satisfy. This thesis develops and examines the concept of multicriteria optimisation in the ESO process. Taking into account the optimisation of numerous criteria simultaneously, Multicriteria ESO allows a more realistic and accurate approach to optimising a model in any given environment. Two traditional methods � the Weighting method and the Global Criterion (Min-max) method have been used, as has two unconventional methods � the Logical AND method and the Logical OR method. These four methods have been examined for different combinations of Finite Element Analysis (FEA) solver types. This has included linear static FEA solver, the natural frequency FEA solver and a recently developed inertia FE solver. Mean compliance minimisation (stiffness maximisation), frequency maximisation and moment of inertia maximisation are an assortment of the specific objectives incorporated. Such a study has provided a platform to use many other criteria and multiple combinations of criteria. In extending the features of ESO, and hence its practical capabilities as a design tool, the creation of another optimisation method based on ESO has been ushered in. This method concerns the betterment of the bending and rotational performance of cross-sectional areas and is known as Evolutionary Moment of Inertia Optimisation (EMIO). Again founded upon a domain of finite elements, the EMIO method seeks to either minimise or maximise the rectangular, product and polar moments of inertia. This dissertation then goes one step further to include the EMIO method as one of the objectives considered in Multicriteria ESO as mentioned above. Most structures, (if not all) in reality are not homogenous as assumed by many structural optimisation methods. In fact, many structures (particularly biological ones) are composed of different materials or the same material with continually varying properties. In this thesis, a new feature called Constant Width Layer (CWL) ESO is developed, in which a distinct layer of material evolves with the developing boundary. During the optimisation process, the width of the outer surrounding material remains constant and is defined by the user. Finally, in verifying its usefulness to the practical aspect of design, the work presented herein applies the CWL ESO and the ESO methods to two dental case studies. They concern the optimisation of an anterior (front of the mouth) ceramic dental bridge and the optimisation of a posterior (back of the mouth) ceramic dental bridge. Comparisons of these optimised models are then made to those developed by other methods.

Le Benchmarking des systèmes client-serveur implique des infrastructures techniques réparties complexes, dont la gestion nécessite une approche autonomique. Cette gestion s'appuie sur une suite d'étapes, observation, analyse et rétroaction, qui correspond au principe d'une boucle de contrôle autonome. Des travaux antérieurs dans le domaine du test de performances ont montré comment introduire des fonctionnalités de test autonome par le biais d'une injection de charge auto-régulée. L'objectif de cette thèse est de suivre cette démarche de calcul autonome (autonomic computing) en y introduisant des fonctionnalités d'optimisation autonome. On peut ainsi obtenir automatiquement des résultats de benchmarks fiables et comparables, mettant en oeuvre l'ensemble des étapes de self-benchmarking. Notre contribution est double. D'une part, nous proposons un algorithme original pour l'optimisation dans un contexte de test de performance, qui vise à diminuer le nombre de solutions potentielles à tester, moyennant une hypothèse sur la forme de la fonction qui lie la valeur des paramètres à la performance mesurée. Cet algorithme est indépendant du système à optimiser. Il manipule des paramètres entiers, dont les valeurs sont comprises dans un intervalle donné, avec une granularité de valeur donnée. D'autre part, nous montrons une approche architecturale à composants et une organisation du benchmark automatique en plusieurs boucles de contrôle autonomes (détection de saturation, injection de charge, calcul d'optimisation), coordonnées de manière faiblement couplée via un mode de communication asynchrone de type publication-souscription. Complétant un canevas logiciel à composants pour l'injection de charge auto-régulée, nous y ajoutons des composants pour reparamétrer et redémarrer automatiquement le système à optimiser.Deux séries d'expérimentations ont été menées pour valider notre dispositif d'auto-optimisation. La première série concerne une application web de type achat en ligne, déployée sur un serveur d'application JavaEE. La seconde série concerne une application à trois tiers effectifs (WEB, métier (EJB JOnAS) et base de données) clusterSample. Les trois tiers sont sur des machines physiques distinctes.

Evolutionary Structural Optimisation (ESO) is a relatively new design tool used to improve and optimise the design of structures. It is a heuristic method where a few elements of an initial design domain of finite elements are iteratively removed. Such a process is carried out repeatedly until an optimum design is achieved, or until a desired given area or volume is reached. There have been many contributions to the ESO procedure since its conception back in 1992. For example, a provision known as Bi-Directional ESO (BESO) has now been incorporated where elements may not only be removed, but added. Also, rather than deal with elements where they are either present or not, the designer now has the option to change the element's properties in a progressive fashion. This includes the modulus of elasticity, the density of the material and the thickness of plate elements, and is known as Morphing ESO. In addition to the algorithmic aspects of ESO, a large preference exists to optimise a structure based on a selection of criteria for various physical processes. Such examples include stress minimisation, buckling and electromagnetic problems. In a changing world that demands the enhancement of design tools and methods that incorporate optimisation, the development of methods like ESO to accommodate this demand is called for. It is this demand that this thesis seeks to satisfy. This thesis develops and examines the concept of multicriteria optimisation in the ESO process. Taking into account the optimisation of numerous criteria simultaneously, Multicriteria ESO allows a more realistic and accurate approach to optimising a model in any given environment. Two traditional methods � the Weighting method and the Global Criterion (Min-max) method have been used, as has two unconventional methods � the Logical AND method and the Logical OR method. These four methods have been examined for different combinations of Finite Element Analysis (FEA) solver types. This has included linear static FEA solver, the natural frequency FEA solver and a recently developed inertia FE solver. Mean compliance minimisation (stiffness maximisation), frequency maximisation and moment of inertia maximisation are an assortment of the specific objectives incorporated. Such a study has provided a platform to use many other criteria and multiple combinations of criteria. In extending the features of ESO, and hence its practical capabilities as a design tool, the creation of another optimisation method based on ESO has been ushered in. This method concerns the betterment of the bending and rotational performance of cross-sectional areas and is known as Evolutionary Moment of Inertia Optimisation (EMIO). Again founded upon a domain of finite elements, the EMIO method seeks to either minimise or maximise the rectangular, product and polar moments of inertia. This dissertation then goes one step further to include the EMIO method as one of the objectives considered in Multicriteria ESO as mentioned above. Most structures, (if not all) in reality are not homogenous as assumed by many structural optimisation methods. In fact, many structures (particularly biological ones) are composed of different materials or the same material with continually varying properties. In this thesis, a new feature called Constant Width Layer (CWL) ESO is developed, in which a distinct layer of material evolves with the developing boundary. During the optimisation process, the width of the outer surrounding material remains constant and is defined by the user. Finally, in verifying its usefulness to the practical aspect of design, the work presented herein applies the CWL ESO and the ESO methods to two dental case studies. They concern the optimisation of an anterior (front of the mouth) ceramic dental bridge and the optimisation of a posterior (back of the mouth) ceramic dental bridge. Comparisons of these optimised models are then made to those developed by other methods.

Les méthodes d'optimisation à base de métamodèles avec enrichissement adaptatif (type Efficient Global Optimization) peuvent, malgré leurs atouts, être rédhibitoires en temps de calculs lorsqu'elles sont appliquées à des modèles numériques de grande taille avec plusieurs minimums locaux. Elles nécessitent la résolution d'un modèle complet à chaque simulation, ce qui peut conduire à des études irréalisables ou alors dans des durées incompatibles avec les échelles de temps typiques du processus de conception d'un produit. Cette thèse s'inscrit dans la thématique de l'optimisation de simulateurs numériquement couteux par l'utilisation des modèles simplifiés. Ces modèles simplifiés peuvent être notamment de deux types : des métamodèles ou des modèles d'ordre réduit. Nous proposons dans cette thèse une nouvelle méthodologie d'optimisation globale de systèmes mécaniques en couplant les méthodes d'optimisation à base de métamodèles adaptatifs avec les méthodes de réduction d'ordre. Les méthodes d'optimisation à base de métamodèles visent à réduire le nombre d'évaluations de la fonction objectif tandis que les méthodes de réduction d'ordre visent à diminuer la taille des modèles permettant ainsi une réduction de leur temps d'exécution. L'objectif de la méthodologie développée dans cette thèse est alors de réduire le nombre d'évaluations de la fonction objectif tout en diminuant également le temps d'exécution des modèles. L'idée de l'approche développée est de construire le métamodèle de la fonction objectif de manière adaptative. Cette construction fusionne des modèles complets et des modèles d'ordre réduit et adapte ainsi la fidélité aux besoins en précision de l'itération courante de l'optimisation. Les performances des algorithmes développés dans cette thèse ont été illustrées sur deux types d'applications : i. un problème d'identification des quatre propriétés orthotropes d'un composite stratifié à partir de mesures de champs de déplacement d'une plaque trouée en traction. ii. un problème de maximisation de la rigidité d'une plaque composite stratifiée. Les résultats ont permis de montrer que notre méthodologie permet d'obtenir des gains considérables, en temps de calcul, par rapport à un algorithme de type EGO classique. / The surrogate based optimization method with adaptive enrichment (Efficient Global Optimization type approach) may, in spite of its strengths, be prohibitive in terms of computational cost when applied to large scale problems with several local minima. They require the resolution of a full numerical model for each simulation, which can lead to intractable studies or to simulation times incompatible with the times allotted for the design of a product. This PhD thesis falls within the scope of optimizing expensive simulator codes by using substitution models of the simulator. These substitutions models can be of two types: a metamodel or a reduced order model. We have proposed here a new methodology for global optimization of mechanical systems by coupling adaptive surrogate based optimization methods with the reduced order modeling methods. The surrogate based optimization methods aim to reduce the number of objective function evaluations while the reduced order model methods aim to reduce the dimensionality of a model and thus its computational cost. The objective of the methodology proposed in this thesis is thus to reduce the number of the objective function evaluations while at the same time significantly reducing the computational expense to the resolutions of the full mechanical model. The basic idea of the proposed approach resides in the adaptive construction the metamodel of the objective function. This construction fuses full and reduced order models and thus adapts the model fidelity to the accuracy requirements of the optimization at the current iteration. The efficiency of our proposed algorithms was illustrated on two types of applications: i. a problem of identification of orthotropic elastic constants from full field displacement measurements based on a tensile test on a plate with a hole ii. a problem of stiffness maximization of laminated plates. The results have shown that our methodology provides a significant speed-up in terms of computational cost, compared to the traditional EGO algorithm.