Cette thèse aborde la résolution de problèmes d'optimisation sous contraintes et systèmes de contraintes non linéaires sur les réels. Dans un premier temps, nous sommes intéressés à l'optimisation de problèmes dont les données sont connues de façon implicite (résultats d'une simulation informatique), en particulier, dans le cas des systèmes à événements discrets. Les contraintes de la simulation nécessitent de choisir judicieusement les méthodes d'optimisation adaptées à cette démarche de simulation-optimisation. Il est nécessaire d'utiliser des méthodes d'optimisation itératives où la fonction objectif est calculée point par point. Cette étude a donné naissance à l’environnement SimOpt qui consiste à coopérer entre un environnement de simulation et un environnement d’optimisation mathématique constitué d’un ensemble de méthodes d’analyse numérique et de recherche opérationnelle. Ensuite, Nous avons étudié les problèmes d'optimisation globale non linéaires continus basés sur la satisfaction de contraintes et l'arithmétique des intervalles. Dans ce cas, le résultat du calcul est un intervalle qui contient les solutions. Notre contribution consiste d'abord à classifier les techniques de résolution et établir les différentes relations de décomposition et transformation d'un problème d'optimisation dans un processus de résolution. Finalement, la dernière partie est consacrée à la résolution de systèmes de contraintes non linéaires. Nous avons proposé une méthode basée sur une approche symbolique numérique qui permet de limiter le problème de localité des raisonnements dans le processus de résolution basé sur les techniques de consistance locale.

Les travaux qui composent cette thèse s'articulent autour des thèmes suivants : Optimisation globale, Analyse et optimisation non différentiables, Analyse convexe.

LES METHODES DE GRADIENTS. UN ALGORITHME DE POINTS REALISABLES POUR L'OPTIMISATION AVEC CONTRAINTES NON LINEAIRES. LES METRIQUES VARIABLES ET L'OPTIMISATION AVEC CONTRAINTES

The generalized area of multiple criteria decision making (MCDM) can be defined as the body of methods and procedures by which the concern for multiple conflicting criteria can be formally incorporated into the analytical process. MCDM consists mostly of two branches, multiple criteria optimization and multi-criteria decision analysis (MCDA). While MCDA is typically concerned with multiple criteria problems that have a small number of alternatives often in an environment of uncertainty (location of an airport, type of drug rehabilitation program), multiple criteria optimization is typically directed at problems formulated within a mathematical programming framework, but with a stack of objectives instead of just one (river basin management, engineering component design, product distribution). It is about the most modern treatment of multiple criteria optimization that this book is concerned. I look at this book as a nicely organized and well-rounded presentation of what I view as ”new wave” topics in multiple criteria optimization. Looking back to the origins of MCDM, most people agree that it was not until about the early 1970s that multiple criteria optimization c- gealed as a field. At this time, and for about the following fifteen years, the focus was on theories of multiple objective linear programming that subsume conventional (single criterion) linear programming, algorithms for characterizing the efficient set, theoretical vector-maximum dev- opments, and interactive procedures.

This book presents solutions to the general problem of single period portfolio optimization. It introduces different linear models, arising from different performance measures, and the mixed integer linear models resulting from the introduction of real features. Other linear models, such as models for portfolio rebalancing and index tracking, are also covered. The book discusses computational issues and provides a theoretical framework, including the concepts of risk-averse preferences, stochastic dominance and coherent risk measures. The material is presented in a style that requires no background in finance or in portfolio optimization; some experience in linear and mixed integer models, however, is required. The book is thoroughly didactic, supplementing the concepts with comments and illustrative examples.

Computational science can loosely be defined as the endeavor to develop and analyse models for the simulation and control of complex processes. This is achieved by making optimal use of computer resources and by drawing upon a variety of disciplines, techniques and theories. With the emergence of advanced computational and information technologies, computational science now has a significant impact on engineering, physical, biological, management and social sciences. The contributions in this state-of-the-art volume range from theoretical and numerical topics to practical implementations. The subject matter includes modelling, mathematical and numerical analysis, differential equations, linear algebra, optimization, domain decomposition, computational fluid dynamics, computational mechanics, elasticity, structure, computational physics and chemistry, electromagnetic, control theory and other applications. This volume is dedicated to Roland Glowinski on the occasion of his 60th birthday. It is aimed at the next generation of scientists, applied mathematicians, computer scientists, practitioners and engineers who will define computational science within the context of the challenging scientific, industrial, economic and societal problems of the 21st century.