• Refine Query
  • Source
  • Publication year
  • to
  • Language
  • 1
  • 1
  • Tagged with
  • 2
  • 2
  • 2
  • 2
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • About
  • The Global ETD Search service is a free service for researchers to find electronic theses and dissertations. This service is provided by the Networked Digital Library of Theses and Dissertations.
    Our metadata is collected from universities around the world. If you manage a university/consortium/country archive and want to be added, details can be found on the NDLTD website.
1

Acceleration and new analysis of convex optimization algorithms

Liu, Lewis 07 1900 (has links)
Ces dernières années ont vu une résurgence de l’algorithme de Frank-Wolfe (FW) (également connu sous le nom de méthodes de gradient conditionnel) dans l’optimisation clairsemée et les problèmes d’apprentissage automatique à grande échelle avec des objectifs convexes lisses. Par rapport aux méthodes de gradient projeté ou proximal, une telle méthode sans projection permet d’économiser le coût de calcul des projections orthogonales sur l’ensemble de contraintes. Parallèlement, FW propose également des solutions à structure clairsemée. Malgré ces propriétés prometteuses, FW ne bénéficie pas des taux de convergence optimaux obtenus par les méthodes accélérées basées sur la projection. Nous menons une enquête dé- taillée sur les essais récents pour accélérer FW dans différents contextes et soulignons où se situe la difficulté lorsque l’on vise des taux linéaires globaux en théorie. En outre, nous fournissons une direction prometteuse pour accélérer FW sur des ensembles fortement convexes en utilisant des techniques d’intervalle de dualité et une nouvelle notion de régularité. D’autre part, l’algorithme FW est une covariante affine et bénéficie de taux de convergence accélérés lorsque l’ensemble de contraintes est fortement convexe. Cependant, ces résultats reposent sur des hypothèses dépendantes de la norme, entraînant généralement des bornes invariantes non affines, en contradiction avec la propriété de covariante affine de FW. Dans ce travail, nous introduisons de nouvelles hypothèses structurelles sur le problème (comme la régularité directionnelle) et dérivons une analyse affine invariante et indépendante de la norme de Frank-Wolfe. Sur la base de notre analyse, nous proposons une recherche par ligne affine invariante. Fait intéressant, nous montrons que les recherches en ligne classiques utilisant la régularité de la fonction objectif convergent étonnamment vers une taille de pas invariante affine, malgré l’utilisation de normes dépendantes de l’affine dans le calcul des tailles de pas. Cela indique que nous n’avons pas nécessairement besoin de connaître à l’avance la structure des ensembles pour profiter du taux accéléré affine-invariant. Dans un autre axe de recherche, nous étudions les algorithmes au-delà des méthodes du premier ordre. Les techniques Quasi-Newton approchent le pas de Newton en estimant le Hessien en utilisant les équations dites sécantes. Certaines de ces méthodes calculent le Hessien en utilisant plusieurs équations sécantes mais produisent des mises à jour non symétriques. D’autres schémas quasi-Newton, tels que BFGS, imposent la symétrie mais ne peuvent pas satisfaire plus d’une équation sécante. Nous proposons un nouveau type de mise à jour symétrique quasi-Newton utilisant plusieurs équations sécantes au sens des moindres carrés. Notre approche généralise et unifie la conception de mises à jour quasi-Newton et satisfait des garanties de robustesse prouvables. / Recent years have witnessed a resurgence of the Frank-Wolfe (FW) algorithm, also known as conditional gradient methods, in sparse optimization and large-scale machine learning problems with smooth convex objectives. Compared to projected or proximal gradient methods, such projection-free method saves the computational cost of orthogonal projections onto the constraint set. Meanwhile, FW also gives solutions with sparse structure. Despite of these promising properties, FW does not enjoy the optimal convergence rates achieved by projection-based accelerated methods. On the other hand, FW algorithm is affine-covariant, and enjoys accelerated convergence rates when the constraint set is strongly convex. However, these results rely on norm-dependent assumptions, usually incurring non-affine invariant bounds, in contradiction with FW’s affine-covariant property. In this work, we introduce new structural assumptions on the problem (such as the directional smoothness) and derive an affine in- variant, norm-independent analysis of Frank-Wolfe. Based on our analysis, we pro- pose an affine invariant backtracking line-search. Interestingly, we show that typical back-tracking line-search techniques using smoothness of the objective function surprisingly converge to an affine invariant stepsize, despite using affine-dependent norms in the computation of stepsizes. This indicates that we do not necessarily need to know the structure of sets in advance to enjoy the affine-invariant accelerated rate. Additionally, we provide a promising direction to accelerate FW over strongly convex sets using duality gap techniques and a new version of smoothness. In another line of research, we study algorithms beyond first-order methods. Quasi-Newton techniques approximate the Newton step by estimating the Hessian using the so-called secant equations. Some of these methods compute the Hessian using several secant equations but produce non-symmetric updates. Other quasi- Newton schemes, such as BFGS, enforce symmetry but cannot satisfy more than one secant equation. We propose a new type of quasi-Newton symmetric update using several secant equations in a least-squares sense. Our approach generalizes and unifies the design of quasi-Newton updates and satisfies provable robustness guarantees.
2

Aerodynamic Shape Design of Transonic Airfoils Using Hybrid Optimization Techniques and CFD

Xing, X.Q., Damodaran, Murali, Teo, Chung Piaw 01 1900 (has links)
This paper will analyze the effects of using hybrid optimization methods for optimizing objective functions that are determined by computational fluid dynamics solvers for compressible viscous flow for optimal design of airfoils. Previous studies on this topic by the authors had examined the application of deterministic optimization methods and stochastic optimization methods such as Simulated Annealing and Simultaneous Perturbation Stochastic Analysis (SPSA). The studies indicated that SPSA method has a greater or equal efficiency as compared with SA method in reaching optimal airfoil designs for the design problem in question. However, in some situations SPSA method has a tendency to demonstrate an oscillatory behavior in the vicinity of a local optima. To overcome this tendency, a hybrid method designed to take full advantage of SPSA’s high rate of reduction of the objective function at the inception of the design process to drive the design cycles towards the optimal zone at first, and then combining with other methods to perform the final stages of the convergence towards the optimal solutions is considered. SPSA method has been combined with the gradient-based Broydon-Fletcher-Goldfarb-Shanno (BFGS) method as well as Simulated Annealing method for the transonic inverse airfoil design problem that is concerned with the specification of a target airfoil surface pressure distribution and starting from an initial guess of an airfoil shape, the target airfoil shape is reached by way of minimization of a quantity that depends on the difference between the target and current airfoil surface pressure distribution. For a typical transonic flow test case, the effects of using hybrid optimization techniques such as SPSA+BFGS and SPSA+SA as opposed to using SPSA alone can be seen in Figure 1. After 800 design cycles using SPSA, the hybrid SPSA+SA method took 2521 function evaluations of SA while the SPSA+BFGS method took 271 function evaluations to reach similar values which are much better than that reached by using SPSA alone in the entire minimization process. Results indicate that both of the two hybrid methods have capability to find a global optimum more efficiently than the SPSA method. The paper will address issues related to hybridization and its impact on the optimal airfoil shape designs in various contexts. / Singapore-MIT Alliance (SMA)

Page generated in 0.0365 seconds