• Refine Query
  • Source
  • Publication year
  • to
  • Language
  • 2
  • 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

Accelerating convex optimization in machine learning by leveraging functional growth conditions

Xu, Yi 01 August 2019 (has links)
In recent years, unprecedented growths in scale and dimensionality of data raise big computational challenges for traditional optimization algorithms; thus it becomes very important to develop efficient and effective optimization algorithms for solving numerous machine learning problems. Many traditional algorithms (e.g., gradient descent method) are black-box algorithms, which are simple to implement but ignore the underlying geometrical property of the objective function. Recent trend in accelerating these traditional black-box algorithms is to leverage geometrical properties of the objective function such as strong convexity. However, most existing methods rely too much on the knowledge of strong convexity, which makes them not applicable to problems without strong convexity or without knowledge of strong convexity. To bridge the gap between traditional black-box algorithms without knowing the problem's geometrical property and accelerated algorithms under strong convexity, how can we develop adaptive algorithms that can be adaptive to the objective function's underlying geometrical property? To answer this question, in this dissertation we focus on convex optimization problems and propose to explore an error bound condition that characterizes the functional growth condition of the objective function around a global minimum. Under this error bound condition, we develop algorithms that (1) can adapt to the problem's geometrical property to enjoy faster convergence in stochastic optimization; (2) can leverage the problem's structural regularizer to further improve the convergence speed; (3) can address both deterministic and stochastic optimization problems with explicit max-structural loss; (4) can leverage the objective function's smoothness property to improve the convergence rate for stochastic optimization. We first considered stochastic optimization problems with general stochastic loss. We proposed two accelerated stochastic subgradient (ASSG) methods with theoretical guarantees by iteratively solving the original problem approximately in a local region around a historical solution with the size of the local region gradually decreasing as the solution approaches the optimal set. Second, we developed a new theory of alternating direction method of multipliers (ADMM) with a new adaptive scheme of the penalty parameter for both deterministic and stochastic optimization problems with structured regularizers. With LEB condition, the proposed deterministic and stochastic ADMM enjoy improved iteration complexities of $\widetilde O(1/\epsilon^{1-\theta})$ and $\widetilde O(1/\epsilon^{2(1-\theta)})$ respectively. Then, we considered a family of optimization problems with an explicit max-structural loss. We developed a novel homotopy smoothing (HOPS) algorithm that employs Nesterov's smoothing technique and accelerated gradient method and runs in stages. Under a generic condition so-called local error bound (LEB) condition, it can improve the iteration complexity of $O(1/\epsilon)$ to $\widetilde O(1/\epsilon^{1-\theta})$ omitting a logarithmic factor with $\theta\in(0,1]$. Next, we proposed new restarted stochastic primal-dual (RSPD) algorithms for solving the problem with stochastic explicit max-structural loss. We successfully got a better iteration complexity than $O(1/\epsilon^2)$ without bilinear structure assumption, which is a big challenge of obtaining faster convergence for the considered problem. Finally, we consider finite-sum optimization problems with smooth loss and simple regularizer. We proposed novel techniques to automatically search for the unknown parameter on the fly of optimization while maintaining almost the same convergence rate as an oracle setting assuming the involved parameter is given. Under the Holderian error bound (HEB) condition with $\theta\in(0,1/2)$, the proposed algorithm also enjoys intermediate faster convergence rates than its standard counterparts with only the smoothness assumption.
2

Infeasibility detection and regularization strategies in nonlinear optimization / Détection de la non-réalisabilité et stratégies de régularisation en optimisation non linéaire

Tran, Ngoc Nguyen 26 October 2018 (has links)
Dans cette thèse, nous nous étudions des algorithmes d’optimisation non linéaire. D’une part nous proposons des techniques de détection rapide de la non-réalisabilité d’un problème à résoudre. D’autre part, nous analysons le comportement local des algorithmes pour la résolution de problèmes singuliers. Dans la première partie, nous présentons une modification d’un algorithme de lagrangien augmenté pour l’optimisation avec contraintes d’égalité. La convergence quadratique du nouvel algorithme dans le cas non-réalisable est démontrée théoriquement et numériquement. La seconde partie est dédiée à l’extension du résultat précédent aux problèmes d’optimisation non linéaire généraux avec contraintes d’égalité et d’inégalité. Nous proposons une modification d’un algorithme de pénalisation mixte basé sur un lagrangien augmenté et une barrière logarithmique. Les résultats théoriques de l’analyse de convergence et quelques tests numériques montrent l’avantage du nouvel algorithme dans la détection de la non-réalisabilité. La troisième partie est consacrée à étudier le comportement local d’un algorithme primal-dual de points intérieurs pour l’optimisation sous contraintes de borne. L’analyse locale est effectuée sans l’hypothèse classique des conditions suffisantes d’optimalité de second ordre. Celle-ci est remplacée par une hypothèse plus faible basée sur la notion de borne d’erreur locale. Nous proposons une technique de régularisation de la jacobienne du système d’optimalité à résoudre. Nous démontrons ensuite des propriétés de bornitude de l’inverse de ces matrices régularisées, ce qui nous permet de montrer la convergence superlinéaire de l’algorithme. La dernière partie est consacrée à l’analyse de convergence locale de l’algorithme primal-dual qui est utilisé dans les deux premières parties de la thèse. En pratique, il a été observé que cet algorithme converge rapidement même dans le cas où les contraintes ne vérifient l’hypothèse de qualification de Mangasarian-Fromovitz. Nous démontrons la convergence superlinéaire et quadratique de cet algorithme, sans hypothèse de qualification des contraintes. / This thesis is devoted to the study of numerical algorithms for nonlinear optimization. On the one hand, we propose new strategies for the rapid infeasibility detection. On the other hand, we analyze the local behavior of primal-dual algorithms for the solution of singular problems. In the first part, we present a modification of an augmented Lagrangian algorithm for equality constrained optimization. The quadratic convergence of the new algorithm in the infeasible case is theoretically and numerically demonstrated. The second part is dedicated to extending the previous result to the solution of general nonlinear optimization problems with equality and inequality constraints. We propose a modification of a mixed logarithmic barrier-augmented Lagrangian algorithm. The theoretical convergence results and the numerical experiments show the advantage of the new algorithm for the infeasibility detection. In the third part, we study the local behavior of a primal-dual interior point algorithm for bound constrained optimization. The local analysis is done without the standard assumption of the second-order sufficient optimality conditions. These conditions are replaced by a weaker assumption based on a local error bound condition. We propose a regularization technique of the Jacobian matrix of the optimality system. We then demonstrate some boundedness properties of the inverse of these regularized matrices, which allow us to prove the superlinear convergence of our algorithm. The last part is devoted to the local convergence analysis of the primal-dual algorithm used in the first two parts of this thesis. In practice, it has been observed that this algorithm converges rapidly even in the case where the constraints do not satisfy the Mangasarian-Fromovitz constraint qualification. We demonstrate the superlinear and quadratic convergence of this algorithm without any assumption of constraint qualification.

Page generated in 0.091 seconds