Seminar 2015.2
August 14, 2015 · 16:00–17:00
On the complexity of the proximal gradient method for convex minimization
José Yunier Bello Cruz · IME/UFG
In this talk we present the convergence and complexity analysis of the iterates in the proximal gradient method with a linesearch. When the stepsizes generated by the linesearch are bounded below by a positive number, our analysis shows that the expected error from the cost value at the $k$-th iteration to the optimal value is $\mathcal{O}(k^{-1})$ in Hilbert spaces and $o(k^{-1})$ in finite dimensions, which improves the complexity of the first-order algorithm presented in the literature. It is worth emphasizing that the global Lipschitz continuity assumption on the gradient of $f$ is sufficient but not necessary for the boundedness from below of the stepsizes aforementioned. Moreover, we show that if the gradient of $f$ is locally Lipchiptz the stepsizes generated by the linesearch are bounded below by a positive number. Furthermore, we aswer the main question here: "Can we have the complexity $o(k^{-1})$ when $\liminf_{k\to \infty}\alpha_k=0$?" with an example.
August 21, 2015 · 16:00–17:00
Inexact Newton's method to nonlinear functions with values in a cone
Gilson do Nascimento Silva · PhD student, IME/UFG
The problem of finding a solution of nonlinear inclusion problems in Banach space is considered. Using convex optimization techniques introduced by Robinson (Numer. Math., Vol. 19, 1972, pp. 341-347), a robust convergence theorem for inexact Newton's method is proved. As an application, an affine invariant version of Kantorovich's theorem and Smale's $\alpha$-theorem for inexact Newton's method is obtained.
August 28, 2015 · 10:00–11:00 · canceled
The linearized proximal algorithm for convex composite optimization with applications
Chong Li · Zhejiang University, China
We propose a linearized proximal algorithm (LPA) to solve a convex composite optimization problem. Each iteration of the LPA is a proximal minimization on the composition of the outer function and the linearization of the inner function at current iterate. The LPA has the attractive computational advantage in that the solution of each subproblem is a singleton, which avoids the difficulty of finding the whole solution set of the subproblem, as in the Gauss-Newton method (GNM), while it still maintains the same local convergence rate as that of the GNM. Under the assumptions of local weak sharp minima of order $p$ ($p \in [1,2]$) and the quasi-regularity condition, we establish the local superlinear convergence rate for the LPA. We also propose a globalization strategy for LPA based on the backtracking line-search and the inexact LPA, as well as the superlinear convergence results. We further apply the LPA to solve a feasibility problem, as well as a sensor network localization problem. Our numerical results illustrate that the LPA meets the demand for a robust and efficient algorithm for the sensor network localization problem.
September 4, 2015 · 16:00–17:00
Proximal Point Method for Vector Optimization on Hadamard Manifolds
Yuri Rafael Leite Pereira · PhD student, IME/UFG
In this talk, we extend the proximal point algorithm in vector optimization to the context in Riemannian Manifolds, by assuming an iterative process which uses a variable nonlinear scalarization function. Then, we show that any sequence generated by this new algorithm reach a weakly e cient point after a .nite number of iterations under the assumption that the weakly e cient point set is weak sharp for the vectorial problem.
September 11, 2015 · 16:00–17:00
Korpelevich's method for variational inequality problems on Hadamard manifolds - Parte 1
Edvaldo Elias Almeida Batista · PhD student, IME/UFG
A variant of Korpelevich?s method for solving the variational inequality problem is extended from Euclidean spaces to constant curvature Hadamard manifolds. Under a pseudomonotone assumption on the underlying vector ?eld, we prove that the sequence generated by the method converges to a solution of variational inequality, whenever it exists.
September 18, 2015 · 16:00–17:00
Método de Newton para encontrar zeros de uma classe especial de funções semi-suaves
Maurício Silva Louzeiro · Master's student, IME/UFG
Neste seminário farei uso do método de Newton para resolver algumas equações semi-suaves especiais. De início, veremos o problema de minimizar uma função convexa sobre um cone simplicial, problema este que equivale a resolver uma determinada equação semi-suave. Particularmente veremos o caso em que essa função é quadrática. Por fim, serão estudadas algumas equações lineares por partes.
September 25, 2015 · 16:00–17:00
Dual Descent Methods and Inexact Proximal Algorithms
Valdinês Leite de Sousa Júnior · PhD student, IME/UFG
Neste seminário apresentaremos um método abstrato em espaços quase-métricos. Analisaremos sua convergência e provaremos a taxa de convergência do método.
October 2, 2015 · 16:00–17:00
A fast multistep proximal forward-backward method with Linesearch 1
Paulo César da Silva Júnior · PhD student, IME/UFG
We present a Method 2, a faster version of the proximal forward-backward method with Linesearch 1, improving the convergence result for Method 1. Here we modify the method by adding a linesearch and an extra projection step.
October 23, 2015 · 16:00–17:00
An extragradient-type algorithm for non-smooth variational inequalities
Edvaldo Elias Almeida Batista · PhD student, IME/UFG
We study extragradient-type methods to solve variational inequality problems involving maxima! monotone point-to-set npcratorz. Wc will show, xing an example, that, in this case, to achieve convergence to the solution set of sequences generated by extragradient methods it is necessary, in the search of the auxiliary point and its suitable image by the involved operator, to enlarge the set of possible search directions. Such an enlargement is known and we use it to overcome this failure of the extragradient-methods.
November 6, 2015 · 16:00–17:00
Trust-Region Method on Riemannian Manifolds
Yuri Rafael Leite Pereira · PhD student, IME/UFG
In this talk, we make a introduction of the trust-region method and we the extend this concept to Riemannian manifolds environment and make the convergence analysis.
November 13, 2015 · 16:00–17:00
Algoritmo Genético Semidifuso. Aplicação de um modelo epidemiologico em Cuba
Aymee Marrero Severo · Universidad de La Habana, Cuba
Apresentar uma alternativa para tratar a incerteza em parâmetros para resolver o problema de estimativa de parâmetros num sistema dinâmico definido por equações diferenciais ordinárias (e.d.o.), como uma variação da Algoritmo Genético fuzzy. Propone uma representação fuzzy de indivíduos na população e operadores de cruzamento e mutação também difusa, melhorando diversidade populacional do AG para que os problemas de convergência prematuros são superados, com a manipulação menos complicado Sistema de Inferência difusa. Aplicação a um modelo epidemiológico.
November 20, 2015 · 16:00–17:00
A variante de Barzilai-Borwein do Método Gradiente
Ábssan Matuzinhos de Moura · Master's student, IME/UFG
Considerando uma função f: Rn --> R quadrática, estritamente convexa e x* o único minimizador de f, mostraremos que o Método Gradiente com o tamanho do passo dado pela variante de Barzilai-Borwein, converge para x*. Lembrando que para o caso bidimensional foi provado que a convergência acontece, e, além disso, com taxa R-superlinear.
November 27, 2015 · 16:00–17:00
Local Newton's method for generalized equations under majorante condition
Gilson do Nascimento Silva · PhD student, IME/UFG
The Newton?s method under the majorant condition for solving generalized equations will be presented. Examples are given to emphasize the application of this method to generalized equations representing the nonlinear programming problem and the nonlinear complementary problem.