Archive 2018-2019

Comité local d’organisation


 

SPOT 63– Lundi 1 Juillet 2019 –  Amphi des thèses ENSEEIHT

14h – Bachir El Khadir – Princeton Univ, USA

Power and Limitations of Sum of Squares Programming for Stability Analysis of Dynamical Systems

We study the power and limitations of sum of squares optimization and semialgebraic Lyapunov functions for proving asymptotic stability of polynomial dynamical systems. We give the first example of a globally asymptotically stable polynomial vector field with rational coefficients that does not admit a polynomial (or even analytic) Lyapunov function in any neighborhood of the origin. We show, however, that if the polynomial vector field is homogeneous, then its asymptotic stability is equivalent to existence of a rational Lyapunov function whose inequalities have sum of squares proofs. This statement generalizes the classical result in control on the equivalence between asymptotic stability of linear systems and existence of a quadratic Lyapunov function satisfying a certain linear matrix inequality.

15h – Tillmann Weisser – Los Alamos Nat Lab, USA

Relaxations and Uncertainty Quantification for the Power Grid

Optimal scheduling power generation such that the energy needs in the grid are met everywhere and at all time is a challenging problem. In particular due to the significant increase of renewable energies in the last years, the search for a global optimal operating point has become more difficult because the generation capacities change over time and often are not entirely predictable. In this talk I address this problem in two ways: First, I will talk about relaxations of the AC Power Flow problem which are used to detect whether locally found solutions are globally optimal. In my current research we are trying to strengthen these relaxations without increasing their computational cost to an unacceptable level.  In the second half, I propose a strategy to avoid solving expensive robust or stochastic programs online by precomputing deterministic approximations.

SPOT 62 – Lundi 3 Juin 2019 –  Amphi des thèses ENSEEIHT

14h – Lorick Huang (IMT) – Quelques contributions autour des algorithmes de Robbins Monro

Les algorithmes de Robbins-Monro sont des procédures récursives basés sur des simulations qui permettent d’approcher le zéro d’une fonction pouvant s’écrire comme une espérance. Dans cet exposé, je présenterai deux résultats originaux autours de la convergence de ces algorithmes.
Sans trop rentrer dans les détails, je présenterai ces algorithmes et mes résultats; un premier résultat de nature numérique sur l’extrapolation de Richardson Romberg et un second de nature plus théorique sur un théorème limite local.

15h – Yacine Chitour (Paris Saclay/L2S) – Aspects dynamiques des algorithmes de gradient pour l’apprentissage
Dans cet exposé, nous présentons un travail en cours avec Z. Liao et R. Couillet. Nous proposons de décrire le comportement empirique de certains algorithmes de gradient pour l’apprentissage sous la forme d’une conjecture dite  »overfitting conjecture », qui adopte le point de vue des systèmes dynamiques déterministes. Pour l’apprentissage linéaire profond, nous montrons la convergence des trajectoires vers des points critiques et la résolution de la conjecture dans le cas d’un système avec une couche. Une série de problèmes ouverts est proposée.


SPOT 61 – Lundi 13 Mai 2019 –  Amphi des thèses ENSEEIHT

14h – Jérôme Fehrenbach (IMT) – Shape optimization using Gauss-Newton method

Shape optimization aims at determining the shape of a domain where the solution of a partial differential equation is optimal with respect to some objective. In general only local optima can be found. We will overview classical methods, among them the method of choice is a descent using first order information. When the objective is a quadratic cost function we will define a Gauss-Newton method to solve the problem, and we will document this method by showing numerical results. The oscillations shown by the classical shape gradient method are removed and the convergence to a local minimizer is faster.

15h – Jean-Baptiste Hiriart-Urruty (IMT) – Three (small) examples of problems of « optimal curves » which are still open

The so-called open problems, the conjectures … have always been driving forces in the advance of knowledge in mathematics, but also in their learning, at all levels : colleges, high schools, mathematical training in universities. To propose such questions, in the form of challenges “Who will do the best ?” is even a suggested activity in the teaching methods because it is very formative. Added to this is the need for a demonstration or proof of an advanced assertion : to propose the best known solution does not mean that there is no better one ; a solution is the best when it has been proved that it is … Number Theory like Geometry are major problem areas containing problems for which we can propose the best known solutions, but without knowing, for as much, if they are the best possible, in other words without being able to prove the optimality of these solutions (in a sense to be defined).

We propose in this talk three problems of geometry (with variants), with a variatio- nal flavor, of the type “Find the curve of minimum length such as …” ; the first one is somehow classical, the next two are more original. In all three cases, we indicate how far mathematicians have been able to go; none of them has been fully resolved, as yet. Of course, we do not expect the listener to solve these problems … But what we propose can be used as material for a research process, between students, colleagues, on “how to bring and/or improve” a proposed answer to a given problem, each one making his imagination and mathematical ability work.


 

SPOT 60 – Lundi 1 Avril 2019 –  Amphi des thèses ENSEEIHT
14h – Hedy Attouch (Univ. Montpellier) – Fast optimization through inertial dynamics with Hessian driven damping
In a Hilbert space H, we first study the convergence properties, when t→+∞, of the trajectories t → x(t) ∈ H generated by the second-order evolution equation:
edoThen we analyze the fast minimization properties of the algorithms obtained as discrete versions in time. The function f: H → R to be minimized is assumed to be convex, twice continuously differentiable, with non-empty solution set. Extensions to structured optimization problems involving non-smooth functions will be examined next. The parameters α and β are supposed to be non-negative. The above system combines two types of damping with specific properties:
  1. The isotropic viscous damping with vanishing coefficient α/t is related to the Nesterov accelerated gradient method and to the FISTA method, as recently shown by Su-Boyd-Candès. We will review recent results regarding the fast convergence properties of these methods. They depend on the value of α relative to the critical value α =3. For α>3, any trajectory converges weakly to a minimizer of f, and f(x(t))- min(f) = o(1/t2). We will also show the natural link with the Ravine method of Gelfand-Tsetlin (1961).
  2. The Hessian driven damping takes into account the geometry of f. It is linked to the Newton and the Levenberg-Marquardt methods. It was introduced in the context of optimization by Alvarez-Attouch-Bolte-Redont (J. Math. Pures et Appl. 2002). Since then, several variants and associated algorithms have been considered. The above formulation has been studied by Attouch-Peypouquet-Redont (JDE 2016). The introduction of the Hessian induces multiple favorable effects on convergence properties. While maintaining the convergence rate of values of the Nesterov-type methods, it provides the fast convergence to zero of the gradients ∇f (x(t)).  For poorly conditioned minimization problems, it neutralizes wild transverse oscillations.

Surprisingly, the presence of the Hessian driven damping makes the system well-posed for a general proper lower-semicontinuous convex function f : H → R∪{+∞}. This is based on the crucial property that the inertial dynamic with Hessian driven damping can be written equivalently as a first-order system in time and space, allowing it to be extended simply by replacing the gradient with the sub-differential. The extension to the case of the sum of two potential functions « non-smooth + smooth » was exploited by Attouch-Maingé-Redont (DEA 2012) to model non-elastic shocks in mechanics and PDE’s. The study of the associated algorithms for structured convex optimization is an ongoing research topic. We will present a new proximal-based inertial algorithm with fast convergence properties combining Nesterov acceleration with Hessian damping. Based on the dynamical interpretation of the algorithms, we show the effect of temporal scaling on the convergence rates. The study of the dynamic system above has natural connection with control theory. Passing from open-loop control (as in Nesterov’s method) to closed-loop control is an active research topic. For Newton’s regularized method (without inertia), it was examined by Attouch-Redont-Svaiter (JOTA 2013). This study in the context above, as well as the restarting method, is an interesting direction of research.

15h – Emmanuel Soubiès (IRIT-CNRS Toulouse) – Sparse Optimization through Exact Continuous Relaxations of the L2-L0 Functional
Several non-convex continuous relaxations of the l0 pseudo-norm have been proposed and analyzed in the literature. In this talk, considering the l0-regularized least-squares criteria (l2-l0), we will present a way to compare such relaxations from the perspective of their fidelity to the initial l2-l0 problem. More specifically, we will exhibit necessary and sufficient conditions on separable approximations of the l0 pseudo-norm that ensure the two following properties:
a) Global minimizers of the relaxation coincide with those of the l2-l0 criteria;
b) Each local minimizer of the relaxation is a local minimizer of the l2-l0 criteria.
From these conditions, we obtain a class of penalties said to be exact regarding to the above properties. Moreover, the second property (b) suggests that such exact relaxations admit potentially less local (non-global) minimizers than the initial l2-l0 functional. This will be quantified both numerically and theoretically. Finally, we will focus on the inferior limit of this class of relaxations which has special properties and is the one that removes the largest number of local (non-global) minimizers of the initial criteria.
Joint work with Laure Blanc-Féraud and Gilles Aubert

SPOT 59 – Lundi 11 Mars 2019 –  Amphi des thèses ENSEEIHT
14h – Philippe Toint (Univ. Namur) – Recent results in worst-case evaluation complexity for smooth and non-smooth, exact and inexact, nonconvex optimization
We present a review of the results obtained during the last year on the worst-case complexity of minimization algorithm for nonconvex problems using potentially high-degree models. In the smooth (Lipschitz or Hölder) case, global complexity bound will be presented that are valid for any model’s degree and any order of optimality, thereby generalizing all known results for first- and second-order methods. The bound states that an adaptive regularization algorithm using derivatives up to degree p will produce an epsilon -approximate q-th order minimizer in at most O(epsilon^{(p+1)/(p-q+1)}) evaluations. Moreover, these results are shown to be sharp. We will also show how to extend these optimal-complexity results to the case where the problem’s objective function and derivatives are computed inexactly, yielding a bound of O(|log(epsilon)|epsilon^{(p+1)/(p-q+1)}) approximate evaluations. Co-authored with C. Cartis, N. Gould, S. Bellavia, G. Gurioli, B. Morini
15h – Elisa Riccetti (IRIT-ENSEEIHT Toulouse) – High-order multilevel optimization strategies and their application to the training of ANNs
Standard iterative optimization methods are based on the approximation of the objective function by a model, given by a truncated Taylor series. Models of order two are classically used. Recently, in the literature a unifying framework has been proposed, to extend the existing theory also to models of higher order. The use of such models comes along with higher costs. We propose a multilevel extension of such methods, to reduce the major cost per iteration, represented by the model minimization. The proposed methods rely on the knowledge of a sequence of approximations to the original objective function, defined on spaces of reduced dimension and cheaper to optimize. We also investigate the application of such techniques to problems where the variables are not related by geometrical constraints. We choose as an important representative of this class the training of artificial neural networks.

SPOT 58 – Lundi 4 Février 2019 –  Amphi des thèses ENSEEIHT
14h – Gersende Fort (IMT CNRS Toulouse)

Convergence de méthodes de gradient stochastiques à biais persistant

Les algorithmes d’Approximation Stochastique, dont l’algorithme de gradient à pas décroissant est un exemple, sont des procédures itératives pour la recherche de zero d’une fonction non explicite. Ils reposent sur une approximation Monte Carlo de cette fonction objectif. Néanmoins, dans beaucoup d’applications, cette approximation stochastique est biaisée avec un biais qui, pour des raisons de coût de calcul, ne disparaît pas au fil des itérations. La persistance de ce biais pose la question du bien-fondé de telles procédures stochastiques.

Dans cet exposé, nous motiverons ce contexte par des applications en Apprentissage Statistique, et illustrerons les résultats théoriques par une extension de l’algorithme de gradient stochastique à l’optimisation de fonctions convexes composites. Nous démontrerons la convergence de ces algorithmes et de versions accélérées, même dans le cas d’un biais persistant. Nous commenterons aussi le parallèle entre ces méthodes et des algorithmes Expectation-Maximization stochastiques (Monte Carlo EM, Stochastic Approximation EM).

15h – Adrien Taylor (INRIA ENS Paris)

Computer-assisted analyses of first-order methods (via semidefinite programming)

We present the performance estimation approach. The methodology aims at automatically analyzing worst-case convergence properties of first-order methods for solving (mostly convex) optimization problems. In particular, it allows obtaining (tight) guarantees for fixed-step first-order methods involving a variety of oracles – that includes explicit, projected, proximal, conditional, inexact, or stochastic (sub)gradient steps – and a variety of convergence measures. This presentation aims at providing a high-level introduction on the current state of the framework, mostly through examples. The talk is based on joint works with great collaborators: François Glineur (UCLouvain), Julien Hendrickx (UCLouvain), Etienne de Klerk (Tilburg/TU Delft), Yoel Drori (Google), Pontus Giselsson (Lund), Carolina Bergeling (Lund), Ernest Ryu (UCLA) and Francis Bach (Inria/ENS Paris). The methodology is implemented within the package « PESTO » (for « Performance EStimation TOolbox », avaible at https://github.com/AdrienTaylor/Performance-Estimation-Toolbox), which allows using the framework without the SDP modelling steps.


SPOT 57 – Lundi 7 Janvier 2019 –  Amphi des thèses ENSEEIHT
 14h – Hristo Sendov (University of Western Ontario) – Polar convexity and critical points of polynomials

A set A, in the extended complex plane, is called convex with respect to a pole u, if for any x,y in A the arc on the unique circle through x,y and u, that connects x and y but does not contain u, is in A. If the pole u is taken at infinity, then this notion reduces to the usual convexity. Polar convexity is connected with the classical Gauss-Lucas’ and Laguerre’s theorems for complex polynomials. If a set is convex with respect to u and contains the zeros of a polynomial, then it contains the zeros of its polar derivative with respect to u. A set may be convex with respect to more than one pole. The main goal of this talk is to present the main relationships between sets and their poles.

This is a joint work with Blagovest Sendov from the Bulgarian Academy of Sciences.

15h – Charles Dossal (INSA and IMT Toulouse) – Exact decay rate for Nesterov Acceleration
 
 In this presentation, we will give new and optimal decay rates for the Nesterov Acceleration scheme of classical gradient descent depending on the local geometry of the function to minimize. Only bounds on the rates are known for convex or strongly convex functions and we will give a more precise and complete description of this rates using Lojasievicz and flatness conditions.
We will explain how these decays can be obtained studying an ODE.
This work is a Joint work with V. Apidopoulos, JF. Aujol and A. Rondepierre.

SPOT 56 – Lundi 3 décembre – Amphi A001 ENSEEIHT

14h – Gérard Dupont et Jayant Sen Gupta (AIRBUS) – Data science and AI at AIRBUS central research
Within AIRBUS Central Research & Technology, the XRD team tackle high value and long-term opportunities to use data analysis and artificial intelligence on AIRBUS engineering challenges. From optimizing flight routes or satellite plan to detecting anomalies in sensors time-series, 2018 has seen the start of a series of new projects which will be presented from the industrial point of view but also highlighting the research difficulties that are targeted.We will dive into the ADVISED project dealing with anomaly detection in large amount of time-series data from sensors. Anticipation unseen failures on a fleet of aircraft or satellite, automatic detection of faulty sensor without prior knowledge are among the industrial objectives. On the research part, the project tries to combine best approaches allowing unsupervised learning with time series data.
15h – Stéphane Canu (INSA Rouen) – Variable selection and outlier detection as a MIP
Dimension reduction or feature selection is an effective strategy to handle contaminated data and to deal with high dimensionality while providing better prediction. Also, outliers can have a considerable bad influence on estimators. Outliers are understood to be observations that have been corrupted, incorrectly measured, mis-recorded, drawn under different conditions than those intended, or so atypical as to require separate modeling. To deal with outlier proneness and spurious variables, we propose a method performing the outright rejection of discordant observations together with the selection of relevant variables. A reformulation of this problem as a mixed integer program allows  the use of an effective commercial solver to solve it. In addition, a proximal relaxation provides  a suboptimal initial solution at a lower computational cost,  accelerating the overall optimization process.

SPOT 55. Lundi 1 Octobre 2018. Lieu : N7, salle des thèses

14h – Salma Kuhlmann (Univ Constance, Allemagne) - Positive polynomials and moment problems

Hilbert’s 17th problem asked whether a real polynomial p(x1,..,xn) which takes non-negative values as a function on R^n is a finite sum of squares (SOS) of real rational functions q(x1,..,xn)/r(x1,..,xn). A complete positive answer was provided by Artin and Schreier (1927), giving birth to real algebraic geometry. The question when the (SOS) representation is denominator free is however of particular interest for applications. In his pioneering 1888 paper, Hilbert gave a general answer (in terms of degree and number of variables). Subsequent general results, such as Krivine’s Positivstellensatz, pertain to a relative situation, where one considers polynomials non-negative on a basic closed semi-algebraic set K and SOSs weighted with inequalities defining K. Stronger results hold when K is compact; the Archimedean Positivstellensatz of Putinar and Jacobi-Prestel is a fundamental tool
in theory and applications. By the classical Riesz-Haviland theorem (1930s), the problem of characterizing positive polynomials on a given closed subset K of R^n is dual to the finite dimensional moment problem (i.e. that of representing a linear functional on the polynomial algebra R[x1,..,xn] as integration with respect to a Borel measure). An algebraic approach was taken in a series of papers by Ghasemi-Kuhlmann-Marshall (2013-2016) who study the moment problem on a general not necessarily finitely generated commutative unital real algebra, a context adapted to infinite dimensional moment problems. In this talk I will survey (with examples) various Positivstellensaetze and their corresponding moment problem interpretations.

15h – Jean-Claude Yakoubsohn (Univ Paul Sabatier Toulouse) – Certified and Fast Computation of SVD

This work is in progress with the collaboration  Joris Van Der Hoeven of the Laboratory LIX.  We define and study a new  method to approximate locally the SVD of a general  matrix: this method  has  a local quadratic convergence.  This solves the problem of the fast certified approximation of the multiple singular values or clusters of singular values.  We also give result of complexity of this approximation using such a type of homotopy method.


 

Juin 2018 : Le séminaire SPOT est terminé pour l’année 2017-2018. De juin à septembre 2018 a lieu le semestre thématique Optimisation financé par le Labex CIMI. Plus de détails à l’adresse : http://cimi.univ-toulouse.fr/optimisation/

Evènements :

Juin 4 – 5, 2018 Master Class sur les Méthodes hybrides pour l’Optimisation Combinatoire / Mixte
Juin 27 – 29, 2018 Workshop Optimisation: aspects fondamentaux et exploitation de la structure.
Septembre 3 – 7, 2018 Spring School Contrôle optimal numérique.
Septembre 10 – 13, 2018 Workshop Optimization and Learning.