Advances in scaling deep learning algorithms

Les algorithmes d'apprentissage profond forment un nouvel ensemble de méthodes puissantes pour l'apprentissage automatique. L'idée est de combiner des couches de facteurs latents en hierarchies. Cela requiert souvent un coût computationel plus elevé et augmente aussi le nombre de paramètres du modèle. Ainsi, l'utilisation de ces méthodes sur des problèmes à plus grande échelle demande de réduire leur coût et aussi d'améliorer leur régularisation et leur optimization. Cette thèse adresse cette question sur ces trois perspectives. Nous étudions tout d'abord le problème de réduire le coût de certains algorithmes profonds. Nous proposons deux méthodes pour entrainer des machines de Boltzmann restreintes et des auto-encodeurs débruitants sur des distributions sparses à haute dimension. Ceci est important pour l'application de ces algorithmes pour le traitement de langues naturelles. Ces deux méthodes (Dauphin et al., 2011; Dauphin and Bengio, 2013) utilisent l'échantillonage par importance pour échantilloner l'objectif de ces modèles. Nous observons que cela réduit significativement le temps d'entrainement. L'accéleration atteint 2 ordres de magnitude sur plusieurs bancs d'essai. Deuxièmement, nous introduisont un puissant régularisateur pour les méthodes profondes. Les résultats expérimentaux démontrent qu'un bon régularisateur est crucial pour obtenir de bonnes performances avec des gros réseaux (Hinton et al., 2012). Dans Rifai et al. (2011), nous proposons un nouveau régularisateur qui combine l'apprentissage non-supervisé et la propagation de tangente (Simard et al., 1992). Cette méthode exploite des principes géometriques et permit au moment de la publication d'atteindre des résultats à l'état de l'art. Finalement, nous considérons le problème d'optimiser des surfaces non-convexes à haute dimensionalité comme celle des réseaux de neurones. Tradionellement, l'abondance de minimum locaux était considéré comme la principale difficulté dans ces problèmes. Dans Dauphin et al. (2014a) nous argumentons à partir de résultats en statistique physique, de la théorie des matrices aléatoires, de la théorie des réseaux de neurones et à partir de résultats expérimentaux qu'une difficulté plus profonde provient de la prolifération de points-selle. Dans ce papier nous proposons aussi une nouvelle méthode pour l'optimisation non-convexe.

An application of a modified constructive heuristic algorithm to transmission expansion planning

Transmission expansion planning (TEP) is a non-convex optimization problem that can be solved via different heuristic algorithms. A variety of classical as well as heuristic algorithms in literature are addressed to solve TEP problem. In this paper a modified constructive heuristic algorithm (CHA) is proposed for solving such a crucial problem. Most of research papers handle TEP problem by linearization of the non-linear mathematical model while in this research TEP problem is solved via CHA using non-linear model. The proposed methodology is based upon Garver's algorithm capable of applying to a DC model. Simulation studies and tests results on the well known transmission network such as: Garver and IEEE 24-bus systems are carried out to show the significant performance as well as the effectiveness of the proposed algorithm. © 2011 IEEE.

An application of CHA to concurrent short-term transmission expansion & reactive power planning

In this paper a heuristic technique for solving simultaneous short-term transmission network expansion and reactive power planning problem (TEPRPP) via an AC model is presented. A constructive heuristic algorithm (CHA) aimed to obtaining a significant quality solution for such problem is employed. An interior point method (IPM) is applied to solve TEPRPP as a nonlinear programming (NLP) during the solution steps of the algorithm. For each proposed network topology, an indicator is deployed to identify the weak buses for reactive power sources placement. The objective function of NLP includes the costs of new transmission lines, real power losses as well as reactive power sources. By allocating reactive power sources at load buses, the circuit capacity may increase while the cost of new lines can be decreased. The proposed methodology is tested on Garver's system and the obtained results shows its capability and the viability of using AC model for solving such non-convex optimization problem. © 2011 IEEE.

Algoritmo tabu search especializado para o problema de planejamento da expansao de sistemas de transmissão

Pós-graduação em Engenharia Elétrica - FEIS

Low rank subspace clustering (LRSC)

We consider the problem of fitting a union of subspaces to a collection of data points drawn from one or more subspaces and corrupted by noise and/or gross errors. We pose this problem as a non-convex optimization problem, where the goal is to decompose the corrupted data matrix as the sum of a clean and self-expressive dictionary plus a matrix of noise and/or gross errors. By self-expressive we mean a dictionary whose atoms can be expressed as linear combinations of themselves with low-rank coefficients. In the case of noisy data, our key contribution is to show that this non-convex matrix decomposition problem can be solved in closed form from the SVD of the noisy data matrix. The solution involves a novel polynomial thresholding operator on the singular values of the data matrix, which requires minimal shrinkage. For one subspace, a particular case of our framework leads to classical PCA, which requires no shrinkage. For multiple subspaces, the low-rank coefficients obtained by our framework can be used to construct a data affinity matrix from which the clustering of the data according to the subspaces can be obtained by spectral clustering. In the case of data corrupted by gross errors, we solve the problem using an alternating minimization approach, which combines our polynomial thresholding operator with the more traditional shrinkage-thresholding operator. Experiments on motion segmentation and face clustering show that our framework performs on par with state-of-the-art techniques at a reduced computational cost.

Steady-state global optimization of metabolic non-linear dynamic models through recasting into power-law canonical models

Background: Design of newly engineered microbial strains for biotechnological purposes would greatly benefit from the development of realistic mathematical models for the processes to be optimized. Such models can then be analyzed and, with the development and application of appropriate optimization techniques, one could identify the modifications that need to be made to the organism in order to achieve the desired biotechnological goal. As appropriate models to perform such an analysis are necessarily non-linear and typically non-convex, finding their global optimum is a challenging task. Canonical modeling techniques, such as Generalized Mass Action (GMA) models based on the power-law formalism, offer a possible solution to this problem because they have a mathematical structure that enables the development of specific algorithms for global optimization. Results: Based on the GMA canonical representation, we have developed in previous works a highly efficient optimization algorithm and a set of related strategies for understanding the evolution of adaptive responses in cellular metabolism. Here, we explore the possibility of recasting kinetic non-linear models into an equivalent GMA model, so that global optimization on the recast GMA model can be performed. With this technique, optimization is greatly facilitated and the results are transposable to the original non-linear problem. This procedure is straightforward for a particular class of non-linear models known as Saturable and Cooperative (SC) models that extend the power-law formalism to deal with saturation and cooperativity. Conclusions: Our results show that recasting non-linear kinetic models into GMA models is indeed an appropriate strategy that helps overcoming some of the numerical difficulties that arise during the global optimization task.

Dynamic, convex, and robust optimization with Bayesian learning for response-guided dosing

Thesis (Ph.D.)--University of Washington, 2016-08

Sensing Cloud Optimization to Solve ED of Units with Valve-Point Effects and Multi-fuels

In this paper a solution to an highly constrained and non-convex economical dispatch (ED) problem with a meta-heuristic technique named Sensing Cloud Optimization (SCO) is presented. The proposed meta-heuristic is based on a cloud of particles whose central point represents the objective function value and the remaining particles act as sensors "to fill" the search space and "guide" the central particle so it moves into the best direction. To demonstrate its performance, a case study with multi-fuel units and valve- point effects is presented.

Direct search optimization application programming interface with remote access

Search Optimization methods are needed to solve optimization problems where the objective function and/or constraints functions might be non differentiable, non convex or might not be possible to determine its analytical expressions either due to its complexity or its cost (monetary, computational, time,...). Many optimization problems in engineering and other fields have these characteristics, because functions values can result from experimental or simulation processes, can be modelled by functions with complex expressions or by noise functions and it is impossible or very difficult to calculate their derivatives. Direct Search Optimization methods only use function values and do not need any derivatives or approximations of them. In this work we present a Java API that including several methods and algorithms, that do not use derivatives, to solve constrained and unconstrained optimization problems. Traditional API access, by installing it on the developer and/or user computer, and remote API access to it, using Web Services, are also presented. Remote access to the API has the advantage of always allow the access to the latest version of the API. For users that simply want to have a tool to solve Nonlinear Optimization Problems and do not want to integrate these methods in applications, also two applications were developed. One is a standalone Java application and the other a Web-based application, both using the developed API.

Revisiting optimization algorithms for maximum likelihood estimation

Parmi les méthodes d’estimation de paramètres de loi de probabilité en statistique, le maximum de vraisemblance est une des techniques les plus populaires, comme, sous des conditions l´egères, les estimateurs ainsi produits sont consistants et asymptotiquement efficaces. Les problèmes de maximum de vraisemblance peuvent être traités comme des problèmes de programmation non linéaires, éventuellement non convexe, pour lesquels deux grandes classes de méthodes de résolution sont les techniques de région de confiance et les méthodes de recherche linéaire. En outre, il est possible d’exploiter la structure de ces problèmes pour tenter d’accélerer la convergence de ces méthodes, sous certaines hypothèses. Dans ce travail, nous revisitons certaines approches classiques ou récemment d´eveloppées en optimisation non linéaire, dans le contexte particulier de l’estimation de maximum de vraisemblance. Nous développons également de nouveaux algorithmes pour résoudre ce problème, reconsidérant différentes techniques d’approximation de hessiens, et proposons de nouvelles méthodes de calcul de pas, en particulier dans le cadre des algorithmes de recherche linéaire. Il s’agit notamment d’algorithmes nous permettant de changer d’approximation de hessien et d’adapter la longueur du pas dans une direction de recherche fixée. Finalement, nous évaluons l’efficacité numérique des méthodes proposées dans le cadre de l’estimation de modèles de choix discrets, en particulier les modèles logit mélangés.

Multi-objective optimization scheme for static and dynamic multicast flows

Muchas de las nuevas aplicaciones emergentes de Internet tales como TV sobre Internet, Radio sobre Internet,Video Streamming multi-punto, entre otras, necesitan los siguientes requerimientos de recursos: ancho de banda consumido, retardo extremo-a-extremo, tasa de paquetes perdidos, etc. Por lo anterior, es necesario formular una propuesta que especifique y provea para este tipo de aplicaciones los recursos necesarios para su buen funcionamiento. En esta tesis, proponemos un esquema de ingeniería de tráfico multi-objetivo a través del uso de diferentes árboles de distribución para muchos flujos multicast. En este caso, estamos usando la aproximación de múltiples caminos para cada nodo egreso y de esta forma obtener la aproximación de múltiples árboles y a través de esta forma crear diferentes árboles multicast. Sin embargo, nuestra propuesta resuelve la fracción de la división del tráfico a través de múltiples árboles. La propuesta puede ser aplicada en redes MPLS estableciendo rutas explícitas en eventos multicast. En primera instancia, el objetivo es combinar los siguientes objetivos ponderados dentro de una métrica agregada: máxima utilización de los enlaces, cantidad de saltos, el ancho de banda total consumido y el retardo total extremo-a-extremo. Nosotros hemos formulado esta función multi-objetivo (modelo MHDB-S) y los resultados obtenidos muestran que varios objetivos ponderados son reducidos y la máxima utilización de los enlaces es minimizada. El problema es NP-duro, por lo tanto, un algoritmo es propuesto para optimizar los diferentes objetivos. El comportamiento que obtuvimos usando este algoritmo es similar al que obtuvimos con el modelo. Normalmente, durante la transmisión multicast los nodos egresos pueden salir o entrar del árbol y por esta razón en esta tesis proponemos un esquema de ingeniería de tráfico multi-objetivo usando diferentes árboles para grupos multicast dinámicos. (en el cual los nodos egresos pueden cambiar durante el tiempo de vida de la conexión). Si un árbol multicast es recomputado desde el principio, esto podría consumir un tiempo considerable de CPU y además todas las comuicaciones que están usando el árbol multicast serán temporalmente interrumpida. Para aliviar estos inconvenientes, proponemos un modelo de optimización (modelo dinámico MHDB-D) que utilice los árboles multicast previamente computados (modelo estático MHDB-S) adicionando nuevos nodos egreso. Usando el método de la suma ponderada para resolver el modelo analítico, no necesariamente es correcto, porque es posible tener un espacio de solución no convexo y por esta razón algunas soluciones pueden no ser encontradas. Adicionalmente, otros tipos de objetivos fueron encontrados en diferentes trabajos de investigación. Por las razones mencionadas anteriormente, un nuevo modelo llamado GMM es propuesto y para dar solución a este problema un nuevo algoritmo usando Algoritmos Evolutivos Multi-Objetivos es propuesto. Este algoritmo esta inspirado por el algoritmo Strength Pareto Evolutionary Algorithm (SPEA). Para dar una solución al caso dinámico con este modelo generalizado, nosotros hemos propuesto un nuevo modelo dinámico y una solución computacional usando Breadth First Search (BFS) probabilístico. Finalmente, para evaluar nuestro esquema de optimización propuesto, ejecutamos diferentes pruebas y simulaciones. Las principales contribuciones de esta tesis son la taxonomía, los modelos de optimización multi-objetivo para los casos estático y dinámico en transmisiones multicast (MHDB-S y MHDB-D), los algoritmos para dar solución computacional a los modelos. Finalmente, los modelos generalizados también para los casos estático y dinámico (GMM y GMM Dinámico) y las propuestas computacionales para dar slución usando MOEA y BFS probabilístico.

Non-existence of competitive equilibria with dynamically inconsistent preferences

This paper shows the robust non-existence of competitive equilibria even in a simple three period representative agent economy with dynamically inconsistent preferences. We distinguish between a sophisticated and naive representative agent. Even when underlying preferences are monotone and convex, at given prices, we show by example that the induced preference of the sophisticated representative agent over choices in first-period markets is both non-convex and satiated. Even allowing for negative prices, the market-clearing allocation is not contained in the convex hull of demand. Finally, with a naive representative agent, we show that perfect foresight is incompatible with market clearing and individual optimization at given prices.

Lower-semicontinuity and optimization of convex functionals

The result that we treat in this article allows to the utilization of classic tools of convex analysis in the study of optimality conditions in the optimal control convex process for a Volterra-Stietjes linear integral equation in the Banach space G([a, b],X) of the regulated functions in [a, b], that is, the functions f : [a, 6] → X that have only descontinuity of first kind, in Dushnik (or interior) sense, and with an equality linear restriction. In this work we introduce a convex functional Lβf(x) of Nemytskii type, and we present conditions for its lower-semicontinuity. As consequence, Weierstrass Theorem garantees (under compacity conditions) the existence of solution to the problem min{Lβf(x)}. © 2009 Academic Publications.

Mathematical programming model for heat exchanger design through optimization of partial objectives

Mathematical programming can be used for the optimal design of shell-and-tube heat exchangers (STHEs). This paper proposes a mixed integer non-linear programming (MINLP) model for the design of STHEs, following rigorously the standards of the Tubular Exchanger Manufacturers Association (TEMA). Bell–Delaware Method is used for the shell-side calculations. This approach produces a large and non-convex model that cannot be solved to global optimality with the current state of the art solvers. Notwithstanding, it is proposed to perform a sequential optimization approach of partial objective targets through the division of the problem into sets of related equations that are easier to solve. For each one of these problems a heuristic objective function is selected based on the physical behavior of the problem. The global optimal solution of the original problem cannot be ensured even in the case in which each of the sub-problems is solved to global optimality, but at least a very good solution is always guaranteed. Three cases extracted from the literature were studied. The results showed that in all cases the values obtained using the proposed MINLP model containing multiple objective functions improved the values presented in the literature.

Primal Attainment in Convex Infinite Optimization Duality

This article provides results guarateeing that the optimal value of a given convex infinite optimization problem and its corresponding surrogate Lagrangian dual coincide and the primal optimal value is attainable. The conditions ensuring converse strong Lagrangian (in short, minsup) duality involve the weakly-inf-(locally) compactness of suitable functions and the linearity or relative closedness of some sets depending on the data. Applications are given to different areas of convex optimization, including an extension of the Clark-Duffin Theorem for ordinary convex programs.