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.

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.

Stability of Random Sums and Extremes

This study is about the stability of random sums and extremes.The difficulty in finding exact sampling distributions resulted in considerable problems of computing probabilities concerning the sums that involve a large number of terms.Functions of sample observations that are natural interest other than the sum,are the extremes,that is , the minimum and the maximum of the observations.Extreme value distributions also arise in problems like the study of size effect on material strengths,the reliability of parallel and series systems made up of large number of components,record values and assessing the levels of air pollution.It may be noticed that the theories of sums and extremes are mutually connected.For instance,in the search for asymptotic normality of sums ,it is assumed that at least the variance of the population is finite.In such cases the contributions of the extremes to the sum of independent and identically distributed(i.i.d) r.vs is negligible.

Analysis and Optimization of Machining Process Using Evolutionary Algorithms

To ensure quality of machined products at minimum machining costs and maximum machining effectiveness, it is very important to select optimum parameters when metal cutting machine tools are employed. Traditionally, the experience of the operator plays a major role in the selection of optimum metal cutting conditions. However, attaining optimum values each time by even a skilled operator is difficult. The non-linear nature of the machining process has compelled engineers to search for more effective methods to attain optimization. The design objective preceding most engineering design activities is simply to minimize the cost of production or to maximize the production efficiency. The main aim of research work reported here is to build robust optimization algorithms by exploiting ideas that nature has to offer from its backyard and using it to solve real world optimization problems in manufacturing processes.In this thesis, after conducting an exhaustive literature review, several optimization techniques used in various manufacturing processes have been identified. The selection of optimal cutting parameters, like depth of cut, feed and speed is a very important issue for every machining process. Experiments have been designed using Taguchi technique and dry turning of SS420 has been performed on Kirlosker turn master 35 lathe. Analysis using S/N and ANOVA were performed to find the optimum level and percentage of contribution of each parameter. By using S/N analysis the optimum machining parameters from the experimentation is obtained.Optimization algorithms begin with one or more design solutions supplied by the user and then iteratively check new design solutions, relative search spaces in order to achieve the true optimum solution. A mathematical model has been developed using response surface analysis for surface roughness and the model was validated using published results from literature.Methodologies in optimization such as Simulated annealing (SA), Particle Swarm Optimization (PSO), Conventional Genetic Algorithm (CGA) and Improved Genetic Algorithm (IGA) are applied to optimize machining parameters while dry turning of SS420 material. All the above algorithms were tested for their efficiency, robustness and accuracy and observe how they often outperform conventional optimization method applied to difficult real world problems. The SA, PSO, CGA and IGA codes were developed using MATLAB. For each evolutionary algorithmic method, optimum cutting conditions are provided to achieve better surface finish.The computational results using SA clearly demonstrated that the proposed solution procedure is quite capable in solving such complicated problems effectively and efficiently. Particle Swarm Optimization (PSO) is a relatively recent heuristic search method whose mechanics are inspired by the swarming or collaborative behavior of biological populations. From the results it has been observed that PSO provides better results and also more computationally efficient.Based on the results obtained using CGA and IGA for the optimization of machining process, the proposed IGA provides better results than the conventional GA. The improved genetic algorithm incorporating a stochastic crossover technique and an artificial initial population scheme is developed to provide a faster search mechanism. Finally, a comparison among these algorithms were made for the specific example of dry turning of SS 420 material and arriving at optimum machining parameters of feed, cutting speed, depth of cut and tool nose radius for minimum surface roughness as the criterion. To summarize, the research work fills in conspicuous gaps between research prototypes and industry requirements, by simulating evolutionary procedures seen in nature that optimize its own systems.

Mean Squared Residue Based Biclustering Algorithms for the Analysis of Gene Expression Data

Computational Biology is the research are that contributes to the analysis of biological data through the development of algorithms which will address significant research problems.The data from molecular biology includes DNA,RNA ,Protein and Gene expression data.Gene Expression Data provides the expression level of genes under different conditions.Gene expression is the process of transcribing the DNA sequence of a gene into mRNA sequences which in turn are later translated into proteins.The number of copies of mRNA produced is called the expression level of a gene.Gene expression data is organized in the form of a matrix. Rows in the matrix represent genes and columns in the matrix represent experimental conditions.Experimental conditions can be different tissue types or time points.Entries in the gene expression matrix are real values.Through the analysis of gene expression data it is possible to determine the behavioral patterns of genes such as similarity of their behavior,nature of their interaction,their respective contribution to the same pathways and so on. Similar expression patterns are exhibited by the genes participating in the same biological process.These patterns have immense relevance and application in bioinformatics and clinical research.Theses patterns are used in the medical domain for aid in more accurate diagnosis,prognosis,treatment planning.drug discovery and protein network analysis.To identify various patterns from gene expression data,data mining techniques are essential.Clustering is an important data mining technique for the analysis of gene expression data.To overcome the problems associated with clustering,biclustering is introduced.Biclustering refers to simultaneous clustering of both rows and columns of a data matrix. Clustering is a global whereas biclustering is a local model.Discovering local expression patterns is essential for identfying many genetic pathways that are not apparent otherwise.It is therefore necessary to move beyond the clustering paradigm towards developing approaches which are capable of discovering local patterns in gene expression data.A biclusters is a submatrix of the gene expression data matrix.The rows and columns in the submatrix need not be contiguous as in the gene expression data matrix.Biclusters are not disjoint.Computation of biclusters is costly because one will have to consider all the combinations of columans and rows in order to find out all the biclusters.The search space for the biclustering problem is 2 m+n where m and n are the number of genes and conditions respectively.Usually m+n is more than 3000.The biclustering problem is NP-hard.Biclustering is a powerful analytical tool for the biologist.The research reported in this thesis addresses the problem of biclustering.Ten algorithms are developed for the identification of coherent biclusters from gene expression data.All these algorithms are making use of a measure called mean squared residue to search for biclusters.The objective here is to identify the biclusters of maximum size with the mean squared residue lower than a given threshold. All these algorithms begin the search from tightly coregulated submatrices called the seeds.These seeds are generated by K-Means clustering algorithm.The algorithms developed can be classified as constraint based,greedy and metaheuristic.Constarint based algorithms uses one or more of the various constaints namely the MSR threshold and the MSR difference threshold.The greedy approach makes a locally optimal choice at each stage with the objective of finding the global optimum.In metaheuristic approaches particle Swarm Optimization(PSO) and variants of Greedy Randomized Adaptive Search Procedure(GRASP) are used for the identification of biclusters.These algorithms are implemented on the Yeast and Lymphoma datasets.Biologically relevant and statistically significant biclusters are identified by all these algorithms which are validated by Gene Ontology database.All these algorithms are compared with some other biclustering algorithms.Algorithms developed in this work overcome some of the problems associated with the already existing algorithms.With the help of some of the algorithms which are developed in this work biclusters with very high row variance,which is higher than the row variance of any other algorithm using mean squared residue, are identified from both Yeast and Lymphoma data sets.Such biclusters which make significant change in the expression level are highly relevant biologically.

Multi-objective assembly job shop scheduling using genetic algorithm and tabu search

Assembly job shop scheduling problem (AJSP) is one of the most complicated combinatorial optimization problem that involves simultaneously scheduling the processing and assembly operations of complex structured products. The problem becomes even more complicated if a combination of two or more optimization criteria is considered. This thesis addresses an assembly job shop scheduling problem with multiple objectives. The objectives considered are to simultaneously minimizing makespan and total tardiness. In this thesis, two approaches viz., weighted approach and Pareto approach are used for solving the problem. However, it is quite difficult to achieve an optimal solution to this problem with traditional optimization approaches owing to the high computational complexity. Two metaheuristic techniques namely, genetic algorithm and tabu search are investigated in this thesis for solving the multiobjective assembly job shop scheduling problems. Three algorithms based on the two metaheuristic techniques for weighted approach and Pareto approach are proposed for the multi-objective assembly job shop scheduling problem (MOAJSP). A new pairing mechanism is developed for crossover operation in genetic algorithm which leads to improved solutions and faster convergence. The performances of the proposed algorithms are evaluated through a set of test problems and the results are reported. The results reveal that the proposed algorithms based on weighted approach are feasible and effective for solving MOAJSP instances according to the weight assigned to each objective criterion and the proposed algorithms based on Pareto approach are capable of producing a number of good Pareto optimal scheduling plans for MOAJSP instances.

Global Optimization Algorithms and their Application to Distributed Systems

In this report, we discuss the application of global optimization and Evolutionary Computation to distributed systems. We therefore selected and classified many publications, giving an insight into the wide variety of optimization problems which arise in distributed systems. Some interesting approaches from different areas will be discussed in greater detail with the use of illustrative examples.

Evolving Distributed Algorithms with Genetic Programming

Distributed systems are one of the most vital components of the economy. The most prominent example is probably the internet, a constituent element of our knowledge society. During the recent years, the number of novel network types has steadily increased. Amongst others, sensor networks, distributed systems composed of tiny computational devices with scarce resources, have emerged. The further development and heterogeneous connection of such systems imposes new requirements on the software development process. Mobile and wireless networks, for instance, have to organize themselves autonomously and must be able to react to changes in the environment and to failing nodes alike. Researching new approaches for the design of distributed algorithms may lead to methods with which these requirements can be met efficiently. In this thesis, one such method is developed, tested, and discussed in respect of its practical utility. Our new design approach for distributed algorithms is based on Genetic Programming, a member of the family of evolutionary algorithms. Evolutionary algorithms are metaheuristic optimization methods which copy principles from natural evolution. They use a population of solution candidates which they try to refine step by step in order to attain optimal values for predefined objective functions. The synthesis of an algorithm with our approach starts with an analysis step in which the wanted global behavior of the distributed system is specified. From this specification, objective functions are derived which steer a Genetic Programming process where the solution candidates are distributed programs. The objective functions rate how close these programs approximate the goal behavior in multiple randomized network simulations. The evolutionary process step by step selects the most promising solution candidates and modifies and combines them with mutation and crossover operators. This way, a description of the global behavior of a distributed system is translated automatically to programs which, if executed locally on the nodes of the system, exhibit this behavior. In our work, we test six different ways for representing distributed programs, comprising adaptations and extensions of well-known Genetic Programming methods (SGP, eSGP, and LGP), one bio-inspired approach (Fraglets), and two new program representations called Rule-based Genetic Programming (RBGP, eRBGP) designed by us. We breed programs in these representations for three well-known example problems in distributed systems: election algorithms, the distributed mutual exclusion at a critical section, and the distributed computation of the greatest common divisor of a set of numbers. Synthesizing distributed programs the evolutionary way does not necessarily lead to the envisaged results. In a detailed analysis, we discuss the problematic features which make this form of Genetic Programming particularly hard. The two Rule-based Genetic Programming approaches have been developed especially in order to mitigate these difficulties. In our experiments, at least one of them (eRBGP) turned out to be a very efficient approach and in most cases, was superior to the other representations.

Optimierung von Betriebsstrategien für elektrifizierte Antriebskonzepte

Im Rahmen dieser Arbeit wird eine gemeinsame Optimierung der Hybrid-Betriebsstrategie und des Verhaltens des Verbrennungsmotors vorgestellt. Die Übernahme von den im Steuergerät verwendeten Funktionsmodulen in die Simulationsumgebung für Fahrzeuglängsdynamik stellt eine effiziente Applikationsmöglichkeit der Originalparametrierung dar. Gleichzeitig ist es notwendig, das Verhalten des Verbrennungsmotors derart nachzubilden, dass das stationäre und das dynamische Verhalten, inklusive aller relevanten Einflussmöglichkeiten, wiedergegeben werden kann. Das entwickelte Werkzeug zur Übertragung der in Ascet definierten Steurgerätefunktionen in die Simulink-Simulationsumgebung ermöglicht nicht nur die Simulation der relevanten Funktionsmodule, sondern es erfüllt auch weitere wichtige Eigenschaften. Eine erhöhte Flexibilität bezüglich der Daten- und Funktionsstandänderungen, sowie die Parametrierbarkeit der Funktionsmodule sind Verbesserungen die an dieser Stelle zu nennen sind. Bei der Modellierung des stationären Systemverhaltens des Verbrennungsmotors erfolgt der Einsatz von künstlichen neuronalen Netzen. Die Auswahl der optimalen Neuronenanzahl erfolgt durch die Betrachtung des SSE für die Trainings- und die Verifikationsdaten. Falls notwendig, wird zur Sicherstellung der angestrebten Modellqualität, das Interpolationsverhalten durch Hinzunahme von Gauß-Prozess-Modellen verbessert. Mit den Gauß-Prozess-Modellen werden hierbei zusätzliche Stützpunkte erzeugt und mit einer verminderten Priorität in die Modellierung eingebunden. Für die Modellierung des dynamischen Systemverhaltens werden lineare Übertragungsfunktionen verwendet. Bei der Minimierung der Abweichung zwischen dem Modellausgang und den Messergebnissen wird zusätzlich zum SSE das 2σ-Intervall der relativen Fehlerverteilung betrachtet. Die Implementierung der Steuergerätefunktionsmodule und der erstellten Steller-Sensor-Streckenmodelle in der Simulationsumgebung für Fahrzeuglängsdynamik führt zum Anstieg der Simulationszeit und einer Vergrößerung des Parameterraums. Das aus Regelungstechnik bekannte Verfahren der Gütevektoroptimierung trägt entscheidend zu einer systematischen Betrachtung und Optimierung der Zielgrößen bei. Das Ergebnis des Verfahrens ist durch das Optimum der Paretofront der einzelnen Entwurfsspezifikationen gekennzeichnet. Die steigenden Simulationszeiten benachteiligen Minimumsuchverfahren, die eine Vielzahl an Iterationen benötigen. Um die Verwendung einer Zufallsvariablen, die maßgeblich zur Steigerung der Iterationanzahl beiträgt, zu vermeiden und gleichzeitig eine Globalisierung der Suche im Parameterraum zu ermöglichen wird die entwickelte Methode DelaunaySearch eingesetzt. Im Gegensatz zu den bekannten Algorithmen, wie die Partikelschwarmoptimierung oder die evolutionären Algorithmen, setzt die neu entwickelte Methode bei der Suche nach dem Minimum einer Kostenfunktion auf eine systematische Analyse der durchgeführten Simulationsergebnisse. Mit Hilfe der bei der Analyse gewonnenen Informationen werden Bereiche mit den bestmöglichen Voraussetzungen für ein Minimum identifiziert. Somit verzichtet das iterative Verfahren bei der Bestimmung des nächsten Iterationsschrittes auf die Verwendung einer Zufallsvariable. Als Ergebnis der Berechnungen steht ein gut gewählter Startwert für eine lokale Optimierung zur Verfügung. Aufbauend auf der Simulation der Fahrzeuglängsdynamik, der Steuergerätefunktionen und der Steller-Sensor-Streckenmodelle in einer Simulationsumgebung wird die Hybrid-Betriebsstrategie gemeinsam mit der Steuerung des Verbrennungsmotors optimiert. Mit der Entwicklung und Implementierung einer neuen Funktion wird weiterhin die Verbindung zwischen der Betriebsstrategie und der Motorsteuerung erweitert. Die vorgestellten Werkzeuge ermöglichten hierbei nicht nur einen Test der neuen Funktionalitäten, sondern auch eine Abschätzung der Verbesserungspotentiale beim Verbrauch und Abgasemissionen. Insgesamt konnte eine effiziente Testumgebung für eine gemeinsame Optimierung der Betriebsstrategie und des Verbrennungsmotorverhaltens eines Hybridfahrzeugs realisiert werden.

Reinforcement Learning by Policy Search

One objective of artificial intelligence is to model the behavior of an intelligent agent interacting with its environment. The environment's transformations can be modeled as a Markov chain, whose state is partially observable to the agent and affected by its actions; such processes are known as partially observable Markov decision processes (POMDPs). While the environment's dynamics are assumed to obey certain rules, the agent does not know them and must learn. In this dissertation we focus on the agent's adaptation as captured by the reinforcement learning framework. This means learning a policy---a mapping of observations into actions---based on feedback from the environment. The learning can be viewed as browsing a set of policies while evaluating them by trial through interaction with the environment. The set of policies is constrained by the architecture of the agent's controller. POMDPs require a controller to have a memory. We investigate controllers with memory, including controllers with external memory, finite state controllers and distributed controllers for multi-agent systems. For these various controllers we work out the details of the algorithms which learn by ascending the gradient of expected cumulative reinforcement. Building on statistical learning theory and experiment design theory, a policy evaluation algorithm is developed for the case of experience re-use. We address the question of sufficient experience for uniform convergence of policy evaluation and obtain sample complexity bounds for various estimators. Finally, we demonstrate the performance of the proposed algorithms on several domains, the most complex of which is simulated adaptive packet routing in a telecommunication network.

Towards Direct Policy Search Reinforcement Learning for Robot Control

This paper proposes a high-level reinforcement learning (RL) control system for solving the action selection problem of an autonomous robot. Although the dominant approach, when using RL, has been to apply value function based algorithms, the system here detailed is characterized by the use of direct policy search methods. Rather than approximating a value function, these methodologies approximate a policy using an independent function approximator with its own parameters, trying to maximize the future expected reward. The policy based algorithm presented in this paper is used for learning the internal state/action mapping of a behavior. In this preliminary work, we demonstrate its feasibility with simulated experiments using the underwater robot GARBI in a target reaching task

Implementing a generic systolic array for genetic algorithms

We have designed a highly parallel design for a simple genetic algorithm using a pipeline of systolic arrays. The systolic design provides high throughput and unidirectional pipelining by exploiting the implicit parallelism in the genetic operators. The design is significant because, unlike other hardware genetic algorithms, it is independent of both the fitness function and the particular chromosome length used in a problem. We have designed and simulated a version of the mutation array using Xilinix FPGA tools to investigate the feasibility of hardware implementation. A simple 5-chromosome mutation array occupies 195 CLBs and is capable of performing more than one million mutations per second. I. Introduction Genetic algorithms (GAs) are established search and optimization techniques which have been applied to a range of engineering and applied problems with considerable success [1]. They operate by maintaining a population of trial solutions encoded, using a suitable encoding scheme.

A novel two pass hexagonal search algorithm for motion estimation

This paper presents a novel two-pass algorithm constituted by Linear Hashtable Motion Estimation Algorithm (LHMEA) and Hexagonal Search (HEXBS). compensation. for block base motion On the basis of research from previous algorithms, especially an on-the-edge motion estimation algorithm called hexagonal search (HEXBS), we propose the LHMEA and the Two-Pass Algorithm (TPA). We introduce hashtable into video compression. In this paper we employ LHMEA for the first-pass search in all the Macroblocks (MB) in the picture. Motion Vectors (MV) are then generated from the first-pass and are used as predictors for second-pass HEXBS motion estimation, which only searches a small number of MBs. The evaluation of the algorithm considers the three important metrics being time, compression rate and PSNR. The performance of the algorithm is evaluated by using standard video sequences and the results are compared to current algorithms. Experimental results show that the proposed algorithm can offer the same compression rate as the Full Search. LHMEA with TPA has significant improvement on HEXBS and shows a direction for improving other fast motion estimation algorithms, for example Diamond Search.

Optimization algorithms in FMRF model-based segmentation for LIDAR data and co-registered bands

In this paper, a fuzzy Markov random field (FMRF) model is used to segment land-objects into free, grass, building, and road regions by fusing remotely, sensed LIDAR data and co-registered color bands, i.e. scanned aerial color (RGB) photo and near infra-red (NIR) photo. An FMRF model is defined as a Markov random field (MRF) model in a fuzzy domain. Three optimization algorithms in the FMRF model, i.e. Lagrange multiplier (LM), iterated conditional mode (ICM), and simulated annealing (SA), are compared with respect to the computational cost and segmentation accuracy. The results have shown that the FMRF model-based ICM algorithm balances the computational cost and segmentation accuracy in land-cover segmentation from LIDAR data and co-registered bands.

Search space pruning and global optimisation of Multiple Gravity Assist spacecraft trajectories

We introduce and describe the Multiple Gravity Assist problem, a global optimisation problem that is of great interest in the design of spacecraft and their trajectories. We discuss its formalization and we show, in one particular problem instance, the performance of selected state of the art heuristic global optimisation algorithms. A deterministic search space pruning algorithm is then developed and its polynomial time and space complexity derived. The algorithm is shown to achieve search space reductions of greater than six orders of magnitude, thus reducing significantly the complexity of the subsequent optimisation.