Dans cette thèse nous abordons deux problématiques différentes : la prédiction et la classification de trafic et les mécanismes d'accès dans les réseaux MANETs. Dans la première partie de la thèse, nous abordons le problème de la prédiction et la classification du trafic. Sur la base des observations du passé et sans considérer aucun modèle en particulier, nous analysons le problème de la prédiction en ligne de la charge sur un lien. Concernant la classification du trafic, nous nous concentrons principalement sur des applications P2P, et particulièrement la télévision P2P (P2P-TV). Dans les deux cas, nous employons la technique de Support Vector Machines (SVM). Les algorithmes que nous proposons fournissent des résultats très précis. De plus, ils sont robustes et leur coût est extrêmement bas. Ces propriétés font que nos solutions soient particulièrement adaptées à des applications en temps réel. Dans la deuxième partie de la thèse, nous abordons deux problèmes différents liés aux mécanismes d'accès dans les réseaux MANETs, et en particulier, nous nous concentrons sur CSMA. Nous présentons d'abord les différents modèles existants pour CSMA et nous identifions leurs principaux points faibles. Des solutions possibles sont proposées, bases sur les outils de la géométrie aléatoire. Nous abordons ensuite le problème de QoS dans CSMA et nous proposons deux mécanismes différents permettant de garantir un débit minimum pour chaque transmission admise. Le but principal étant d'identifier le meilleur mécanisme dans un scénario donné comparé au protocole CSMA.

近幾年無線寬頻網路崛起，寄望WiMAX可以取代最後一哩，雖然WiMAX有QoS的設計，但是對於Call Admission Control、Bandwidth Allocation、Scheduler並沒有實際定義，給予廠商彈性設計。本篇論文提出以機器學習的方式依據網路狀態動態配置頻寬，以符合實際頻寬需求。 由於BS在配置頻寬的時候並沒有SS佇列的訊息，使得BS無法配置適合的頻寬，達到較好的效能，尤其是有期限的rtPS封包最為明顯。在系統負載較高的環境下，容易導致封包遺失提升，吞吐量降低的情形發生。因此本研究提出了支持向量機的方式，收集大量Training Data，訓練成動態頻寬配置模組；以動態配置適合的頻寬給rtPS，使rtPS在負載高的環境下的封包遺失率降低，且延遲能夠維持一定水準。搭配適應性頻寬配置策略，在低負載的環境下可以保留少許頻寬給Non Real Time Traffic，在高負載環境下，先滿足Real Time Traffic為原則。模擬工具採用NS 2-2.29、長庚大學-資策會的WiMAX模組，以及台大林智仁老師開發的支持向量機函式庫libSVM。 / In recent years, the rise of wireless broadband access networks. Hope that WiMAX can solve the last mile problem. Although WiMAX has QoS design, but for call admission control, bandwidth allocation, scheduler are not defined in standard. In this paper, we proposed a machine learning approach dynamic bandwidth allocation based on network state. BS because of the bandwidth allocation at a time when there is no message of SS’s queue. Enables BS can not configure a more suitable bandwidth to achieve better performance. In particular, there is the deadline of rtPS packets. At the higher loading on the system environment, easily lead to packet loss raise, lower throughput situations happen. In this study, a support vector machine approach to collect a large number of training data. Training modules into a dynamic bandwidth allocation. We can dynamically allocate bandwidth to fit rtPS. Adaptive bandwidth allocation strategy, at the low loading environment can keep some bandwidth for non real time traffic. At a high loading environment must first meet the real time traffic. We use Network Simulator 2-2.29, CGU-III WiMAX module, libSVM library.

Les capteurs RSO (Radar à Synthèse d'Ouverture) fournissent des observations des surfaces terrestres de manière continue depuis 1991 avec la mise en orbite du satellite ERS-1. Les données acquises jusqu'à peu, principalement basées sur l'exploitation de l'intensité du signal acquis selon une configuration de polarisation particulière, ont été l'objet de nombreuses études, notamment sur le suivi de la déforestation. Depuis 2007, de nouveaux capteurs RSO polarimétriques (PALSAR, RADARSAT-2, TerraSAR-X...) permettent la caractérisation polarimétrique des surfaces observées. Ces données nécessitent des traitements adpatés afin d'en extraire l'information la plus pertinente pour la thématique considérée. L'objet de ces travaux a été d'évaluer leur potentiel pour la cartographie de surfaces naturelles en milieu tropical. L'apport des multiples indices polarimétriques a été évalué à partir de l'algorithme de classification SVM (Machines à Vecteurs de Support). Cet algorithme est spécialement adapté pour prendre en compte un grand nombre d'indices non forcément homogènes. Les données utilisées ont été acquises par le capteur aéroporté AIRSAR sur une île en Polynésie Française. De nombreux relevés in situ ont permis la validation des résultats obtenus. Les résultats montrent que la sensibilité de ces données à la structure géométrique des surfaces observées permet une bonne discrimination entre les différents couvert végétaux étudiés, en particulier des types de forêts. De plus, la classification obtenue à partir de la méthode SVM est particulièrement plus performante que la classification usuelle basée sur la distribution de Wishart vérifiée a priori par les données radar. Ces résultats laissent présager de l'apport significatif des données radar polarimétriques futures pour le suivi des surfaces naturelles

Ingeniero Civil Eléctrico / La energía eléctrica es una de las formas de energía más usadas por el ser humano. Sin ella muchas de las comodidades a las que se está habituado desaparecerían. Sin embargo uno de sus mayores consumidores es la industria, donde tan solo la idea de una mala calidad de suministro sostenida en el tiempo podría causar una gran conmoción. Es por eso que los sistemas de generación de esta energía deben ser monitoreados constantemente en búsqueda de posibles fallas o anomalías que pongan en peligro la disponibilidad de los equipos eléctricos ahí usados, en especial de las máquinas usadas para la generación. Los generadores sincrónicos son las máquinas rotatorias más usadas en la industria de la generación de energía eléctrica, es por eso que el diagnóstico de fallas para estos equipos toma gran importancia a nivel mundial. En el presente trabajo de título se diseña un algoritmo de diagnóstico de fallas orientado a detectar y clasificar fallas de tipo monofásicas para máquinas sincrónicas de polos salientes, monitoreando las corrientes de estator trifásicas y la corriente de campo. Está basado en el uso de una novedosa técnica de aprendizaje supervisado llamada Máquina de Vectores de Soporte (SVM), la cual, mediante su sistema de implementación uno contra el resto es capaz de clasificar el estado de la máquina en 4 clases distintas: sano , falla clase 1 , falla clase 2 y falla clase 3 . La SVM recibe como entrada los llamados atributos de falla, variables que se obtienen a partir de las corrientes monitoreadas y se caracterizan por poseer la información suficiente para que la SVM pueda resolver el problema de clasificación planteado. Los atributos son obtenidos a través del análisis de las corrientes de estator y de campo. Consisten en un conjunto formado por distintas frecuencias de falla (obtenidas mediante la Transformada de Fourier de las distintas señales de entrada) como también de amplitudes o características de las corrientes en el tiempo. Los datos de operación de la máquina sincrónica que son usados para entrenar, probar y validar el algoritmo de diagnóstico se obtienen a partir de simulaciones del modelo basado en la representación del Voltaje detrás de la Reactancia, este modelo implementa una novedosa forma de subdividir los devanados de estator de la máquina con el fin de simular fallas internas. El algoritmo es validado usando datos contaminados con ruido blanco Gaussiano en distintos niveles, logrando una tasa correcta de clasificación del 97:5% para datos contaminados con ruido S=N = 30[dB], lo que indica que el método propuesto es robusto ante perturbaciones y podría ser aplicado experimentalmente en el diagnóstico de fallas monofásicas en máquinas sincrónicas de polos salientes.

Text classification is used in information extraction and retrieval from a given text, and text classification has been considered as an important step to manage a vast number of records given in digital form that is far-reaching and expanding. This thesis addresses patent document classification problem into fifteen different categories or classes, where some classes overlap with other classes for practical reasons. For the development of the classification model using machine learning techniques, useful features have been extracted from the given documents. The features are used to classify patent document as well as to generate useful tag-words. The overall objective of this work is to systematize NASA’s patent management, by developing a set of automated tools that can assist NASA to manage and market its portfolio of intellectual properties (IP), and to enable easier discovery of relevant IP by users. We have identified an array of methods that can be applied such as k-Nearest Neighbors (kNN), two variations of the Support Vector Machine (SVM) algorithms, and two tree based classification algorithms: Random Forest and J48. The major research steps in this work consist of filtering techniques for variable selection, information gain and feature correlation analysis, and training and testing potential models using effective classifiers. Further, the obstacles associated with the imbalanced data were mitigated by adding synthetic data wherever appropriate, which resulted in a superior SVM classifier based model.

The area of misfire detection is important because of the effects of misfires on both the environment and the exhaust system. Increasing requirements on the detection performance means that improvements are always of interest. In this thesis, potential improvements to an existing misfire detection algorithm are evaluated. The improvements evaluated are: using Gaussian processes to model the classifier, alternative signal treatments for detection of multiple misfires, and effects of where flywheel tooth angle error estimation is performed. The improvements are also evaluated for their suitability for use on-line. Both the use of Gaussian processes and the detection of multiple misfires are hard problems to solve while maintaining detection performance. Gaussian processes most likely loses performance due to loss of dependence between the weights of the classifier. It can give performance similar to the original classifier, but with greatly increased complexity. For multiple misfires, the performance can be slightly improved without loss of single misfire performance. Greater improvements are possible, but at the cost of single misfire performance. The decision is in the end down to the desired trade-off. The flywheel tooth angle error compensation gives nearly identical performance regardless of where it is estimated. Consequently the error estimation can be separated from the signal processing, allowing the implementation to be modular. Using an EKF for estimating the flywheel errors on-line is found to be both feasible and give good performance. Combining the separation of the error estimation from the signal treatment with a, after initial convergence, heavily restricted EKF gives a vastly reduced computational load for only a moderate loss of performance.

This dissertation deals with developing parallel processing algorithms for Graphic Processing Unit (GPU) in order to solve machine learning problems for large datasets. In particular, it contributes to the development of fast GPU based algorithms for calculating distance (i.e. similarity, affinity, closeness) matrix. It also presents the algorithm and implementation of a fast parallel Support Vector Machine (SVM) using GPU. These application tools are developed using Compute Unified Device Architecture (CUDA), which is a popular software framework for General Purpose Computing using GPU (GPGPU). Distance calculation is the core part of all machine learning algorithms because the closer the query is to some samples (i.e. observations, records, entries), the more likely the query belongs to the class of those samples. K-Nearest Neighbors Search (k-NNS) is a popular and powerful distance based tool for solving classification problem. It is the prerequisite for training local model based classifiers. Fast distance calculation can significantly improve the speed performance of these classifiers and GPUs can be very handy for their accelerations. Meanwhile, several GPU based sorting algorithms are also included to sort the distance matrix and seek for the k-nearest neighbors. The speed performances of the sorting algorithms vary depending upon the input sequences. The GPUKNN proposed in this dissertation utilizes the GPU based distance computation algorithm and automatically picks up the most suitable sorting algorithm according to the characteristics of the input datasets. Every machine learning tool has its own pros and cons. The advantage of SVM is the high classification accuracy. This makes SVM possibly the best classification tool. However, as in many other machine learning algorithms, SVM's slow training phase slows down when the size of the input datasets increase. The GPU version of parallel SVM based on parallel Sequential Minimal Optimization (SMO) implemented in this dissertation is proposed to reduce the time cost in both training and predicting phases. This implementation of GPUSVM is original. It utilizes many parallel processing techniques to accelerate and minimize the computations of kernel evaluation, which are considered as the most time consuming operations in SVM. Although the many-core architecture of GPU performs the best in data level parallelism, multi-task (aka. task level parallelism) processing is also integrated into the application to improve the speed performance of tasks such as multiclass classification and cross-validation. Furthermore, the procedure of finding worst violators is distributed to multiple blocks on the CUDA model. This reduces the time cost for each iteration of SMO during the training phase. All of these violators are shared among different tasks in multiclass classification and cross-validation to reduce the duplicate kernel computations. The speed performance results have shown that the achieved speedup of both the training phase and predicting phase are ranging from one order of magnitude to three orders of magnitude times faster compared to the state of the art LIBSVM software on some well known benchmarking datasets.

Support Vector Machines (SVMs) are used for a growing number of applications. A fundamental constraint on SVM learning is the management of the training set. This is because the order of computations goes as the square of the size of the training set. Typically, training sets of 1000 (500 positives and 500 negatives, for example) can be managed on a PC without hard-drive thrashing. Training sets of 10,000 however, simply cannot be managed with PC-based resources. For this reason most SVM implementations must contend with some kind of chunking process to train parts of the data at a time (10 chunks of 1000, for example, to learn the 10,000). Sequential and multi-threaded chunking methods provide a way to run the SVM on large datasets while retaining accuracy. The multi-threaded distributed SVM described in this thesis is implemented using Java RMI, and has been developed to run on a network of multi-core/multi-processor computers.

This work is an investigation into reconstructing fragmented ASCII files based on content analysis motivated by a desire to demonstrate machine learning's applicability to Digital Forensics. Using a categorized corpus of Usenet, Bulletin Board Systems, and other assorted documents a series of experiments are conducted using machine learning techniques to train classifiers which are able to identify fragments belonging to the same original file. The primary machine learning method used is the Support Vector Machine with a variety of feature extractions to train from. Additional work is done in training committees of SVMs to boost the classification power over the individual SVMs, as well as the development of a method to tune SVM kernel parameters using a genetic algorithm. Attention is given to the applicability of Information Retrieval techniques to file fragments, as well as an analysis of textual artifacts which are not present in standard dictionaries.

There is a significant need to identify approaches for classifying chemical sensor array data with high success rates that would enhance sensor detection capabilities. The present study attempts to fill this need by investigating six machine learning methods to classify a dataset collected using a chemical sensor array: K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Classification and Regression Trees (CART), Random Forest (RF), Naïve Bayes Classifier (NB), and Principal Component Regression (PCR). A total of 10 predictors that are associated with the response from 10 sensor channels are used to train and test the classifiers. A training dataset of 4 classes containing 136 samples is used to build the classifiers, and a dataset of 4 classes with 56 samples is used for testing. The results generated with the six different methods are compared and discussed. The RF, CART, and KNN are found to have success rates greater than 90%, and to outperform the other methods.