[pt] O aumento da participação de fontes de energias renováveis variáveis (ERVs) na matriz elétrica do Brasil traz diversos desafios ao planejamento e à operação do Sistema Elétrico Brasileiro (SEB), devido à estocasticidade das ERVs. Tais desafios envolvem a modelagem e simulação dos processos intermitentes de geração e, dessa forma, um volume considerável de pesquisas tem sido direcionado ao tema. Nesse contexto, um tópico de crescente importância na literatura relaciona-se ao desenvolvimento de metodologias para simulação estocástica conjunta de recursos intermitentes com características complementares, como, por exemplo, as fontes eólica e solar. Visando contribuir com essa temática, este trabalho propõe melhorias a um modelo de simulação já estabelecido na literatura, avaliando sua aplicabilidade a partir de dados do Nordeste brasileiro. A metodologia proposta baseia-se em discretização das séries temporais de energia aplicando a técnica de machine learning k-means, construção de matrizes de transição de estados com base nos clusters identificados e simulação de Monte Carlo para obtenção dos cenários. As séries sintéticas obtidas são comparadas aos resultados gerados pelo modelo já estabelecido na literatura a partir de técnicas estatísticas. Quanto ao alcance dos objetivos da pesquisa, a modelagem proposta se mostrou mais eficiente, gerando cenários que reproduziram satisfatoriamente todas as características dos dados históricos avaliadas. / [en] The increased participation of variable renewable energy sources (VRES) in Brazil s electricity matrix brings several challenges to the planning and operation of the Brazilian Power System (BPS), due to the VRES stochasticity. Such challenges involve the modeling and simulation of intermittent generation processes and, in this context, a considerable amount of research has been directed to the theme. In this context, a topic of increasing importance in the literature is related to the development of methodologies for joint stochastic simulation of intermittent resources with complementary characteristics, such as wind and solar sources. Aiming to contribute to this theme, this work proposes improvements in a simulation model already established in the literature, evaluating its applicability based on Brazilian Northeast data. The proposed methodology is based on the discretization of energy time series applying the kmeans machine learning technique, construction of state transition matrices based on the identified clusters, and Monte Carlo simulation to obtain the scenarios. The synthetic series obtained are compared to the results generated by the model already established in the literature from statistical techniques. Regarding the scope of the research objectives, the proposed modeling demonstrated more promising results, generating scenarios that satisfactorily reproduced all the evaluated characteristics of the historical data.

Master of Science / Department of Electrical and Computer Engineering / Christopher L. Lewis / This thesis describes the development of an automated computer vision system that identifies and inventories road signs from imagery acquired from the Kansas Department of Transportation's road profiling system that takes images every 26.4 feet on highways through out the state. Statistical models characterizing the typical size, color, and physical location of signs are used to help identify signs from the imagery. First, two phases of a computationally efficient K-Means clustering algorithm are applied to the images to achieve over-segmentation. The novel second phase ensures over-segmentation without excessive computation. Extremely large and very small segments are rejected. The remaining segments are then classified based on color. Finally, the frame to frame trajectories of sign colored segments are analyzed using triangulation and Bundle adjustment to determine their physical location relative to the road video log system. Objects having the appropriate color, and physical placement are entered into a sign database. To develop the statistical models used for classification, a representative set of images was segmented and manually labeled determining the joint probabilistic models characterizing the color and location typical to that of road signs. Receiver Operating Characteristic curves were generated and analyzed to adjust the thresholds for the class identification. This system was tested and its performance characteristics are presented.

L'atténuation des photons est un phénomène qui affecte directement et de façon profonde la qualité et l'information quantitative obtenue d'une image en Tomographie d'Emission par Positrons (TEP). De sévères artefacts compliquant l'interprétation visuelle ainsi que de profondes erreurs d'exactitudes sont présents lors de l'évaluation quantitative des images TEP, biaisant la vérification de la corrélation entre les concentrations réelles et mesurées.L' atténuation est due aux effets photoélectrique et Compton pour l'image de transmission (30 keV - 140 keV), et majoritairement à l'effet Compton pour l'image d'émission (511 keV). La communauté en médecine nucléaire adhère largement au fait que la correction d'atténuation constitue une étape cruciale pour l'obtention d'images sans artefacts et quantitativement exactes. Pour corriger les images d'émission TEP pour l'atténué, l'approche proposée consiste concrètement à segmenter une image de transmission à l'aide d'algorithmes de segmentation: K-means (KM), Fuzzy C-means (FCM), Espérance-Maximisation (EM), et EM après une transformation en ondelettes (OEM). KM est un algorithme non supervisé qui partitionne les pixels de l'image en agrégats tels que chaque agrégat de la partition soit défini par ses objets et son centroïde. FCM est un algorithme de classification non-supervisée qui introduit la notion d'ensemble flou dans la définition des agrégats, et chaque pixel de l'image appartient à chaque agrégat avec un certain degré, et tous les agrégats sont caractérisés par leur centre de gravité.L'algorithme EM est une méthode d'estimation permettant de déterminer les paramètres du maximum de vraisemblance d'un mélange de distributions avec comme paramètres du modèle à estimer la moyenne, la covariance et le poids du mélange correspondant à chaque agrégat. Les ondelettes forment un outil pour la décomposition du signal en une suite de signaux dits d'approximation de résolution décroissante suivi d'une suite de rectifications appelées détails.L' image à laquelle a été appliquée les ondelettes est segmentée par EM. La correction d'atténuation nécessite la conversion des intensités des images de transmission segmentées en coefficients d'atténuation à 511 keV. Des facteurs de correction d' atténuation (FCA) pour chaque ligne de réponse sont alors obtenus, lesquels représentent le rapport entre les photons émis et transmis. Ensuite il s'agit de multiplier le sinogramme, formé par l'ensemble des lignes de réponses, des FCA par le sinogramme de l'image d'émission pour avoir le sinogramme corrigé pour l'atténuation, qui est par la suite reconstruit pour générer l'image d'émission TEP corrigée. Nous avons démontré l'utilité de nos méthodes proposées dans la segmentation d'images médicales en les appliquant à la segmentation des images du cerveau, du thorax et de l'abdomen humains. Des quatre processus de segmentation, la décomposition par les ondelettes de Haar suivie de l'Espérance-Maximisation (OEM) semble donner un meilleur résultat en termes de contraste et de résolution. Les segmentations nous ont permis une réduction claire de la propagation du bruit des images de transmission dans les images d'émission, permettant une amélioration de la détection des lésions, et améliorant les diagnostics en médecine nucléaire.

This thesis examines the performance of the support vector machine and the random forest models in the context of binary classification. The two techniques are compared and the outstanding one is used to construct a final parsimonious model. The data set consists of 33 observations and 89 biomarkers as features with no known dependent variable. The dependent variable is generated through k-means clustering, with a predefined final solution of two clusters. The training of the algorithms is performed using five-fold cross-validation repeated twenty times. The outcome of the training process reveals that the best performing versions of the models are a linear support vector machine and a random forest with six randomly selected features at each split. The final results of the comparison on the test set of these optimally tuned algorithms show that the random forest outperforms the linear kernel support vector machine. The former classifies all observations in the test set correctly whilst the latter classifies all but one correctly. Hence, a parsimonious random forest model using the top five features is constructed, which, to conclude, performs equally well on the test set compared to the original random forest model using all features.

Methods for determining clusters of data under- specified constraints have recently gained popularity. Although general constraints may be used, we focus on clustering methods with the constraint of a minimal cluster size. In this dissertation, we propose two constrained k-means algorithms: Linear Programming Algorithm (LPA) and Genetic Constrained K-means Algorithm (GCKA). Linear Programming Algorithm modifies the k-means algorithm into a linear programming problem with constraints requiring that each cluster have m or more subjects. In order to achieve an acceptable clustering solution, we run the algorithm with a large number of random sets of initial seeds, and choose the solution with minimal Root Mean Squared Error (RMSE) as our final solution for a given data set. We evaluate LPA with both generic data and simulated data and the results indicate that LPA can obtain a reasonable clustering solution. Genetic Constrained K-Means Algorithm (GCKA) hybridizes the Genetic Algorithm with a constrained k-means algorithm. We define Selection Operator, Mutation Operator and Constrained K-means operator. Using finite Markov chain theory, we prove that the GCKA converges in probability to the global optimum. We test the algorithm with several datasets. The analysis shows that we can achieve a good clustering solution by carefully choosing parameters such as population size, mutation probability and generation. We also propose a Bi-Nelder algorithm to search for an appropriate cluster number with minimal RMSE.

Mémoire numérisé par la Direction des bibliothèques de l'Université de Montréal.

Clustering algorithms are widely used in pattern recognition and data mining applications. Due to their computational efficiency, partitional clustering algorithms are better suited for applications with large datasets than hierarchical clustering algorithms. K-means is among the most popular partitional clustering algorithm, but has a major shortcoming: it is extremely sensitive to the choice of initial centers used to seed the algorithm. Unless k-means is carefully initialized, it converges to an inferior local optimum and results in poor quality partitions. Developing improved method for selecting initial centers for k-means is an active area of research. Genetic algorithms (GAs) have been successfully used to evolve a good set of initial centers. Among the most promising GA-based methods are those that exchange neighboring centers between candidate partitions in their crossover operations. K-means is best suited to work when datasets have well-separated non-overlapping clusters. Fuzzy c-means (FCM) is a popular variant of k-means that is designed for applications when clusters are less well-defined. Rather than assigning each point to a unique cluster, FCM determines the degree to which each point belongs to a cluster. Like k-means, FCM is also extremely sensitive to the choice of initial centers. Building on GA-based methods for initial center selection for k-means, this dissertation developed an evolutionary program for center selection in FCM called FCMGA. The proposed algorithm utilized region-based crossover and other mechanisms to improve the GA. To evaluate the effectiveness of FCMGA, three independent experiments were conducted using real and simulated datasets. The results from the experiments demonstrate the effectiveness and consistency of the proposed algorithm in identifying better quality solutions than extant methods. Moreover, the results confirmed the effectiveness of region-based crossover in enhancing the search process for the GA and the convergence speed of FCM. Taken together, findings in these experiments illustrate that FCMGA was successful in solving the problem of initial center selection in partitional clustering algorithms.

Data clustering, which partitions data points into clusters, has many useful applications in economics, science and engineering. Data clustering algorithms can be partitional or hierarchical. The k-means algorithm is the most widely used partitional clustering algorithm because of its simplicity and efficiency. One problem with the k-means algorithm is that the quality of partitions produced is highly dependent on the initial selection of centers. This problem has been tackled using genetic algorithms (GA) where a set of centers is encoded into an individual of a population and solutions are generated using evolutionary operators such as crossover, mutation and selection. Of the many GA methods, the region-based genetic algorithm (RBGA) has proven to be an effective technique when the centroid was used as the representative object of a cluster (ROC) and the Euclidean distance was used as the distance metric. The RBGA uses a region-based crossover operator that exchanges subsets of centers that belong to a region of space rather than exchanging random centers. The rationale is that subsets of centers that occupy a given region of space tend to serve as building blocks. Exchanging such centers preserves and propagates high-quality partial solutions. This research aims at assessing the RBGA with a variety of ROCs and distance metrics. The RBGA was tested along with other GA methods, on four benchmark datasets using four distance metrics, varied number of centers, and centroids and medoids as ROCs. The results obtained showed the superior performance of the RBGA across all datasets and sets of parameters, indicating that region-based crossover may prove an effective strategy across a broad range of clustering problems.

<div> <div> <div> <p>Organizations today collect and aggregate huge amounts of data from individuals under various scenarios and for different purposes. Such aggregation of individuals’ data when combined with techniques of data analytics allows organizations to make informed decisions and predictions. But in many situations, different portions of the data associated with individuals are collected and curated by different organizations. To derive more accurate conclusions and predictions, those organization may want to conduct the analysis based on their joint data, which cannot be simply accomplished by each organization exchanging its own data with other organizations due to the sensitive nature of data. Developing approaches for collaborative privacy-preserving data analytics, however, is a nontrivial task. At least two major challenges have to be addressed. The first challenge is that the security of the data possessed by each organization should always be properly protected during and after the collaborative analysis process, whereas the second challenge is the high computational complexity usually accompanied by cryptographic primitives used to build such privacy-preserving protocols. </p><p><br></p><p> </p><div> <div> <div> <p>In this dissertation, based on widely adopted primitives in cryptography, we address the aforementioned challenges by developing techniques for data analytics that not only allow multiple mutually distrustful parties to perform data analysis on their joint data in a privacy-preserving manner, but also reduce the time required to complete the analysis. More specifically, using three common data analytics tasks as concrete examples, we show how to construct the respective privacy-preserving protocols under two different scenarios: (1) the protocols are executed by a collaborative process only involving the participating parties; (2) the protocols are outsourced to some service providers in the cloud. Two types of optimization for improving the efficiency of those protocols are also investigated. The first type allows each participating party access to a statistically controlled leakage so as to reduce the amount of required computation, while the second type utilizes the parallelism that could be incorporated into the task and pushes some computation to the offline phase to reduce the time needed for each participating party without any additional leakage. Extensive experiments are also conducted on real-world datasets to demonstrate the effectiveness of our proposed techniques.<br></p> <p> </p> </div> </div> </div> </div> </div> </div>

The main task of this thesis is to perform image segmentation on images of fingers to partition the image into two parts, one with the fingers and one with all that is not fingers. First, we present the theory behind several well-used image segmentation methods, such as SNIC superpixels, the k-means algorithm, and the normalised cut algorithm. These have then been implemented and tested on images of fingers and the results are shown. The implementations are unfortunately not stable and give segmentations of varying results.