La thèse consiste en l'apprentissage d'une tâche complexe de robotique de manipulation en utilisant très peu d'aprioris. Plus précisément, la tâche apprise consiste à atteindre un objet avec un robot série. L'objectif est de réaliser cet apprentissage sans paramètres de calibrage des caméras, modèles géométriques directs, descripteurs faits à la main ou des démonstrations d'expert. L'apprentissage par renforcement profond est une classe d'algorithmes particulièrement intéressante dans cette optique. En effet, l'apprentissage par renforcement permet d’apprendre une compétence sensori-motrice en se passant de modèles dynamiques. Par ailleurs, l'apprentissage profond permet de se passer de descripteurs faits à la main pour la représentation d'état. Cependant, spécifier les objectifs sans supervision humaine est un défi important. Certaines solutions consistent à utiliser des signaux de récompense informatifs ou des démonstrations d'experts pour guider le robot vers les solutions. D'autres consistent à décomposer l'apprentissage. Par exemple, l'apprentissage "petit à petit" ou "du simple au compliqué" peut être utilisé. Cependant, cette stratégie nécessite la connaissance de l'objectif en termes d'état. Une autre solution est de décomposer une tâche complexe en plusieurs tâches plus simples. Néanmoins, cela n'implique pas l'absence de supervision pour les sous tâches mentionnées. D'autres approches utilisant plusieurs robots en parallèle peuvent également être utilisés mais nécessite du matériel coûteux. Pour notre approche, nous nous inspirons du comportement des êtres humains. Ces derniers généralement regardent l'objet avant de le manipuler. Ainsi, nous décomposons la tâche d'atteinte en 3 sous tâches. La première tâche consiste à apprendre à fixer un objet avec un système de deux caméras pour le localiser dans l'espace. Cette tâche est apprise avec de l'apprentissage par renforcement profond et un signal de récompense faiblement supervisé. Pour la tâche suivante, deux compétences sont apprises en parallèle : la fixation d'effecteur et une fonction de coordination main-oeil. Comme la précédente tâche, un algorithme d'apprentissage par renforcement profond est utilisé avec un signal de récompense faiblement supervisé. Le but de cette tâche est d'être capable de localiser l'effecteur du robot à partir des coordonnées articulaires. La dernière tâche utilise les compétences apprises lors des deux précédentes étapes pour apprendre au robot à atteindre un objet. Cet apprentissage utilise les mêmes aprioris que pour les tâches précédentes. En plus de la tâche d'atteinte, un predicteur d'atteignabilité d'objet est appris. La principale contribution de ces travaux est l'apprentissage d'une tâche de robotique complexe en n'utilisant que très peu de supervision. / The thesis is focused on learning a complex manipulation robotics task using little knowledge. More precisely, the concerned task consists in reaching an object with a serial arm and the objective is to learn it without camera calibration parameters, forward kinematics, handcrafted features, or expert demonstrations. Deep reinforcement learning algorithms suit well to this objective. Indeed, reinforcement learning allows to learn sensori-motor mappings while dispensing with dynamics. Besides, deep learning allows to dispense with handcrafted features for the state spacerepresentation. However, it is difficult to specify the objectives of the learned task without requiring human supervision. Some solutions imply expert demonstrations or shaping rewards to guiderobots towards its objective. The latter is generally computed using forward kinematics and handcrafted visual modules. Another class of solutions consists in decomposing the complex task. Learning from easy missions can be used, but this requires the knowledge of a goal state. Decomposing the whole complex into simpler sub tasks can also be utilized (hierarchical learning) but does notnecessarily imply a lack of human supervision. Alternate approaches which use several agents in parallel to increase the probability of success can be used but are costly. In our approach,we decompose the whole reaching task into three simpler sub tasks while taking inspiration from the human behavior. Indeed, humans first look at an object before reaching it. The first learned task is an object fixation task which is aimed at localizing the object in the 3D space. This is learned using deep reinforcement learning and a weakly supervised reward function. The second task consists in learning jointly end-effector binocular fixations and a hand-eye coordination function. This is also learned using a similar set-up and is aimed at localizing the end-effector in the 3D space. The third task uses the two prior learned skills to learn to reach an object and uses the same requirements as the two prior tasks: it hardly requires supervision. In addition, without using additional priors, an object reachability predictor is learned in parallel. The main contribution of this thesis is the learning of a complex robotic task with weak supervision.

Cette thèse présente un cadre formel pour l'interaction Homme-robot (HRI), qui reconnaître un important lexique de gestes statiques et dynamiques mesurés par des capteurs portatifs. Gestes statiques et dynamiques sont classés séparément grâce à un processus de segmentation. Les tests expérimentaux sur la base de données de gestes UC2017 ont montré une haute précision de classification. La classification pas à pas en ligne utilisant des données brutes est fait avec des réseaux de neurones profonds « Long-Short Term Memory » (LSTM) et à convolution (CNN), et sont plus performants que les modèles statiques entraînés avec des caractéristiques spécialement conçues, au détriment du temps d'entraînement et d'inférence. La classification en ligne des gestes permet une classification prédictive avec réussit. Le rejet des gestes hors vocabulaire est proposé par apprentissage semi-supervisé par un réseau de neurones du type « Auxiliary Conditional Generative Adversarial Networks ». Le réseau propose a atteint une haute précision de rejet de les gestes non entraînés de la base de données UC2018 DualMyo. / This thesis presents a human-robot interaction (HRI) framework to classify large vocabularies of static and dynamic hand gestures, captured with wearable sensors. Static and dynamic gestures are classified separately thanks to the segmentation process. Experimental tests on the UC2017 hand gesture dataset showed high accuracy. In online frame-by-frame classification using raw incomplete data, Long Short-Term Memory (LSTM) deep networks and Convolutional Neural Networks (CNN) performed better than static models with specially crafted features at the cost of training and inference time. Online classification of dynamic gestures allows successful predictive classification. The rejection of out-of-vocabulary gestures is proposed to be done through semi-supervised learning of a network in the Auxiliary Conditional Generative Adversarial Networks framework. The proposed network achieved a high accuracy on the rejection of untrained patterns of the UC2018 DualMyo dataset.

Aprendizado supervisionado tem sido principalmente utilizado para classificação. Neste trabalho são mostrados os benefícios do uso de rankings ao invés de classificação de exemplos isolados. Um rankeador é um algoritmo que ordena um conjunto de exemplos de tal modo que eles são apresentados do exemplo de maior para o exemplo de menor expectativa de ser positivo. Um ranking é o resultado dessa ordenação. Normalmente, um ranking é obtido pela ordenação do valor de confiança de classificação dado por um classificador. Este trabalho tem como objetivo procurar por novas abordagens para promover o uso de rankings. Desse modo, inicialmente são apresentados as diferenças e semelhanças entre ranking e classificação, bem como um novo algoritmo de ranking que os obtém diretamente sem a necessidade de obter os valores de confiança de classificação, esse algoritmo é denominado de LEXRANK. Uma área de pesquisa bastante importante em rankings é a análise ROC. O estudo de árvores de decisão e análise ROC é bastante sugestivo para o desenvolvimento de uma visualização da construção da árvore em gráficos ROC. Para mostrar passo a passo essa visualização foi desenvolvido uma sistema denominado PROGROC. Ainda do estudo de análise ROC, foi observado que a inclinação (coeficiente angular) dos segmentos que compõem o fecho convexo de curvas ROC é equivalente a razão de verossimilhança que pode ser convertida para probabilidades. Essa conversão é denominada de calibração por fecho convexo de curvas ROC que coincidentemente é equivalente ao algoritmo PAV que implementa regressão isotônica. Esse método de calibração otimiza Brier Score. Ao explorar essa medida foi encontrada uma relação bastante interessante entre Brier Score e curvas ROC. Finalmente, também foram explorados os rankings construídos durante o método de seleção de exemplos do algoritmo de aprendizado semi-supervisionado multi-descrição CO-TRAINING / Supervised learning has been used mostly for classification. In this work we show the benefits of a welcome shift in attention from classification to ranking. A ranker is an algorithm that sorts a set of instances from highest to lowest expectation that the instance is positive, and a ranking is the outcome of this sorting. Usually a ranking is obtained by sorting scores given by classifiers. In this work, we are concerned about novel approaches to promote the use of ranking. Therefore, we present the differences and relations between ranking and classification followed by a proposal of a novel ranking algorithm called LEXRANK, whose rankings are derived not from scores, but from a simple ranking of attribute values obtained from the training data. One very important field which uses rankings as its main input is ROC analysis. The study of decision trees and ROC analysis suggested an interesting way to visualize the tree construction in ROC graphs, which has been implemented in a system called PROGROC. Focusing on ROC analysis, we observed that the slope of segments obtained from the ROC convex hull is equivalent to the likelihood ratio, which can be converted into probabilities. Interestingly, this ROC convex hull calibration method is equivalent to Pool Adjacent Violators (PAV). Furthermore, the ROC convex hull calibration method optimizes Brier Score, and the exploration of this measure leads us to find an interesting connection between the Brier Score and ROC Curves. Finally, we also investigate rankings build in the selection method which increments the labelled set of CO-TRAINING, a semi-supervised multi-view learning algorithm

Les nouvelles techniques de fabrication nanométriques comme l’auto-assemblage ou la nanoimpression permettent de réaliser des matrices régulières (crossbars) atteignant des densités extrêmes (jusqu’à 1012 nanocomposants/cm2) tout en limitant leur coût de fabrication. Cependant, il est attendu que ces technologies s’accompagnent d’une augmentation significative du nombre de défauts et de dispersions de caractéristiques. La capacité à exploiter ces crossbars est alors conditionnée par le développement de nouvelles techniques de calcul capables de les spécialiser et de tolérer une grande densité de défauts. Dans ce contexte, l’approche neuromimétique qui permet tout à la fois de configurer les nanodispositifs et de tolérer leurs défauts et dispersions de caractéristiques apparaît spécialement pertinente. L’objectif de cette thèse est de démontrer l’efficacité d’une telle approche et de quantifier la fiabilité obtenue avec une architecture neuromimétique à base de crossbar de memristors, ou neurocrossbar (NC). Tout d’abord la thèse introduit des algorithmes permettant l’apprentissage de fonctions logiques sur un NC. Par la suite, la thèse caractérise la tolérance du modèle NC aux défauts et aux variations de caractéristiques des memristors. Des modèles analytiques probabilistes de prédiction de la convergence de NC ont été proposés et confrontés à des simulations Monte-Carlo. Ils prennent en compte l’impact de chaque type de défaut et de dispersion. Grâce à ces modèles analytiques il devient possible d’extrapoler cette étude à des circuits NC de très grande taille. Finalement, l’efficacité des méthodes proposées est expérimentalement démontrée à travers l’apprentissage de fonctions logiques par un NC composé de transistors à nanotube de carbone à commande optique (OG-CNTFET). / Novel manufacturing techniques, such as nanoscale self-assembly or nanoimprint, allow a cost-efficient way to fabricate high-density crossbar matrices (1012 nanodevices/cm2). However, it is expected that these technologies will be accompanied by a significant increase of defects and dispersion in device characteristics. Thus, programming these crossbars require new computational techniques that possess high tolerance for such variations. In this context, approaches based on neural networks are promising for configuring nanodevices, since they provide a natural way for tolerating low yields and device variations. The main objective of this thesis is to explore such a neural-network approach, by examining factors such as efficiency and reliability, using the memristor crossbar architecture or neurocrossbar (NC). We introduce algorithms for learning the logic functions on the NC, and the tolerance of NC against static defects (stuck-defect) and dispersion of device properties is discussed. Probabilistic analytical models for predicting the convergence of NC are proposed and compared with Monte Carlo simulations, which take into account the impact of each type of defect and dispersion. These analytical models can be extrapolated to study large-sized NCs. Finally, the effectiveness of the proposed methods is experimentally demonstrated through the learning of logic functions by a real NC made of Optically Gated Carbon Nanotube Field Effect Transistor (OG-CNTFET).

A busca progressiva no espaço de estados é uma das abordagens mais populares de Planejamento Automatizado. O desempenho dos algoritmos de busca progressiva é influenciado pela heurística independente de domínio utilizada para guiá-lo. Nesse contexto, o foco do presente trabalho consiste em investigar técnicas de aprendizado de máquina supervisionadas que possibilitaram agregar à heurística do plano relaxado, comumente utilizada em abordagens atuais de planejamento, informações sobre o domínio em questão que viessem a ser úteis ao algoritmo de busca. Essas informações foram representadas por meio de um espaço de características do problema de planejamento e uma rede neural MLP foi aplicada para estimar uma nova função heurística para guiar a busca por meio de um processo de regressão não linear. Uma vez que o conjunto de características disponíveis para a construção da nova função heurística é grande, foi necessário a definição de um processo de seleção de características capaz de determinar qual conjunto de características de entrada da rede resultaria em melhor desempenho para o modelo de regressão. Portanto, para a seleção de características, aplicou-se uma abordagem de algoritmos genéticos. Como principal resultado, tem-se uma análise comparativa do desempenho entre a utilização da heurística proposta neste trabalho e a utilização da heurística do plano relaxado para guiar o algoritmo de busca na tarefa de planejamento. Para a análise empírica foram utilizados domínios de diferentes complexidades disponibilizados pela Competições Internacionais de Planejamento. Além dos resultados empíricos e análises comparativas, as contribuições deste trabalho envolvem o desenvolvimento de um novo planejador independente de domínio, denominado LearnInPlanner. Esse planejador utiliza a nova função heurística estimada por meio do processo de aprendizado e o algoritmo de Busca Gulosa para solucionar os problemas de planejamento. / The forward state-space search is one of the most popular Automated Planning approaches. The performance of forward search algorithms is affected by the domain-independent heuristic being used. In this context, the focus of this work consisted on investigating techniques of supervised machine learning that make possible to agregate to the relaxed plan heuristic, commonly used in current planning approaches, information about the domain which could be useful to the search algorithm. This information has been represented through a feature space of planning problem and a MLP neural network has been applied to estimate a new heuristic function for guiding the search through a non-linear regression process. Once the set of features available for the construction of the new heuristic function is large, it was necessary to define a feature selection process capable of determining which set of neural network input features would result in the best performance for the regression model. Therefore, for selecting features, an approach of genetic algorithms has been applied. As the main result, one has obtained a comparative performance analysis between the use of heuristic proposed in this work and the use of the relaxed plan heuristic to guide the search algorithm in the planning task. For the empirical analysis were used domains with different complexities provided by the International Planning Competitions. In addition to the empirical results and comparative analysis, the contributions of this work involves the development of a new domain-independent planner, named LearnInPlanner. This planner uses the new heuristic function estimated by the learning process and the Greedy Best-First search algorithm to solve planning problems.

Nos últimos 30 anos, a área de aprendizagem de máquina desenvolveu-se de forma comparável com a Física no início do século XX. Esse avanço tornou possível a resolução de problemas do mundo real que anteriormente não poderiam ser solucionados por máquinas, devido à dificuldade de modelos puramente estatísticos ajustarem-se de forma satisfatória aos dados de treinamento. Dentre tais avanços, pode-se citar a utilização de técnicas de aprendizagem de máquina na área de Química Medicinal, envolvendo métodos de análise, representação e predição de informação molecular por meio de recursos computacionais. Os dados utilizados no contexto biológico possuem algumas características particulares que podem influenciar no resultado de sua análise. Dentre estas, pode-se citar a complexidade das informações moleculares, o desbalanceamento das classes envolvidas e a existência de dados incompletos ou rotulados de forma incerta. Tais adversidades podem prejudicar o processo de identificação de compostos candidatos a novos fármacos, se não forem tratadas de forma adequada. Neste trabalho, foi abordada uma técnica de aprendizagem de máquina semi-supervisionada capaz de reduzir o impacto causado pelo problema da incerteza na rotulação dos dados, aplicando um método para estimar rótulos mais confiáveis para os compostos químicos existentes no conjunto de treinamento. Na tentativa de evitar os efeitos causados pelo desbalanceamento dos dados, foi incorporada ao processo de estimação de rótulos uma abordagem sensível ao custo, com o objetivo de evitar o viés em benefício da classe majoritária. Após o tratamento do problema da incerteza na rotulação, classificadores baseados em Máquinas de Aprendizado Extremo foram construídos, almejando boa capacidade de aproximação em um tempo de processamento reduzido em relação a outras abordagens de classificação comumente aplicadas. Por fim, o desempenho dos classificadores construídos foi avaliado por meio de análises dos resultados obtidos, confrontando o cenário com os dados originais e outros com as novas rotulações obtidas durante o processo de estimação semi-supervisionado / In the last 30 years, the area of machine learning has developed in a way comparable to Physics in the early twentieth century. This breakthrough has made it possible to solve real-world problems that previously could not be solved by machines because of the difficulty of purely statistical models to fit satisfactorily with training data. Among these advances, one can cite the use of machine learning techniques in the area of Medicinal Chemistry, involving methods for analysing, representing and predicting molecular information through computational resources. The data used in the biological context have some particular characteristics that can influence the result of its analysis. These include the complexity of molecular information, the imbalance of the classes involved, and the existence of incomplete or uncertainly labeled data. If they are not properly treated, such adversities may affect the process of identifying candidate compounds for new drugs. In this work, a semi-supervised machine learning technique was considered to reduce the impact caused by the problem of uncertainty in the data labeling, by applying a method to estimate more reliable labels for the chemical compounds in the training set. In an attempt to reduce the effects caused by data imbalance, a cost-sensitive approach was incorporated to the label estimation process, in order to avoid bias in favor of the majority class. After addressing the uncertainty problem in labeling, classifiers based on Extreme Learning Machines were constructed, aiming for good approximation ability in a reduced processing time in relation to other commonly applied classification approaches. Finally, the performance of the classifiers constructed was evaluated by analyzing the results obtained, comparing the scenario with the original data and others with the new labeling obtained by the semi-supervised estimation process

Com o aumento da capacidade de armazenamento, as bases de dados são cada vez maiores e, em muitas situações, apenas um pequeno subconjunto de itens de dados pode ser rotulado. Isto acontece devido ao processo de rotulagem ser frequentemente caro, demorado e necessitar do envolvimento de especialistas humanos. Com isso, diversos algoritmos semissupervisionados foram propostos, mostrando que é possível obter bons resultados empregando conhecimento prévio, relativo à pequena fração de dados rotulados. Dentre esses algoritmos, os que têm ganhado bastante destaque na área têm sido aqueles baseados em redes. Tal interesse, justifica-se pelas vantagens oferecidas pela representação via redes, tais como, a possibilidade de capturar a estrutura topológica dos dados, representar estruturas hierárquicas, bem como modelar manifolds no espaço multi-dimensional. No entanto, existe uma grande quantidade de dados representados em tabelas atributo-valor, nos quais não se poderia aplicar os algoritmos baseados em redes sem antes construir uma rede a partir desses dados. Como a geração das redes, assim como sua relação com o desempenho dos algoritmos têm sido pouco estudadas, esta tese investigou esses aspectos e propôs novos métodos para construção de redes, considerando características ainda não exploradas na literatura. Foram propostos três métodos para construção de redes com diferentes topologias: 1) S-kNN (Sequential k Nearest Neighbors), que gera redes regulares; 2) GBILI (Graph Based on the Informativeness of Labeled Instances) e RGCLI (Robust Graph that Considers Labeled Instances), que exploram os rótulos disponíveis gerando redes com distribuição de grau lei de potência; 3) GBLP (Graph Based on Link Prediction), que se baseia em medidas de predição de links gerando redes com propriedades mundo-pequeno. As estratégias de construção de redes propostas foram analisadas por meio de medidas de teoria dos grafos e redes complexas e validadas por meio da classificação semissupervisionada. Os métodos foram aplicados em benchmarks da área e também na classificação de gêneros musicais e segmentação de imagens. Os resultados mostram que a topologia da rede influencia diretamente os algoritmos de classificação e as estratégias propostas alcançam boa acurácia. / With the increase capacity of storage, databases are getting larger and, in many situations, only a small subset of data items can be labeled. This happens because the labeling process is often expensive, time consuming and requires the involvement of human experts. Hence, several semi-supervised algorithms have been proposed, showing that it is possible to achieve good results by using prior knowledge. Among these algorithms, those based on graphs have gained prominence in the area. Such interest is justified by the benefits provided by the representation via graphs, such as the ability to capture the topological structure of the data, represent hierarchical structures, as well as model manifold in high dimensional spaces. Nevertheless, most of available data is represented by attribute-value tables, making necessary the study of graph construction techniques in order to convert these tabular data into graphs for applying such algorithms. As the generation of the weight matrix and the sparse graph, and their relation to the performance of the algorithms have been little studied, this thesis investigated these aspects and proposed new methods for graph construction with characteristics litle explored in the literature yet. We have proposed three methods for graph construction with different topologies: 1) S-kNN (Sequential k Nearest Neighbors) that generates regular graphs; 2) GBILI (Graph Based on the informativeness of Labeled Instances) and RGCLI (Robust Graph that Considers Labeled Instances), which exploit the labels available generating power-law graphs; 3) GBLP (Graph Based on Link Prediction), which are based on link prediction measures and generates small-world graphs. The strategies proposed were analyzed by graph theory and complex networks measures and validated in semi-supervised classification tasks. The methods were applied in benchmarks of the area and also in the music genre classification and image segmentation. The results show that the topology of the graph directly affects the classification algorithms and the proposed strategies achieve good accuracy.

Cette étude a pour sujet les hypergraphes. / This study has for subject the hypergraphs.

<p>Cilj &ndash; Cilj istraživanja u okviru doktorske disertacije je razvoj sistema za automatsku polu-nadgledanu klasifikaciju podataka. Sistem bi trebao biti primenljiv na &scaron;irokom spektru domena gde je neophodna klasifikacija podataka, a te&scaron;ko je, ili čak nemoguće, doći do dovoljno velikog i raznovrsnog obučavajućeg skupa podataka<br />Metodologija &ndash; Modeli opisani u disertaciji se baziraju na kombinaciji ko-trening algoritma i tehnika učenja sa grupom hipoteza. Prvi korak jeste obučavanje grupe klasifikatora velike raznolikosti i kvaliteta. Sa ovim ciljem modeli eksploati&scaron;u primenu različitih konfiguracija ko-trening algoritma na isti skup podataka. Prednost ovog pristupa je mogućnost kori&scaron;ćenja značajno manjeg anotiranog obučavajućeg skupa za inicijalizaciju algoritma.<br />Skup nezavisno obučenih ko-trening klasifikatora se kreira generisanjem predefinisanog broja slučajnih podela obeležja polaznog skupa podataka. Nakon toga se, polazeći od istog inicijalnog obučavajućeg skupa, ali kori&scaron;ćenjem različitih kreiranih podela obeležja, obučava grupa ko-trening klasifikatora. Nakon ovoga, neophodno je kombinovati predikcije nezavisno obučenih klasifikatora.<br />Predviđena su dva načina kombinovanja predikcija. Prvi način se zasniva na klasifikaciji zapisa na osnovu većine glasova grupe ko-trening klasifikatora. Na ovaj način se daje predikcija za svaki od zapisa koji su pripadali grupi neanotiranih primera kori&scaron;ćenih u toku obuke ko-treninga. Potom se primenjuje genetski algoritam u svrhu selekcije najpouzdanije klasifikovanih zapisa ovog skupa. Konačno,<br />163<br />najpouzdanije klasifikovani zapisi se koriste za obuku finalnog klasifikatora. Ovaj finalni klasifikator se koristi za predikciju klase zapisa koje je neophodno klasifikovati. Opisani algoritam je nazvan Algoritam Statistike Slučajnih Podela (Random Split Statistics algorithm, RSSalg).<br />Drugi način kombinovanja nezavisno obučenih ko-trening klasifikatora se zasniva na GMM-MAPML tehnici estimacije tačnih klasnih obeležja na osnovu vi&scaron;estrukih obeležja pripisanih od strane različitih anotatora nepoznatog kvaliteta. U ovom algoritmu, nazvanom Integracija Vi&scaron;estrukih Ko-treninranih Klasifikatora (Integration of Multiple Co-trained Classifiers, IMCC), svaki od nezavisno treniranih ko-trening klasifikatora daje predikciju klase za svaki od zapisa koji je neophodno klasifikovati. U ovoj postavci se svaki od ko-trening klasifikatora tretira kao jedan od anotatora čiji je kvalitet nepoznat, a svakom zapisu, za koga je neophodno odrediti klasno obeležje, se dodeljuje vi&scaron;e klasnih obeležja. Na kraju se primenjuje GMM-MAPML tehnika, kako bi se na osnovu dodeljenih vi&scaron;estrukih klasnih obeležja za svaki od zapisa izvr&scaron;ila estimacija stvarnog klasnog obeležja zapisa.<br />Rezultati &ndash; U disertaciji su razvijena dva modela, Integracija Vi&scaron;estrukih Ko-treninranih Klasifikatora (IMCC) i Algoritam Statistike Slučajnih Podela (RSSalg), bazirana na ko-trening algoritmu, koja re&scaron;avaju zadatak automatske klasifikacije u slučaju nepostojanja dovoljno velikog anotiranog korpusa za obuku. Modeli predstavljeni u disertaciji dizajnirani su tako da omogućavaju primenu ko-trening algoritma na skupove podataka bez prirodne podele obeležja, kao i da unaprede njegove performanse. Modeli su na vi&scaron;e skupova podataka različite veličine, dimenzionalnosti i redudantnosti poređeni sa postojećim ko-trening alternativama. Pokazano je da razvijeni modeli na testiranim skupovima podataka postižu bolje performanse od testiranih ko-trening alternativa.<br />Praktična primena &ndash; Razvijeni modeli imaju &scaron;iroku mogućnost primene u svim domenima gde je neophodna klasifikacija podataka, a anotiranje podataka dugotrajno i skupo. U disertaciji je prikazana i primena razvijenih modela u nekoliko konkretnih<br />164<br />situacija gde su modeli od posebne koristi: detekcija subjektivnosti, vi&scaron;e-kategorijska klasifikacija i sistemi za davanje preporuka.<br />Vrednost &ndash; Razvijeni modeli su korisni u &scaron;irokom spektru domena gde je neophodna klasifikacija podataka, a anotiranje podataka dugotrajno i skupo. Njihovom primenom se u značajnoj meri smanjuje ljudski rad neophodan za anotiranje velikih skupova podataka. Pokazano je da performanse razvijenih modela prevazilaze performanse postojećih alternativa razvijenih sa istim ciljem relaksacije problema dugotrajne i mukotrpne anotacije velikih skupova podataka.</p> / <p>Aim &ndash; The research presented in this thesis is aimed towards the development of the system for automatic semi-supervised classification. The system is designed to be applicable on the broad spectrum of practical domains where automatic classification of data is needed but it is hard or impossible to obtain a large enough training set.<br />Methodology &ndash; The described models combine co-training algorithm with ensemble learning with the aim to overcome the problem of co-training application on the datasets without the natural feature split. The first step is to create the ensemble of co-training classifiers. For this purpose the models presented in this thesis apply different configurations of co-training on the same training set. Compared to existing similar approaches, this approach requires a significantly smaller initial training set.<br />The ensemble of independently trained co-training classifiers is created by generating a predefined number of random feature splits of the initial training set. Using the same initial training set, but different feature splits, a group of co-training classifiers is trained. The two models differ in the way the predictions of different co-training classifiers are combined.<br />The first approach is based on majority voting: each instance recorded in the enlarged training sets resulting from co-training application is classified by majority voting of the group of obtained co-training classifiers. After this, the genetic algorithm is applied in order to select the group of most reliably classified instances from this set. The most reliable instances are used in<br />167<br />order to train a final classifier which is used to classify new instances. The described algorithm is called Random Split Statistic Algorithm (RSSalg).<br />The other approach of combining single predictions of the group of co-training classifiers is based on GMM-MAPML technique of estimating the true hidden label based on the multiple labels assigned by multiple annotators of unknown quality. In this model, called the Integration of Multiple Co-trained Classifiers (IMCC), each of the independently trained co-training classifiers predicts the label for each test instance. Each co-training classifier is treated as one of the annotators of unknown quality and each test instance is assigned multiple labels (one by each of the classifiers). Finally, GMM-MAPML technique is applied in order to estimate the true hidden label in the multi-annotator setting.<br />Results &ndash; In the dissertation the two models are developed: the Integration of Multiple Co-trained Classifiers (IMCC) and Random Split Statistic Algorithm (RSSalg). The models are based on co-training and aimed towards enabling automatic classification in the cases where the existing training set is insufficient for training a quality classification model. The models are designed to enable the application of co-training algorithm on datasets that lack the natural feature split needed for its application, as well as with the goal to improve co-training performance. The models are compared to their co-training alternatives on multiple datasets of different size, dimensionality and feature redundancy. It is shown that the developed models exhibit superior performance compared to considered co-training alternatives.<br />Practical application &ndash; The developed models are applicable on the wide spectrum of domains where there is a need for automatic classification and training data is insufficient. The dissertation presents the successful application of models in several concrete situations where they are highly<br />168<br />beneficial: subjectivity detection, multicategory classification and recommender systems.<br />Value &ndash; The models can greatly reduce the human effort needed for long and tedious annotation of large datasets. The conducted experiments show that the developed models are superior to considered alternatives.</p>

Neste trabalho é proposta uma plataforma robótica móvel, concebida de maneira modular e hierárquica, visando o estudo de diversos aspectos aplicados à navegação, tanto autônoma quanto semi-autônoma, em ambientes internos. O sistema proposto possibilita a implementação de arquiteturas reativas e híbridas com aprendizagem, sendo a importância e limitações desta última discutidas. Utilizando a plataforma desenvolvida, uma aplicação de navegação robótica com aprendizagem supervisionada é realizada, usando sensores de ultra-som e através de tele-operação. O objetivo é fazer com que o agente associe, em tempo real, suas próprias respostas sensoriais com as ações motoras realizadas pelo tele-operador, permitindo que a tarefa seja repetida autonomamente com alguma generalização. Para realizar tal mapeamento, uma rede de função de base radial (RBF), usando um algoritmo de aprendizado seqüencial, é apresentada e utilizada. / This work presents a mobile robotic platform, built as a modular and hierarchical approach, aiming at the study of several aspects of indoor navigation. The proposed system allows the implementation of reactive and hybrid architectures with learning, for autonomous or semi-autonomous navigation. The importance and limitations of the learning characteristics are discussed. An application of robotic navigation with supervised learning is implemented using ultrasonic sensors and teleoperation. The aim is the agent to associate, in real time, its own sensorial perception to the motor actions realized by a teleoperator, allowing the task to be repeated in an autonomous way, with some generalization. To make the corresponding mapping, a radial basis function network (RBF), trained by a sequential learning algorithm, is presented and used.