There are basically three areas which can be explored to improve an HMM based recognizer, namely, parameter extraction, training methods and vocabulary representation. / The goal of parameter extraction is not only to find a compact and robust parametric representation of the speech signal, but also to find one which allows the HMMs to obtain the best possible recognition performance. Historically, improvements at this level have usually been obtained on a trial and error basis, using as much knowledge as possible about both the speech production and speech perception mechanisms. That is, the acoustic parameter extraction module has always been viewed as a separate module from the HMMs. This thesis will explore the concept of performing parameter extraction with a connectionist model, whose parameters can be learned from training data. / Two HMM training techniques are used in this thesis, namely MLE and MMIE. Parameter initialization, of critical importance for both, will be investigated for discrete, semi-continuous and continuous HMMs. Training processes involving a combination of MLE and MMIE training are studied. Other issues such as codebook exponents and the use of pause and silence models will also be explored. / Even if the vocabulary contains only 11 words, its representation is a very important issue. The effects of vocabulary representation with phoneme based, word based (with no sharing) and inter-word models will be experimentally evaluated. It will be shown how a word error rate of 0.23% and a string error rate of 0.68% can be achieved on the TIDIGITS corpus--a performance rivalling the best results ever reported by any group of researchers.

The development of algorithms to efficiently determine an optimal path through a complex environment is a continuing area of research within Computer Science. When such environments can be represented as a graph, established graph search algorithms, such as Dijkstra’s shortest path and A*, can be used. However, many environments are constructed from a set of regions that do not conform to a discrete graph. The Weighted Region Problem was proposed to address the problem of finding the shortest path through a set of such regions, weighted with values representing the cost of traversing the region. Robust solutions to this problem are computationally expensive since finding shortest paths across a region requires expensive minimisation. Sampling approaches construct graphs by introducing extra points on region edges and connecting them with edges criss-crossing the region. Dijkstra or A* are then applied to compute shortest paths. The connectivity of these graphs is high and such techniques are thus not particularly well suited to environments where the weights and representation frequently change. The Field D* algorithm, by contrast, computes the shortest path across a grid of weighted square cells and has replanning capabilites that cater for environmental changes. However, representing an environment as a weighted grid (an image) is not space-efficient since high resolution is required to produce accurate paths through areas containing features sensitive to noise. In this work, we extend Field D* to weighted simplicial complexes – specifically – triangulations in 2D and tetrahedral meshes in 3D. Such representations offer benefits in terms of space over a weighted grid, since fewer triangles can represent polygonal objects with greater accuracy than a large number of grid cells. By exploiting these savings, we show that Triangulated Field D* can produce an equivalent path cost to grid-based Multi-resolution Field D*, using up to an order of magnitude fewer triangles over grid cells and visiting an order of magnitude fewer nodes. Finally, as a practical demonstration of the utility of our formulation, we show how Field D* can be used to approximate a distance field on the nodes of a simplicial complex, and how this distance field can be used to weight the simplicial complex to produce contour-following behaviour by shortest paths computed with Field D*.

In the field of Evolutionary Robotics, the design, development and application of artificial neural networks as controllers have derived their inspiration from biology. Biologists and artificial intelligence researchers are trying to understand the effects of neural network learning during the lifetime of the individuals on evolution of these individuals by qualitative and quantitative analyses. The conclusion of these analyses can help develop optimized artificial neural networks to perform any given task. The purpose of this thesis is to study the effects of learning on evolution. This has been done by applying Temporal Difference Reinforcement Learning methods to the evolution of Artificial Neural Tissue controller. The controller has been assigned the task to collect resources in a designated area in a simulated environment. The performance of the individuals is measured by the amount of resources collected. A comparison has been made between the results obtained by incorporating learning in evolution and evolution alone. The effects of learning parameters: learning rate, training period, discount rate, and policy on evolution have also been studied. It was observed that learning delays the performance of the evolving individuals over the generations. However, the non zero learning rate throughout the evolution process signifies natural selection preferring individuals possessing plasticity.

In the field of Evolutionary Robotics, the design, development and application of artificial neural networks as controllers have derived their inspiration from biology. Biologists and artificial intelligence researchers are trying to understand the effects of neural network learning during the lifetime of the individuals on evolution of these individuals by qualitative and quantitative analyses. The conclusion of these analyses can help develop optimized artificial neural networks to perform any given task. The purpose of this thesis is to study the effects of learning on evolution. This has been done by applying Temporal Difference Reinforcement Learning methods to the evolution of Artificial Neural Tissue controller. The controller has been assigned the task to collect resources in a designated area in a simulated environment. The performance of the individuals is measured by the amount of resources collected. A comparison has been made between the results obtained by incorporating learning in evolution and evolution alone. The effects of learning parameters: learning rate, training period, discount rate, and policy on evolution have also been studied. It was observed that learning delays the performance of the evolving individuals over the generations. However, the non zero learning rate throughout the evolution process signifies natural selection preferring individuals possessing plasticity.

This study presents new techniques for the analysis of data acquired by focal plane array Fourier transform infrared (FPA-FTIR) spectroscopy. FPA-FTIR spectrometers are capable of acquiring several orders of magnitude more data than conventional FTIR spectrometers, necessitating the use of novel data analysis techniques to exploit the information-rich nature of these infrared imaging systems. The techniques investigated in this study are demonstrated in the context of bacteria identification by FPA-FTIR spectroscopy. Initially, an examination is made of the image fidelity of three FPA-FTIR instruments and demonstrates the high degree of within-image and between-image variability that is encountered with this technology. This is followed by a description of the development of pixel filtration routines that allow for the extraction of the most representative data from the infrared images of non-uniform samples. A genetic algorithm (GA) approach is introduced for determining the relevancy of spectral features in relation to bacterial classification and is compared to other forms of classifier optimizations. A proof-of-concept study demonstrating the potential use of infrared imaging to detect bacterial samples originating from a mixed culture is then presented. Finally, an overall methodology involving the combination of these data analysis techniques and including additional approaches towards the development, maintenance, and validation of databases based on infrared imaging data is described. This methodology has been developed with an emphasis on accessibility by implementing the elements of an expert system which allows for this technology to be employed by a non-technical user. / Cette étude présente une nouvelle approche d'analyse de données spectrales résultant de l'utilisation de la spectroscopie infrarouge à transformée de Fourier couplée à un détecteur type «matrice à plan focal» (FPA-FTIR) à balayage rapide. Les spectromètres FPA-FTIR ont une capacité de capture de données de plusieurs ordres de grandeur supérieurs aux spectromètres traditionnels et nécessitent donc des techniques avancées d'analyse de données pour exploiter cette mine d'information que représente l'imagerie infrarouge. La spectroscopie FPA-FTIR a été utilisée dans cette étude pour l'identification des bactéries. L'étape initiale, celle de la comparaison de trois spectromètres FPA-FTIR sur les points de vue fidélité de l'image, tant image-image qu'entre images, a révélé de grandes variabilités qui sont propres à cette technologie. Cette étape est suivie du développement de routines de filtration de pixels permettant d'extraire les données caractéristiques de l'imagerie infrarouge des échantillons non-uniformes. Un algorithme génétique (GA) est ensuite introduit pour déterminer la pertinence des caractéristiques spectrales sur le plan de la classification bactérienne et a été comparé à d'autres formes de classification optimisée. Une étude de démonstration de la capacité de la technologie d'imagerie infrarouge pour la détection des échantillons de bactéries provenant de cultures mixtes s'en est suivie. Pour terminer, une méthodologie globale combinant ces techniques d'analyse de données et incluant d'autres étapes telles le développement, la mise à niveau et la validation des bases de données d'imagerie infrarouge a été présentée. Cette méthodologie met l'emphase sur le développement et l'implantation d'un système expert accessible d'utilisation à de non-experts.

In this thesis, we analyze a context-dependent movie recommendation system using a Hierarchical Bayesian Network. Unlike most other recommender systems which either do not consider context or do so using collaborative filtering, our approach is content-based. This allows users to individually interpret contexts or invent their own contexts and continue to get good recommendations. By using a Hierarchical Bayesian Network, we can provide context recommendations when users have only provided a small amount of information about their preferences per context. At the same time, our model has enough degrees of freedom to handle users with different preferences in different contexts. We show on a real data set that using a Bayesian Network to model contexts reduces the error on cross-validation over models that do not link contexts together or ignore context altogether. / Dans cette thèse, nous analysons un système de recommandations de films dépendant du contexte en utilisant un réseau Bayésien hiérarchique. Contrairement à la plupart des systèmes de recommendations qui, soit ne considère pas le contexte, soit le considère en utilisant le filtrage collaboratif, notre approche est basée sur le contenu. Ceci permet aux utilisateurs d'interpréter les contextes individuellement ou d'inventer leurs propres contextes tout en obtenant toujours de bonnes recommandations. En utilisant le rèseau Bayésien hiérarchique, nous pouvons fournir des recommendations en contexte quand les utilisateurs n'ont fourni que quelques informations par rapport à leurs préférences dans différents contextes. De plus, notre modèle a assez de degrés de liberté pour prendre en charge les utilisateurs avec des préférences différentes dans différents contextes. Nous démontrons sur un ensemble de données réel que l'utilisation d'un réseau Bayésien pour modéliser les contextes réduit l'erreur de validation croisée par rapport aux modèles qui ne lient pas les contextes ensemble ou qui ignore tout simplement le contexte.

We explore the use of co-training to improve the performance of image classification in the setting where multiple classifiers are used and several types of features are available. Features are assigned to classifiers in an optimal manner using hierarchical clustering with a distance metric based on conditional mutual information. The effect of increasing the number of classifiers is then evaluated by co-training using the assigned feature sets. Experimental results indicate that the feature assignments chosen by the clustering approach afford superior co-training performance in comparison to other logical assignment choices. The results also indicate that increasing the number of classifiers beyond two leads to improved performance provided that the classifiers are sufficiently independent, and are reasonable well balanced in terms of labeling ability.Additionally, we explore the effect that the initial training set selectionhas on co-training performance. We find that the quality of training imageshas a profound effect on performance and provide recommendations for howbest to select these images. / Nous explorons l'utilisation de la co-formation pour améliorer la performance de classification d'image dans un milieu où multiples classificateurs s'emploient et plusieurs types de caractéristiques sont disponibles. Les caractéristiques sont associés aux classificateurs d'une manière optimal en employant le groupage hiérarchique avec une mesure de distance basée sur l'information mutuelle conditionnelle. L'effet d'augmenter le nombre de classificateurs est alors evalué par la co-formation, en employant les ensembles de caractéristiques attribués. Les résultats de nos expériences indique que si on augmente le nombre de classificateurs au-delà de deux, la performance s'améliore pourvu que les caractéristiques soient suffisamment indépendantes et assez bien équilibrées en termes de compétence d'étiquetage. En plus, nous explorons l'effet de l'ensemble choisi pour l'entraînement initial sur la performance en co-formation. Nous trouvons que la qualité d'images dans l'entraînement a un effet profond sur la performance, et nous fournissons des recommandations sur comment sélectionner ces images pour le meilleur effet.

Since the emergence of the object technology, organizations have accumulated a tremendous amount of object-oriented (OO) code. Instead of continuing to recreate components similar to existing artifacts, and considering the rising costs of development, many organizations would like to decrease software development costs and cycle time by reusing existing OO components. The difficulty of finding reusable components is that reuse is a complex and thus less quantifiable measure. In this research, we first proposed three reuse hypotheses about the impact of three internal characteristics (inheritance, coupling, and complexity) of OO software artifacts on reusability. Corresponding metrics suites were then selected and extracted. We used C4.5, a machine learning algorithm, to build predictive models from the learning data set that we have obtained from a medium sized software system developed in C++. Each predictive models was then verified according to its completeness, correctness and global accuracy. The verification results proved that the proposed hypotheses were correct. The uniqueness of this research work is that we have combined the state of the art of three different subjects (reuse detection and prediction, OO metrics and their extraction, and applied machine learning algorithm) to form a process of finding interesting properties of OO software components that affect reusability.

Reinforcement Learning has emerged as a useful framework for learning to perform a task optimally from experience in unknown systems. A major problem for such learning algorithms is how to balance optimally the exploration of the system, to gather knowledge, and the exploitation of current knowledge, to complete the task. Model-based Bayesian Reinforcement Learning (BRL) methods provide an optimal solution to this problem by formulating it as a planning problem under uncertainty. However, the complexity of these methods has so far limited their applicability to small and simple domains. To improve the applicability of model-based BRL, this thesis presents several extensions to more complex and realistic systems, such as partially observable and continuous domains. To improve learning efficiency in large systems, this thesis includes another extension to automatically learn and exploit the structure of the system. Approximate algorithms are proposed to efficiently solve the resulting inference and planning problems. / L'apprentissage par renforcement a émergé comme une technique utile pour apprendre à accomplir une tâche de façon optimale à partir d'expérience dans les systèmes inconnus. L'un des problèmes majeurs de ces algorithmes d'apprentissage est comment balancer de façon optimale l'exploration du système, pour acquérir des connaissances, et l'exploitation des connaissances actuelles, pour compléter la tâche. L'apprentissage par renforcement bayésien avec modèle permet de résoudre ce problème de façon optimale en le formulant comme un problème de planification dans l'incertain. La complexité de telles méthodes a toutefois limité leur applicabilité à de petits domaines simples. Afin d'améliorer l'applicabilité de l'apprentissage par renforcement bayésian avec modèle, cette thèse presente plusieurs extensions de ces méthodes à des systèmes beaucoup plus complexes et réalistes, où le domaine est partiellement observable et/ou continu. Afin d'améliorer l'efficacité de l'apprentissage dans les gros systèmes, cette thèse inclue une autre extension qui permet d'apprendre automatiquement et d'exploiter la structure du système. Des algorithmes approximatifs sont proposés pour résoudre efficacement les problèmes d'inference et de planification résultants.

An original way to define, analyze and design mechanical systems with inherently lifelike dynamic properties is presented. The construction of robotic manipulators which embody a complete set of technologically relevant biological principles is outlined. The ultimate objective is to develop a new class of mobile, autonomous, and interactive machines which dynamically emulate live musculoskeletal systems. / This study introduces the mathematical models and algorithms to transform and synthesize the results of research in musculoskeletal physiology into explicit engineering design specifications. The application of a new contractile muscle-like viscoelastic motor, as a servomechanical drive for articulated rigid link mechanisms as well as for a novel flexible trunk-like manipulator, is investigated. Key features of the neuromuscular force control by twitch summation are combined to formulate a pulse stream control method suitable for fluid powered mechanisms.