Ces dernières années, des recherches ont été menées sur plusieurs techniques reliées à la collection des déchets. Plusieurs découvertes centrales pour le ramassage de miettes par copie ont été réalisées. Cependant, des améliorations sont encore possibles. Dans ce mémoire, nous introduisons des nouvelles techniques et de nouveaux algorithmes pour améliorer le ramassage de miettes. En particulier, nous introduisons une technique utilisant des cadres délimités pour marquer et retracer les pointeurs racines. Cette technique permet un calcul efficace de l'ensemble des racines. Elle réutilise des concepts de deux techniques existantes, card marking et remembered sets, et utilise une configuration bidirectionelle des objets pour améliorer ces concepts en stabilisant le surplus de mémoire utilisée et en réduisant la charge de travail lors du parcours des pointeurs. Nous présentons aussi un algorithme pour marquer récursivement les objets rejoignables sans utiliser de pile (éliminant le gaspillage de mémoire habituel). Nous adaptons cet algorithme pour implémenter un ramasse-miettes copiant en profondeur et améliorer la localité du heap. Nous améliorons l'algorithme de collection des miettes older-first et sa version générationnelle en ajoutant une phase de marquage garantissant la collection de toutes les miettes, incluant les structures cycliques réparties sur plusieurs fenêtres. Finalement, nous introduisons une technique pour gérer les gros objets. Pour tester nos idées, nous avons conçu et implémenté, dans la machine virtuelle libre Java SableVM, un cadre de développement portable et extensible pour la collection des miettes. Dans ce cadre, nous avons implémenté des algorithmes de collection semi-space, older-first et generational. Nos expérimentations montrent que la technique du cadre délimité procure des performances compétitives pour plusieurs benchmarks. Elles montrent aussi que, pour la plupart des benchmarks, notre algorithme de parcours en profondeur améliore la localité et augmente ainsi la performance. Nos mesures de la performance générale montrent que, utilisant nos techniques, un ramasse-miettes peut délivrer une performance compétitive et surpasser celle des ramasses-miettes existants pour plusieurs benchmarks. ______________________________________________________________________________ MOTS-CLÉS DE L’AUTEUR : Ramasse-Miettes, Machine Virtuelle, Java, SableVM.

Le projet porte sur la communication homme-machine. Homme et machine sont pris comme deux espaces cognitifs distincts, naturel et artificiel. Quelle perception ont-ils l'un de l'aulre? Comment interagissent-ils? L'expression sonore constitue un point commun aux deux espèces. Cnidaria est une installation robotique interactive et immersive. Une communauté de robots occupent un territoire public. Le public peut interagir avec l'installation et entrer dans la représentation en usant d'un langage rudimentaire. Le spectacle se crée par la dynamique entre les interacteurs et un système informatique autonome et génératif

Les systèmes tutoriels intelligents (STI) sont un grand pas vers une réforme dans l'éducation. Ces systèmes offrent une souplesse d'enseignement que les autres aides pédagogiques informatiques n'ont pas. De ce fait, ils pourraient, s'ils sont bien intégrés dans un programme éducatif, décharger les professeurs pour qu'ils consacrent une attention particulière aux étudiants plus faibles. Les systèmes tutoriels intelligents atteignent cette souplesse grâce à la combinaison de sous-systèmes; l'un d'entre eux est la communication. Plusieurs recherches ont été effectuées dans ce sens notamment pour la communication en langage naturel. Cette communication peut être divisée en trois parties soit la compréhension du langage naturel, la génération de texte en langage naturel et la planification des dialogues. Cette dernière représente la base de ce type de communication. CTS (Cognitive Tutoring System) est un moteur de système tutoriel intelligent basé sur la conscience d'accès développée par le GDAC. CTS a été intégré à Canadarm Tutor pour son développement. Ce mémoire traite de l'ajout d'un système de planification du dialogue basé sur les travaux effectués sur Beetle. Dans un premier temps, plusieurs correctifs seront apportés au fonctionnement du Réseau des Actes pour tenter de stabiliser son comportement; d'autres amèneront le système plus près de ses fondements notamment l'ajout de la délibération. L'ajout du planificateur tel que décrit dans le STI de Beetle s'effectuera dans un second temps et utilisera l'architecture unique de CTS pour le faire. Cette combinaison d'architecture apportera plusieurs avantages et donnera un système de planification de dialogue générique et augmentable. ______________________________________________________________________________

In this thesis I present research in two fields: machine learning and computational biology. First, I develop new machine learning methods for graphical models that can be applied to protein problems. Then I apply graphical model algorithms to protein problems, obtaining improvements in protein structure prediction and protein structure alignment. First,in the machine learning work, I focus on a special kind of graphical model---conditional random fields (CRFs). Here, I present a new semi-supervised training procedure for CRFs that can be used to train sequence segmentors and labellers from a combination of labeled and unlabeled training data. Such learning algorithms can be applied to protein and gene name entity recognition problems. This work provides one of the first semi-supervised discriminative training methods for structured classification. Second, in my computational biology work, I focus mainly on protein problems. In particular, I first propose a tree decomposition method for solving the protein structure prediction and protein structure alignment problems. In so doing, I reveal why tree decomposition is a good method for many protein problems. Then, I propose a computational framework for detection of similar structures of a target protein with sparse NMR data, which can help to predict protein structure using experimental data. Finally, I propose a new machine learning approach---LS_Boost---to solve the protein fold recognition problem, which is one of the key steps in protein structure prediction. After a thorough comparison, the algorithm is proved to be both more accurate and more efficient than traditional z-Score method and other machine learning methods.

We analyze videos in which a hand interacts with a basketball. In this work, we present a computational system which detects and classifies hand-ball events, given the trajectories of a hand and ball. Our approach is to determine non-gravitational parts of the ball's motion using only the motion of the hand as a reliable cue for hand-ball events. This thesis makes three contributions. First, we show that hand motion can be segmented using piecewise fifth-order polynomials inspired by work in motor control. We make the remarkable experimental observation that hand-ball events have a phenomenal correspondence to the segmentation breakpoints. Second, by fitting a context-dependent gravitational model to the ball over an adaptive window, we can isolate places where the hand is causing non-gravitational motion of the ball. Finally, given a precise segmentation, we use the measured velocity steps (force impulses) on the ball to detect and classify various event types.

Uptime and maintenance planning are important issues for vehicle operators (e.g.operators of bus fleets). Unplanned downtime can cause a bus operator to be fined if the vehicle is not on time. Supervised classification methods for detecting faults in vehicles are compared in this thesis. Data has been collected by a vehicle manufacturer including three kinds of faulty states in vehicles (i.e. charge air cooler leakage, radiator and air filter clogging). The problem consists of differentiating between the normal data and the three different categories of faulty data. Evaluated methods include linear model, neural networks model, 1-nearest neighbor and random forest model. For every kind of model, a variable selection method should be used. In our thesis we try to find the best model for this problem, and also select the most important input signals. After we compare these four models, we found that the best accuracy (96.9% correct classifications) was achieved with the random forest model.

With malware becoming more and more diffused and at the same time more sophisticated in its attack techniques, countermeasures need to be set up so that new kinds of threats can be identified and dismantled in the shortest possible time, before they cause harm to the system under attack. With new behaviour patterns like the one shown by polymorphic and metamorphic viruses, static analysis is not any more a reliable way to detect those threats, and behaviour analysis seems a good candidate to fight against the next-generation families of viruses. In this project, we describe a methodology to analyze and categorize binaries solely on the basis of their behaviour, in terms of their interaction with the Operating System, other processes and network. The approach can strengten host-based intrusion detection systems by a timely classification of unkown but similar malware code. It has been evaluated on a dataset from the research community and tried on a smaller data set from local companies collected at University of Mondragone.

In this thesis I present research in two fields: machine learning and computational biology. First, I develop new machine learning methods for graphical models that can be applied to protein problems. Then I apply graphical model algorithms to protein problems, obtaining improvements in protein structure prediction and protein structure alignment. First,in the machine learning work, I focus on a special kind of graphical model---conditional random fields (CRFs). Here, I present a new semi-supervised training procedure for CRFs that can be used to train sequence segmentors and labellers from a combination of labeled and unlabeled training data. Such learning algorithms can be applied to protein and gene name entity recognition problems. This work provides one of the first semi-supervised discriminative training methods for structured classification. Second, in my computational biology work, I focus mainly on protein problems. In particular, I first propose a tree decomposition method for solving the protein structure prediction and protein structure alignment problems. In so doing, I reveal why tree decomposition is a good method for many protein problems. Then, I propose a computational framework for detection of similar structures of a target protein with sparse NMR data, which can help to predict protein structure using experimental data. Finally, I propose a new machine learning approach---LS_Boost---to solve the protein fold recognition problem, which is one of the key steps in protein structure prediction. After a thorough comparison, the algorithm is proved to be both more accurate and more efficient than traditional z-Score method and other machine learning methods.

We analyze videos in which a hand interacts with a basketball. In this work, we present a computational system which detects and classifies hand-ball events, given the trajectories of a hand and ball. Our approach is to determine non-gravitational parts of the ball's motion using only the motion of the hand as a reliable cue for hand-ball events. This thesis makes three contributions. First, we show that hand motion can be segmented using piecewise fifth-order polynomials inspired by work in motor control. We make the remarkable experimental observation that hand-ball events have a phenomenal correspondence to the segmentation breakpoints. Second, by fitting a context-dependent gravitational model to the ball over an adaptive window, we can isolate places where the hand is causing non-gravitational motion of the ball. Finally, given a precise segmentation, we use the measured velocity steps (force impulses) on the ball to detect and classify various event types.

Municipalities and contractors in Canada and other parts of the world rely on road surface condition information during and after a snow storm to optimize maintenance operations and planning. With an ever increasing demand for safer and more sustainable road network there is an ever increasing demand for more reliable, accurate and up-to-date road surface condition information while working with the limited available resources. Such high dependence on road condition information is driving more and more attention towards analyzing the reliability of current technology as well as developing new and more innovative methods for monitoring road surface condition. This research provides an overview of the various road condition monitoring technologies in use today. A new machine vision based mobile road surface condition monitoring system is proposed which has the potential to produce high spatial and temporal coverage. The proposed approach uses multiple models calibrated according to local pavement color and environmental conditions potentially providing better accuracy compared to a single model for all conditions. Once fully developed, this system could potentially provide intermediate data between the more reliable xed monitoring stations, enabling the authorities with a wider coverage without a heavy extra cost. The up to date information could be used to better plan maintenance strategies and thus minimizing salt use and maintenance costs.