Artificial neural networks have primarily been utilized to solve problems in pattern recognition, decision-making, signal analysis and controls. This thesis investigates the use of networks in modeling physical systems in fluid mechanics by using the governing partial differential equations to initialize the network parameters. The initialization requires imposing certain constraints on the values of the input, bias, and output weights. The attribution of certain roles to each of these parameters allows for mapping a polynomial approximation into an artificial neural network architecture. This approach is shown to be capable of incorporating smooth neuron transfer functions, such as the popular hyperbolic tangent. Attention is focused on the two-dimensional Navier-Stokes equations for Boussinesq convection that model two-dimensional double diffusive convection. The network used to model this example utilizes an approximation of the Gudermannian function and an application of the pseudospectral method for complete network initiation. Numerical examples are presented illustrating the accuracy and utility of the method.

Closed kinematic chains (CKCs) are constrained multibody systems that contain closed kinematic loops. Nowadays, CKCs are used in a variety of applications ranging from flight simulators to medical instruments, and are becoming increasingly popular in the machine-tool industry and haptic interfaces due to their better performance in terms of accuracy, rigidity and payload capacity as compared to open-chain mechanisms. This document intends to present a novel methodology for the modeling and control of general CKCs. The dynamics of CKCs are characterized by index-3 differential algebraic equations (DAEs). Dynamic models in the form of DAEs pose difficulties in model-based control because most existing control design techniques are devised for explicit state space models. The control methodology presented in this document is based on a singular perturbation formulation (SPF), which has attractive properties including the minimum dimension of its slow dynamics and the large validity domain that contains the entire singularity-free workspace of the CKCs. The key issue of the model approximation error is addressed under different stability conditions. Explicit error bounds are derived and sufficient conditions for the exponential convergence of the approximation errors are established. For the control of CKCs, our approach transfers the control of the original DAE system to the control of an artificially created singularly perturbed system. Compared to control methods which directly solve the nonlinear algebraic constraint equations, the proposed method uses an ODE solver to obtain the dependent coordinates, hence eliminating the need for Newton type iterations and is amenable to real-time implementation. The closed loop system, when controlled by typical open kinematic chain schemes, achieves asymptotic trajectory tracking. The efficacy of the approach is illustrated by simulating the dynamics of a CKC mechanism, the Rice Planar Delta Robot, and then by validating the simulation results with experimental data. Thus, this work establishes a framework in which the control of CKCs can be systematically addressed.

No description available.

This thesis presents a new methodology for automatically learning an optimal neurostimulation strategy for the treatment of epilepsy. The technical challenge is to automatically modulate neurostimulation parameters, as a function of the observed field potential recording, so as to minimize the frequency and duration of seizures. The methodology leverages recent techniques from the machine learning literature, in particular the reinforcement learning paradigm, to formalize this optimization problem. We present an algorithm which is able to learn an adaptive neurostimulation strategy directly from labeled training data acquired from animal brain tissues. Our results suggest that this methodology can be used to automatically find a stimulation strategy which effectively reduces the incidence of seizures, while also minimizing the amount of stimulation applied. This work highlights the crucial role that modern machine learning techniques can play in the optimization of treatment strategies for patients with chronic disorders such as epilepsy. / Cette thèse présente une nouvelle méthodologie pour apprendre, de façon automatique, une stratégie optimale de neurostimulation pour le traitement de l'épilepsie. Le défi technique est de moduler automatiquement les paramètres de stimulation, en fonction de l'enregistrement de potentiel de champ observé, afin de minimiser la fréquence et la durée des crises d'épilepsie. Cette méthodologie fait appel à des techniques récentes développées dans le domaine de l'apprentissage machine, en particulier le paradigme d'apprentissage par renforcement, pour formaliser ce problème d'optimisation. Nous présentons un algorithme qui est capable d'apprendre une stratégie adaptative de neurostimulation, et ce directement à partir de données d'apprentissage, étiquetées, acquises depuis des tissus de cerveaux d'animaux. Nos résultats suggèrent que cette méthodologie peut être utiliser pour trouver, automatiquement, une stratégie de stimulation qui réduit efficacement l'indicence des crises d'épilepsie tout en minimisant le nombre de stimulations appliquées. Ce travail met en évidence le rôle crucial que les techniques modernes d'apprentissage machine peuvent jouer dans l'optimisation de stratégies de traitements pour des patients souffrant de maladies chroniques telle l'épilepsie.

This thesis describes a research effort directed at producing a computational model based on artificially intelligent cellular automata. This model was developed for the purpose of learning a mapping from an input space to an output space. A specific problem that occurs in the mining industry was used to develop and test the model's ability to learn the mapping between a three-dimensional input volume and a three-dimensional output volume. In this case, the mapping was a consequence of the industrial processes used in mining as well as the properties of the material being mined. / Three main computational tools were combined in this work to form the complete mine stope prediction model. The three modules are a learning module, an optimisation module, and an overall network architecture. The overall network architecture is a 3-D lattice of cellular automata (CA) and has the capability to implicitly capture the complexities in shape that render other types of models arduous or inapplicable. The learning module uses a Discrete Time Cellular Neural Network (DTCNN) to store and recall information about a given mapping. The optimisation module uses the Simulated Annealing (SA) algorithm to perform a non-linear optimisation on the set of weights used by the DTCNN. / Variations of the model, and different experiments, were performed to test and explore the model in depth. Concepts such as "Small-Worlds" and "Forgetting Factor" were investigated. The applicability of a Partial Least Squares (PLS) model as an alternative to the DTCNN transition rule was also explored.

Conventional computer speech recognition systems use models of speech acoustics and the language of the recognition task in order to perform recognition. For all but trivial recognition tasks, sub-word units are modeled, typically phonemes. Recognizing words then requires a pronunciation dictionary ( PD) to specify how each word is pronounced in terms of the units modeled. Even if the acoustic modeling component is perfect, the recognizer will still be prone to misrecognition, most often because the speaker can use a pronunciation other than that in the PD. This different pronunciation may be due to the speaker being a non-native speaker of the language being recognized, having 'mispronounced' the word, coarticulatory effects, recognizer errors in phoneme hypothesization, or any combination of these. One way to overcome these misrecognitions is to use a dynamic PD, able to acquire new pronunciations for words as they are encountered and misrecognized. The thesis examines the following questions: can automated methods be found that produce reliable alternate pronunciations? If so, does augmenting a PD (which originally contains only canonical pronunciations) with these alternate pronunciations lead to improved recognizer performance? It shows that using even simple methods, average reductions in word error rate of at least 45% are possible, even with speakers who are not native speakers of the recognition task language.

This thesis addresses the problem of automatically constructing a visual representation of an unknown environment that is useful for robotic navigation, localization and exploration. There are two main contributions. First, the concept of the visual map is developed, a representation of the visual structure of the environment, and a framework for learning this structure is provided. Second, methods for automatically constructing a visual map are presented for the case when limited information is available about the position of the camera during data collection. / The core concept of this thesis is that of the visual map, which models a set of image-domain features extracted from a scene. These are initially selected using a measure of visual saliency, and subsequently modelled and evaluated for their utility for robot pose estimation. Experiments are conducted demonstrating the feature learning process and the inferred models' reliability for pose inference. / The second part of this thesis addresses the problem of automatically collecting training images and constructing a visual map. First, it is shown that visual maps are self-organizing in nature, and the transformation between the image and pose domains is established with minimal prior pose information. Second, it is shown that visual maps can be constructed reliably in the face of uncertainty by selecting an appropriate exploration strategy. A variety of such strategies are presented and these approaches are validated experimentally in both simulated and real-world settings.

Object-Oriented (OO) technology claims faster development and higher quality of software than the procedural paradigm. The quality of the product is the single most important reason that determines its acceptance and success. The basic project management problem is "delivery of a product with targeted quality, within the budget, and on schedule". We propose a state-of-the-art approach that gets closer to the solution by improving the software development process used. An important objective in all software development is to ensure that the delivered product is as fault-free as possible. We proposed three hypotheses that relate the OO design properties---inheritance, cohesion, and coupling---and the fault-proneness as software's quality indicator. We built classification models that predict which components are likely to be faulty, based on an appropriate suite of OO design measures. The models represent empirical evidence that the aforementioned relationships exist. We used the C4.5 machine learning algorithm as a predictive modeling technique, because it is robust, reliable, and allows intelligible interpretation of the results. We defined three new measures that quantify the specific contribution of each of the metrics selected by the model(s), and also provide a deeper insight into the design structure of the product. We evaluated the quality of the predictive models using an objective set of standards. The models built have high quality.

In this thesis I examine a delay differential equation model for an artificial neural network with two neurons. Linear stability analysis is used to determine the stability region of the stationary solutions. They can lose stability through either a pitchfork or a supercritical Hopf bifurcation. It is shown that, for appropriate parameter values, an interaction takes place between the pitchfork and Hopf bifurcations. Conditions are found under which the set of initial conditions that converge to a stable stationary solution is open and dense in the function space. Analytic results are illustrated with numerical simulations.

An analytical approach is obviously practical only when we want to study nonlinear systems of low complexity. An alternative for more complex processes that has raised a lot of interest in recent years relies on Artificial Neural Networks (ANNs). / In this work we will explore an alternative avenue to the problems of control and identification, where Cellular Automata (CAs) will be considered in place of ANNs. CAs not only share ANNs' most valuable characteristics but they also have interesting characteristics of their own, for a structurally simpler architecture. CAs applications so far have been mainly restrained to simulating natural phenomena occuring in a finite homogeneous space. / Concepts relevant to the problems of control and identification will be introduced in the first part of our work. CAs will then be introduced, with a discussion of the issues raised by their application in the context, A working prototype of a CA-based controller is introduced in the last part of the work, that confirms the interest of using CAs to address the problem of nonlinear adaptive control. (Abstract shortened by UMI.)