Machine-learning algorithms are applied to explore the relation between significant flares and their associated CMEs. The NGDC flares catalogue and the SOHO/LASCO CME catalogue are processed to associate X and M-class flares with CMEs based on timing information. Automated systems are created to process and associate years of flare and CME data, which are later arranged in numerical-training vectors and fed to machine-learning algorithms to extract the embedded knowledge and provide learning rules that can be used for the automated prediction of CMEs. Properties representing the intensity, flare duration, and duration of decline and duration of growth are extracted from all the associated (A) and not-associated (NA) flares and converted to a numerical format that is suitable for machine-learning use. The machine-learning algorithms Cascade Correlation Neural Networks (CCNN) and Support Vector Machines (SVM) are used and compared in our work. The machine-learning systems predict, from the input of a flare¿s properties, if the flare is likely to initiate a CME. Intensive experiments using Jack-knife techniques are carried out and the relationships between flare properties and CMEs are investigated using the results. The predictive performance of SVM and CCNN is analysed and recommendations for enhancing the performance are provided. / EPSRC

Thesis (doctorat) -- Universität Ulm, 2005. / Includes bibliographical references (p. 227-248).

Wire Electrical Discharge Machining (WEDM) is a nontraditional machining process used for machining intricate shapes in high strength and temperature resistive (HSTR) materials. WEDM provides high accuracy, repeatability, and a better surface finish; however the tradeoff is a very slow machining rate. Due to the slow machining rate in WEDM, machining tasks take many hours depending on the complexity of the job. Because of this, users of WEDM try to predict machining rate beforehand so that input parameter values can be pre-programmed to achieve automated machining. However, partial success with traditional methodologies such as thermal modeling, artificial neural networks, mathematical, statistical, and empirical models left this problem still open for further research and exploration of alternative methods. Also, earlier efforts in applying the decision tree rule induction algorithms for predicting the machining rate in WEDM had limitations such as use of coarse grained method of discretizing the target and exploration of only C4.5 as the learning algorithm. The goal of this dissertation was to address the limitations reported in literature in using decision tree rule induction algorithms for WEDM. In this study, the three decision tree inductive algorithms C5.0, CART and CHAID have been applied for predicting material removal rate when the target was discretized into varied number of classes (two, three, four, and five classes) by three discretization methods. There were a total of 36 distinct combinations when learning algorithms, discretization methods, and number of classes in the target are combined. All of these 36 models have been developed and evaluated based on the prediction accuracy. From this research, a total of 21 models found to be suitable for WEDM that have prediction accuracy ranging from 71.43% through 100%. The models indentified in the current study not only achieved better prediction accuracy compared to previous studies, but also allows the users to have much better control over WEDM than what was previously possible. Application of inductive learning and development of suitable predictive models for WEDM by incorporating varied number of classes in the target, different learning algorithms, and different discretization methods have been the major contribution of this research.

This paper deals with the application of speed variable pumps in industrial hydraulic systems. The benefit of the natural feedback of the load torque is investigated for the issue of condition monitoring as the development of losses can be taken as evidence of faults. A new approach is proposed to improve the fault detection capabilities by tracking the changes via machine learning techniques. The presented algorithm is an art of adaptive modeling of the torque balance over a range of steady operation in fault free behavior. The aim thereby is to form a numeric reference with acceptable accuracy of the unit used in particular, taking into consideration the manufacturing tolerances and other operation conditions differences. The learned model gives baseline for identification of major possible abnormalities and offers a fundament for fault isolation by continuously estimating and analyzing the deviations.

<p>  </p> <p>The research described in this dissertation began in response to frequent questions from users of several crosswalks near a university campus. At each crosswalk was a sign indicating that motorists should yield to pedestrians in the crosswalk. That this message was not being interpreted uniformly was a concern at locations where heterogeneous road users (pedestrians, cyclists, and motorists) were interacting. Instead of trying to impose a single interpretation on users of each crosswalk, it was decided to observe and analyze interactions between users of the crosswalk. </p> <p>Several hours of video were recorded of pedestrians and motorists “negotiating” the right of way at the crosswalk. Because these crossing locations were marked but not signalized, they were called “semi-controlled crosswalks”. The negotiations took place during what were called pedestrian-motorist interactions (PMIs). The PMIs observed on video can be characterized as a “zebra-crossing” game, as described in Chapter 4 of this dissertation. </p> <p>Recently, computer vision (CV) algorithms have been extensively used in road users’ detection and tracking at an unparalleled spatial-temporal scale. In this study, CV algorithms have been applied to convert the video recordings into a large-scale spatial-temporal trajectory dataset including 800 pedestrians and cyclists interacting with more than 500 vehicles. Utilizing the trajectory dataset, a spatial-temporal graph convolutional network-based sequence to sequence (ST-GCN-Seq2Seq) algorithm has been developed to reasonably forecast heterogeneous road users’ trajectories and behavior in real time. Combining CV and ST-GCN-Seq2Seq algorithms can help both design an intelligent tracking system and achieve a form of “smart” interaction at semi-controlled crosswalks for heterogeneous road users.</p> <p>Based on road users’ arrival patterns detected from CV algorithms, it is likely that a "smart" control strategy can minimize the delay of pedestrians and motorists at crosswalks.  Therefore, another branch of this study is to investigate the “smart” control strategies at crosswalks using traffic signal controllers. A reinforcement learning framework was proposed as the “smart” control strategy, and several experiments were conducted using microsimulation. The proposed reinforcement learning framework is able to reduce traffic delay (efficiency), considering real-time pedestrian flow rates and vehicle flow rates with appropriate sensors.</p>

L’objectif de cette thèse est de développer, d’intégrer et de tester une nouvelle approche algorithmique d’aide au stationnement dans les centres urbains. Considérons différents types d’infrastructures déployées allant de la détection des entrées/sorties des véhicules jusqu'à la variation dans le temps du nombre de places de stationnement disponibles dans chaque portion de rue, nous montrons qu’il est possible de proposer une méthode efficace qui détermine un itinéraire qui minimise l’espérance de temps pour trouver une place de stationnement disponible et de prédire la disponibilité des placesde stationnement.Pour cela, la zone urbaine choisie sera donc considérée comme un ensemble de ressources de stationnement (segments de rues).Nous modélisons d’abord cette zone urbaine par un graphe où les sommets désignent les carrefours et les arcs représentent les portions de rues. Les paramètres essentiels pour notre modèle de réseau urbain sont la capacité de stationnement et le temps de parcours des portions de rue.L’originalité et l’aspect innovant de notre approche s’appuient sur deux principes.Le premier principe concerne le guidage comme une ressource : il ne s’agit pas de guider vers une place libre mais de proposer un parcours qui optimise l’espérance de temps de trouver une telle place. Pour cela nous déterminons dans une zone centrée sur une destination donnée, le parcours à effectuer par un véhicule pour minimiser son espérance de temps de trouver une place destationnement le plus rapidement possible.Ainsi nous avons mis en œuvre un algorithme d’apprentissage par renforcement basée sur la méthode LRI (Linear Reward Inaction) et la méthode Monte Carlo pour minimiser l’espérance de temps de trouver une place de stationnement en zone urbaine.Nous avons comparé cet algorithme avec une approche globale basée sur l’évaluation arborescente à profondeur bornée.Le second principe repose sur la prédiction des places de stationnement disponibles par périodes de temps homogènes où on ne s’intéresse pas à une place de stationnement en temps réel mais aux places de stationnement par zones. Il s’agit alors pour le système de pouvoir prédire le potentiel de places libres dans chacune des ressources pour les prochaines périodes. On ne vise donc pas ici la prédiction de la disponibilité de chaque place ; chaque ressource sera considérée comme une zone de stockage dont la disponibilité sera établie en grande partie en fonction des flux d’entrée et de sortie de la portion. Pour ce principe, nous avons donc déterminé par algorithmes de calculs et d’apprentissages la probabilité qu’il y ait au moins une place libre pour stationner dans un tronçon de rue pour un créneau de temps donné. Les principales données nécessaires pour effectuer ces calculs sont les séries temporelles d’entrée sortie de chaque véhicule aux intersections des rues et les variations des places de stationnement au cours du temps.Nous avons évalué les performances de notre approche par simulations sur des données générées aléatoirement et des données réelles obtenues sur un quartier de Versailles. / The objective of this thesis is to develop, to integrate and to test a new algorithmic approach to help parking in urban centers.Given the different types of deployed infrastructure : from input-output detection of vehicles to time variation of the number of available places within each street segment, we propose an efficient method to determine an itinerary that minimize the time expectation to find an available place and also to predict the availability of the parking places.We have chosen an urban area and we have considered it as a set of parking resources called street segments. More exactly, this urban area is considered as a graph where the vertexes represent the crossroads and the arcs represent the street segments. The essential parameters of our urban area model are the parking capacity and the time crossing of each street segment. The originality and the innovation of our approach are based on two principles.The first one is the guidance as a resource, i.e., it means that the proposed itinerary is not the one that lead to an available parking place but rather the one that minimized the time expectation to find an available parking place. In order to achieve that we determine, in a an area centered on a given destination, the itinerary to follow by the vehicle in order minimize its time expectation to find an available parking place as quickly aspossible.We have designed and realized a reinforcement learning algorithm based on the LRI method (Linear Reward Inaction) and a Monte Carlo method to minimize the time expectation to find an available parking place in the urban area. We have compared this algorithm to a global approach based on tree evaluation with bounded depth. The second principle is based on the prediction of the parking places by homogeneous time period where we are not interestedon a parking place in real time but rather on the parking places byarea. In other terms, the system predict the potential available parkingplaces by resource for the next time periods. Thus, we don’t aim to predict the availability of each parking place, i.e., each resource is considered as stock area and its availability is assessed in major part in function of the street segment input-output flow. For this principle, we have determined by a learning algorithm the probability that there is at least one available parking place in a street segment within a given time. The major data needed to compute this probability are the time series of input-output of each vehicle in street intersections, and the variation of the available parking places through the time.We have evaluated the performance of this approach by simulation based on random generated data and on real data of a district in Versailles.

This thesis studies the generalization ability of machine learning algorithms in a statistical setting. It focuses on the data-dependent analysis of the generalization performance of learning algorithms in order to make full use of the potential of the actual training sample from which these algorithms learn.¶ First, we propose an extension of the standard framework for the derivation of generalization bounds for algorithms taking their hypotheses from random classes of functions. This approach is motivated by the fact that the function produced by a learning algorithm based on a random sample of data depends on this sample and is therefore a random function. Such an approach avoids the detour of the worst-case uniform bounds as done in the standard approach. We show that the mechanism which allows one to obtain generalization bounds for random classes in our framework is based on a “small complexity” of certain random coordinate projections. We demonstrate how this notion of complexity relates to learnability and how one can explore geometric properties of these projections in order to derive estimates of rates of convergence and good confidence interval estimates for the expected risk. We then demonstrate the generality of our new approach by presenting a range of examples, among them the algorithm-dependent compression schemes and the data-dependent luckiness frameworks, which fall into our random subclass framework.¶ Second, we study in more detail generalization bounds for a specific algorithm which is of central importance in learning theory, namely the Empirical Risk Minimization algorithm (ERM). Recent results show that one can significantly improve the high-probability estimates for the convergence rates for empirical minimizers by a direct analysis of the ERM algorithm. These results are based on a new localized notion of complexity of subsets of hypothesis functions with identical expected errors and are therefore dependent on the underlying unknown distribution. We investigate the extent to which one can estimate these high-probability convergence rates in a data-dependent manner. We provide an algorithm which computes a data-dependent upper bound for the expected error of empirical minimizers in terms of the “complexity” of data-dependent local subsets. These subsets are sets of functions of empirical errors of a given range and can be determined based solely on empirical data. We then show that recent direct estimates, which are essentially sharp estimates on the high-probability convergence rate for the ERM algorithm, can not be recovered universally from empirical data.

We present a framework for learning in hidden Markov models with distributed state representations. Within this framework, we derive a learning algorithm based on the Expectation--Maximization (EM) procedure for maximum likelihood estimation. Analogous to the standard Baum-Welch update rules, the M-step of our algorithm is exact and can be solved analytically. However, due to the combinatorial nature of the hidden state representation, the exact E-step is intractable. A simple and tractable mean field approximation is derived. Empirical results on a set of problems suggest that both the mean field approximation and Gibbs sampling are viable alternatives to the computationally expensive exact algorithm.

Statistical machine learning is a branch of mathematics concerned with developing algorithms for data analysis. This thesis presents new mathematical models and statistical software, and is organized into two parts. In the first part, I present several new algorithms for clustering and segmentation. Clustering and segmentation are a class of techniques that attempt to find structures in data. I discuss the following contributions, with a focus on applications to cancer data from bioinformatics. In the second part, I focus on statistical software contributions which are practical for use in everyday data analysis.

Software quality ensures that applications that are developed are failure free. Some modern systems are intricate, due to the complexity of their information processes. Software fault prediction is an important quality assurance activity, since it is a mechanism that correctly predicts the defect proneness of modules and classifies modules that saves resources, time and developers’ efforts. In this study, a model that selects relevant features that can be used in defect prediction was proposed. The literature was reviewed and it revealed that process metrics are better predictors of defects in version systems and are based on historic source code over time. These metrics are extracted from the source-code module and include, for example, the number of additions and deletions from the source code, the number of distinct committers and the number of modified lines. In this research, defect prediction was conducted using open source software (OSS) of software product line(s) (SPL), hence process metrics were chosen. Data sets that are used in defect prediction may contain non-significant and redundant attributes that may affect the accuracy of machine-learning algorithms. In order to improve the prediction accuracy of classification models, features that are significant in the defect prediction process are utilised. In machine learning, feature selection techniques are applied in the identification of the relevant data. Feature selection is a pre-processing step that helps to reduce the dimensionality of data in machine learning. Feature selection techniques include information theoretic methods that are based on the entropy concept. This study experimented the efficiency of the feature selection techniques. It was realised that software defect prediction using significant attributes improves the prediction accuracy. A novel MICFastCR model, which is based on the Maximal Information Coefficient (MIC) was developed to select significant attributes and Fast Correlation Based Filter (FCBF) to eliminate redundant attributes. Machine learning algorithms were then run to predict software defects. The MICFastCR achieved the highest prediction accuracy as reported by various performance measures. / School of Computing / Ph. D. (Computer Science)