Prior Knowledge in Kernel Methods

High-order adaptive mesh refinement offers the potential for accurate and efficient resolution of problems in fluid dynamics and other fields where a wide range of length scales is present. A critical requirement for the interface closures used with these methods is stability in the context of hyperbolic systems of partial differential equations. In this study, a class of energy-stable high-order finite-difference interface closures is presented for grids with step resolution changes in one dimension. Asymptotic stability in time for these schemes is achieved by imposing a summation-by-parts condition on the interface closure, which is thus also nondissipative. Interface closures compatible with interior fourth- and sixth-order explicit, and fourth-order implicit centered schemes are presented. Validation tests include linear and nonlinear problems in one and in two dimensions with tensor-product grid refinement. A second class of stable high-order interface closures is presented for two-dimensional cell-centered grids with patch-refinement and step-changes in resolution. For these grids, coarse and fine nodes are not aligned at the mesh interfaces, resulting in hanging nodes. Stability is achieved by again imposing a summation-by-parts condition, resulting in nondissipative closures, at the cost of accuracy at corner interfaces. Interface stencils for an explicit fourth-order finite-difference scheme are presented for each geometry. Validation tests confirm the stability and accuracy of these closures for linear and nonlinear problems. The Richtmyer-Meshkov instability is investigated using a novel first-order perturbation of the two-dimensional Navier-Stokes equations about a shock-resolved base flow. The computational domain is efficiently resolved using the one-dimensional fourth-order interface scheme. Results are compared to analytic models of the instability, showing agreement with predicted asymptotic growth rates in the inviscid range, while significant discrepancies are noted in the transient growth phase. Viscous effects are found to be poorly predicted by existing models.

Motor intention in the posterior parietal cortex

Brain-computer interfaces (BCIs) present an alternative way of communication for people with severe disabilities. One of the shortcomings in current BCI systems, recently put forward in the fourth BCI competition, is the asynchronous detection of motor imagery versus resting state. We investigated this extension to the three-class case, in which the resting state is considered virtually lying between two motor classes, resulting in a large penalty when one motor task is misclassified into the other motor class. We particularly focus on the behavior of different machine-learning techniques and on the role of multi-class cost-sensitive learning in such a context. To this end, four different kernel methods are empirically compared, namely pairwise multi-class support vector machines (SVMs), two cost-sensitive multi-class SVMs and kernel-based ordinal regression. The experimental results illustrate that ordinal regression performs better than the other three approaches when a cost-sensitive performance measure such as the mean-squared error is considered. By contrast, multi-class cost-sensitive learning enables us to control the number of large errors made between two motor tasks.

Semi-supervised matrixized least squares support vector machine

In the regulatory assessment of chemicals, the use of in silico prediction methods such as (quantitative) structure-activity relationship models ([Q]SARs), is increasingly required or encouraged, in order to increase the efficiency and effectiveness of the risk assessment process, and to minimise the reliance on animal testing. The main question for the assessor concerns the usefulness of the prediction approach, which can be broken down into the practical applicability of the method and the adequacy of the predictions. A framework for assessing and documenting (Q)SAR models and their predictions has been established at the European and international levels. Exactly how the framework is applied in practice will depend on the provisions of the specific legislation and the context in which the nontesting data are being used. This report describes the current framework for documenting (Q)SAR models and their predictions, and discuses how it might be built upon to provide more detailed guidance on the use of (Q)SAR predictions in regulatory decision making. The proposed framework is illustrated by using selected pesticide active compounds as examples. LIST OF ABBREVIATIONS AD Applicability domain ADME Absorption, Distribution, Metabolism and Elimination ANN Artificial Neural Network CRD UK Chemicals Regulations Directorate DfW Derek for Windows ECHA European Chemicals Agency EFSA European Food Safety Authority EU European Union GCMP Good Computer Modelling Practice JRC Joint Research Centre k-NN K Nearest Neighbours LOAEL Lowest Observed Adverse Effect Level LOO Leave-one-out LMO Leave-many-out OECD Organisation for Economic Cooperation and Development PBBK Physiologically Based Biokinetic (model) PPP Plant Protection Product QMRF QSAR Model Reporting Format QPRF QSAR Prediction Reporting Format QSAR Quantitative Structure-Activity Relationship REACH Registration, Evaluation, Authorisation and Restriction of Chemicals ROC Receiver Operating Characteristic SAR Structure-Activity Relationship SVM Support Vector Machine (computational method) TTC Threshold of Toxicological Concern

Transductive Support Vector Machine (TSVM) is a famous model in solving classification problems in semi-supervised learning (SSL). However, traditional methods of TSVM are always time-consuming, which becomes a bottleneck when there is large scale data. In order to improve the efficiency of a training process for TSVM, we propose a fast training method for classification problems in semi-supervised learning. Different from the traditional algorithms, just few labeled instances are regarded as the initial training dataset at the beginning of our method. In each following iteration, only reliable unlabeled instances are selected and then labels are set by using all instances which are the labels set by the previous iteration. The algorithm finishes until no reliable unlabeled instances can be selected. This core idea makes our proposed method consume less running time than traditional TSVM implementation on SVMlight. In addition, for different degrees of reliable unlabeled instances, different label setting strategies will be applied to keep high accuracy in label setting. Simulations on both artificial data and real-world data have proven that our proposed fast training method can greatly reduce the running time compared to traditional TSVM implementation on SVMlight. Meanwhile, it can reach the same accuracy level in classification.

Localization of mobile nodes in wireless sensor network gets more and more important, because many applications need to locate the source of incoming measurements as precise as possible. Many previous approaches to the location-estimation problem need know the theories and experiential signal propagation model and collect a large number of labeled samples. So, these approaches are coarse localization because of the inaccurate model, and to obtain such data requires great effort. In this paper, a semi-supervised manifold learning is used to estimate the locations of mobile nodes in a wireless sensor network. The algorithm is used to compute a subspace mapping function between the signal space and the physical space by using a small amount of labeled data and a large amount of unlabeled data. This mapping function can be used online to determine the location of mobile nodes in a sensor network based on the signals received. We use independent development nodes to setup the network in metallurgical industry environment, outdoor and indoor. Experimental results show that we can achieve a higher accuracy with much less calibration effort as compared with RADAR localization systems.

Coupling Mortar Finite Element and Boundary Element Methods for 2D Navier-Stokes Equations Courant Element: Before and After Iterative Methods for Solving Stiff Elliptic Problems Straight and Curved Finite Elements of Class C1 and Some Applications to Thin Shell Problems Exact Controllability to Solve the Helmholtz Equation with Absorbing Boundary Conditions Cubic Version of FEM in Elliptic Problems with Interfaces and Singularities Least Squares Mixed Finite Elements Turbulence Modelling in Finite Element Industrial Applications Necessary and Sufficient Conditions for the Numerical Approximation of a Partial Differential Equation Depending on a Small Parameter Efficient Solution Methods for Compressible Flow Computations Parallel Finite Volume Algorithms for Solving the Time-Domain Maxwell Equations on Nonstructured Meshes Coupling Between Nonlinear Maxwell and Heat Equations for an Induction Heating Problem: Modelling and Numerical Methods Solving the 3D Harmonic Maxwell Equations with Finite Elements, Lagrange Multipliers, and Iterative Methods Some Applications of the Hierarchic High Order MITC Finite Elements for Reissner-Mindlin Plates Domain Decomposition for Immiscible Displacement in Single Porosity Systems An Error Estimator for Nonconforming Approximations of a Nonlinear Problem Some Observations on Raviart-Thomas Mixed Finite Elements in p Extension for Parabolic Problems Mixed Finite Element Methods in Fluid Structure Systems A Black-Box Solver for the Solution of General Nonlinear Functional Equations by Mixed FEM A Remark on the Asymptotic Behaviour of Parabolic Variational Inequalities and Their Finite Element Approximation by the Courant Element Domain Decomposition vs. Adaptivity Material Optimization of Composites. Part Contents.

With the broader dissemination of digital technologies, visionary concepts like the Internet of Things also affect an increasing number of use cases with interfaces to humans, e.g. manufacturing environments with technical operators monitoring the processes. This leads to additional challenges, as besides the technical issues also human aspects have to be considered for a successful implementation of strategic initiatives like Industrie 4.0. From a technical perspective, complex event processing has proven itself in practice to be capable of integrating and analyzing huge amounts of heterogeneous data and establishing a basic level of situation awareness by detecting situations of interests. Whereas this reactive nature of complex event processing systems may be sufficient for machine-to-machine use cases, the new characteristic of application fields with humans remaining in the control loop leads to an increasing action distance and delayed reactions. Taking human aspects into consideration leads to new requirements, with transparency and comprehensibility of the processing of events being the most important ones. Improving the comprehensibility of complex event processing and extending its capabilities towards an effective support of human operators allows tackling technical and non-technical challenges at the same time. The main contribution of this thesis answers the question of how to evolve state-of-the-art complex event processing from its reactive nature towards a transparent and holistic situation management system. The goal is to improve the interaction among systems and humans in use cases with interfaces between both worlds. Realizing a holistic situation management requires three missing capabilities to be introduced by the contributions of this thesis: First, based on the achieved transparency, the retrospective analysis of situations is enabled by collecting information related to a situation's occurrence and development. Therefore, CEP engine-specific situation descriptions are transformed into a common model, allowing the automatic decomposition of the underlying patterns to derive partial patterns describing the intermediate states of processing. Second, by introducing the psychological model of situation awareness into complex event processing, human aspects of information processing are taken into consideration and introduced into the complex event processing paradigm. Based on this model, an extended situation life-cycle and transition method are derived. The introduced concepts and methods allow the implementation of the controlling function of situation management and enable the effective acquisition and maintenance of situation awareness for human operators to purposefully direct their attention towards upcoming situations. Finally, completing the set of capabilities for situation management, an approach is presented to support the generation and integration of prediction models for predictive situation management. Therefore, methods are introduced to automatically label and extract relevant data for the generation of prediction models and to enable the embedding of the resulting models for an automatic evaluation and execution. The contributions are introduced, applied and evaluated along a scenario from the manufacturing domain.

There is a trend in recent machine learning community to construct a nonlinear version of a linear algorithm using the 'kernel method', e.g. Support Vector Machines (SVMs), kernel principal component analysis, kernel fisher discriminant analysis and the recent kernel clustering algorithms. In unsupervised clustering algorithms using kernel method, typically, a nonlinear mapping is used first to map the data into a potentially much higher feature space, where clustering is then performed. A drawback of these kernel clustering algorithms is thatthe clustering prototypes lie in high dimensional feature space and hence lack clear and intuitive descriptions unless using additional projection approximation from the feature to the data space as done in the existing literatures. In this paper, a novel clustering algorithm using the 'kernel method' based on the classical fuzzy clustering algorithm (FCM) is proposed and called as kernel fuzzy c-means algorithm (KFCM). KFCM adopts a new kernel-induced metric in the data space to replace the original Euclidean norm metric in FCM and the clustered prototypes still lie in the data space so that the clustering results can be reformulated and interpreted in the original space. Our analysis shows that KFCM is robust to noise and outliers and also tolerates unequal sized clusters. And finally this property is utilized to cluster incomplete data. Experiments on two artificial and one real datasets show that KFCM hasbetter clustering performance and more robust than several modifications of FCM for incomplete data clustering.

In recent years, discriminative learning methods have widely been used in various areas of Natural Language Processing (NLP). These methods achieve the best performance, when the set of training and testing samples have the same distribution. However, in many applications of NLP, the lack of labeled datasets for some domains is a serious challenge. In such conditions, we need to develop a model based on domains with rich labeled instances and apply it to the domain with no labeled instances. In this research, a method for sentiment classification of opinions into positive and negative groups, which represent the users' feelings, is offered based on multi-source transfer learning. The proposed method here employs Spectral Feature Alignment algorithm to adapt different domains. Furthermore, according to the Majority Voting, accuracy is assigned to classifications trained on different domains based on the Majority Voting Error. Ultimately, decisions are made for each classification based on the calculated error. The Amazon datasets for four different categories, each of which contains 1000 positive and 1000 negative samples, are exploited to train the proposed model. Meanwhile, each category includes unlabeled samples that are used to select pivot features. The accuracy values of 85.5%, 86.4%, 83.5% and 90.1% obtained for Electronics, DVD, Books and Kitchen domains respectively, show the effectiveness of the proposed method compared with similar methods.

Long term supporting schemes for photovoltaic (PV) system installation have led to accommodating large numbers of PV systems within load pockets in distribution grids. High penetrations of PV systems can cause new technical challenges, such as voltage rise due to reverse power flow during light load and high PV generation conditions. Therefore, new strategies are required to address the associated challenges. Moreover, due to these changes in distribution grids, a different response behavior of the distribution grid on the transmission side can be expected. Hence, a new equivalent model of distribution grids with high penetration of PV systems is needed to be addressed for future power system studies. The thesis contributions lie in three parts. The first part of the thesis copes with the PV modelling. A non-proprietary PV model of a three-phase, single stage PV system is developed in PSCAD/EMTDC and PowerFactory. Three different reactive power regulation strategies are incorporated into the models and their behavior are investigated in both simulation platforms using a distribution system with PV systems. In the second part of the thesis, the voltage rise problem is remedied by use of reactive power. On the other hand, considering large numbers of PV systems in grids, unnecessary reactive power consumption by PV systems first increases total line losses, and second it may also jeopardize the stability of the network in the case of contingencies in conventional power plants, which supply reactive power. Thus, this thesis investigates and develops the novel schemes to reduce reactive power flows while still keeping voltage within designated limits via three different approaches: 1. decentralized voltage control to the pre-defined set-points 2. developing a coordinated active power dependent (APD) voltage regulation Q(P) using local signals 3. developing a multi-objective coordinated droop-based voltage (DBV) regulation Q(V) using local signals In the third part of the thesis, furthermore, a gray-box load modeling is used to develop a new static equivalent model of a complex distribution grid with large numbers of PV systems embedded with voltage support schemes. In the proposed model, variations of voltage at the connection point simulate variations of the model’s active and reactive power. This model can simply be integrated into load-flow programs and replace the complex distribution grid, while still keeping the overall accuracy high. The thesis results, in conclusion, demonstrate: i) using rms-based simulations in PowerFactory can provide us with quite similar results using the time domain instantaneous values in PSCAD platform; ii) decentralized voltage control to specific setpoints through the PV systems in the distribution grid is fundamentally impossible due to the high level voltage control interaction and directionality among the PV systems; iii) the proposed APD method can regulate the voltage under the steady-state voltage limit and consume less total reactive power in contrast to the standard characteristic Cosf (P) proposed by German Grid Codes; iv) the proposed optimized DBV method can directly address voltage and successfully regulate it to the upper steady-state voltage limit by causing minimum reactive power consumption as well as line losses; v) it is beneficial to address PV systems as a separate entity in the equivalencing of distribution grids with high density of PV systems.

This paper investigates the use of semi-supervised learning (SSL) for predicting Alzheimers Disease (AD) conversion in Mild Cognitive Impairment (MCI) patients based on Magnetic Resonance Imaging (MRI). SSL methods differ from standard supervised learning methods in that they make use of unlabeled data - in this case data from MCI subjects whose final diagnosis is not yet known. We compare two widely used semi-supervised methods (low density separation (LDS) and semi-supervised discriminant analysis (SDA)) to the corresponding supervised methods using real and synthetic MRI data of MCI subjects. With simulated data, using SSL instead of supervised learning led to higher classification performance in certain cases, however, the applicability of semi-supervised methods depended strongly on the data distributions. With real MRI data, the SSL methods achieved significantly better classification performances over supervised methods. Moreover, even using a small number of unlabeled samples improved the AD conversion predictions.

This paper reviews the state of the art of using Brain-Computer Interfaces (BCIs) in combination with Augmented Reality (AR). First it introduces the field of AR and its main concepts. Second, it describes the various systems designed so far combining AR and BCI categorized by their application field: medicine, robotics, home automation and brain activity visualization. Finally, it summarizes and discusses the results of the study.

Aiming at the problem of object-based image retrieval, a novel semi-supervised multi-instance learning (MIL) algorithm based on RS (rough set) attribute reduction and Transductive support vector machine (TSVM) has been presented---RSTSVM- MIL algorithm. This algorithm regards the whole image as a bag, and the low-level visual feature of the segmented regions as instances, in order to transform every bag to a single sample, RSTSVM-MIL first generates a collection of visual-word to construct projection space by K-means clustering method in the instance feature space, then a nonlinear mapping is defined using these visual-word and maps every bag to a point in the projection space, extracting all bags' projection feature, which converts MIL problem to a standard supervised learning problem. Finally, in order to improve the efficiency of classifier training and to use a large amount of unlabeled image together with the labeled image to build better classifiers, we use RS (rough set) method to reduce the redundant information in projection feature firstly, then semi-supervised learning algorithm TSVM be used to train classifier. Experimental results on the SIVAL image set show that solving MIL problems using semi-supervised learning is feasible, and the RSTSVM-MIL is competitive with other state-of-the-art MIL algorithms in object- based image retrieval problem.