PhD Students

ABBELOOS Dirk (Starting date : September 2009)
Multigrid methods for the control of time periodic parabolic partial differential equations
Thesis advisors : S. Vandewalle, M. Diehl
A variety of applications within the optimization in engineering center (OPTEC) leads to the simulation and optimization of complex systems of time periodic parabolic partial differential equations (PDEs). Traditional methods are based on transforming the problem into a system of ordinary differential equations (ODEs) and afterwards solving them with existing ODE-solvers. These methods do not incorporate the spatial PDE properties in the solution process and therefore they are slow and incapable of handling large systems. In this research we develop one-shot multigrid methods incorporating the spatial PDE properties in the solution process. One-shot methods solve the optimality system for the states, adjoints and controls at the same time. This approach results in much faster methods.

AKKERMANS Simen (Starting date : October 2013)
Predictive microbiology under stressing conditions: systematic procedures to extend widely used modelling methodologies
Thesis advisor : Jan Van Impe
The goal of this PhD is to develop methods for the extension of microbial models to accurately describe the behaviour of microorganisms under new and stressing environmental conditions. To this end, kinetic models will be extended based on the Gamma concept to take interactions between environmental conditions into account. Next probabilistic models will be extended to include interactions between environmental conditions. Also the performance of data mining techniques will be examined to predict the chance of growth of microorganisms in a specific environment. The main application of these results lies in the field of food technology.

ANSARI Amir Hossein (Starting date : 7 October 2013)
Development of a fully automated neonatal brain monitor
Thesis advisor : Sabine Van Huffel
This PhD project fits within a larger collaboration with the departments of Neonatology at the University Hospitals of Leuven, Middelheim Hospital in Antwerp and the Erasmus Medical Center Rotterdam. Considering the diversity of neonatal brain disorders, in first instance the focus will be on asphyxiated patients. The main goal of this so-called NeoGuard PhD project is to build a decision supportive tool for the diagnosis and management of neonatal seizures (I) and abnormal background patterns (II), embedded in a user-friendly Graphical User Interface (III).

BAKA Maria (Starting date : October 2010)
Development of predictive models for structured food products: a multi-scale approach
Thesis advisor : Jan Van Impe
Experiments performed at different scales, i.e., micro-level, meso-level and macro-level, are combined to unravel the dynamics (growth, inactivation, survival) of pathogens and spoilage micro-organisms in structured systems. The newly gained information is used to build broadly applicable predictive models.

BILLIET Lieven (Starting date : October 2014)
Sensor-based Platform for the Accurate and Remote monitoring of Kine(ma)tics Linked to E-health (SPARKLE)
Thesis advisor : Sabine Van Huffel
Chronic musculoskeletal conditions are highly prevalent, with 30% of European population being affected. An effective treatment asks for a continuous and reliable monitoring. SPARKLE, an interdisciplinary project, aims therefore to develop a wearable system of multiple inertial measurement units (IMU). This PhD focuses on the estimation of kine(ma)tic parameters for the IMU data. Furthermore, new approaches for (continuous) activity estimation and evaluation are being investigated, particularly oriented towards interpretability, e.g. by means of interpretable kernels. In the end, a clinical decision support system will be developed.

BOONS Kathleen (Starting date : October 2010)
Integration of microscopic information in predictive models applied to food safety and food quality
Thesis advisor : Jan Van Impe
The aim of this research is the visualization, characterization and quantification of the microbial dynamics in structured food (model) systems. Microscopic methods are applied and/or developed to study the microbial evolution in gelled systems.

BOUSSE Martijn (Starting date : September 2014)
Optimization of multilinear functions for data analysis
Thesis advisor : Lieven De Lathauwer
Many problems in science and technology are centered around two key concepts from linear algebra: matrix decompositions and linear systems of equations. In multilinear algebra, we consider tensors: a higher-order generalization of matrices and vectors. Naturally, tensor decompositions are a multilinear generalization of classical matrix decompositions such as the well-known singular value decomposition. On the other hand, little is known about a multilinear generalization of linear systems which is exactly the focus of this research. The main goal is to develop an efficient optimization-based framework for structured data fusion based on these multilinear systems, paving the way to a broad array of new applications in signal processing and data analysis in general.

BOUVIN Jeroen (Starting date : October 2012)
Enhanced heterologous protein production in Streptomyces lividans by metabolic flux optimization
Thesis advisors : K. Bernaerts, J. Van Impe, J. Anné
In this PhD research, a combination of flux balance analysis and 13C-metabolic flux analysis on continuous bioreactor experiments will be used to unravel the metabolic burden caused by heterologous protein production. Based on the results, a predictive metabolic model will be identified and used to determine appropriate genetic manipulations which will increase protein production.

CAICEDO DORADO Alexander (Starting date : February 2009)
Signal processing for multimodal perinatal monitoring
Thesis advisor : Sabine Van Huffel
Development of novel multimodal signal processing techniques based on multi-way signal and 3D canonical correlation analysis in order to extract the common dynamics among multiple neonatal recordings (cerebral oxygenation, EEG, ECG, EOG, EMG, respiration, saturation, etc...) in order to monitor and detect risk situations in an automated way. The algorithms developed will be integrated in a software platform to be used for bedside monitoring.

CAO The Anh (Starting date : October 2010)
Monitoring and model based control of biological wastewater treatment processes
Thesis advisor : Ilse Smets
Although a lot of research has already been dedicated to modeling biological wastewater treatment processes, the developed (most often first principles based and, hence, complex models are not yet commonly integrated in monitoring and control strategies in practice. This PhD work concentrates on monitoring (e.g., through image analysis) and model based control of the plants for which reliable but low complexity models are required. Applications in conventional activated sludge systems and membrane bioreactors are envisaged.

CASTRO GARCIA Ricardo (Starting date : 14 October 2013)
Modeling structured dynamical systems using parametric and kernel-based models
Thesis advisor : Johan Suykens
In model structures for nonlinear system identification support vector machines and kernel methods have been successfully applied in the past for certain classes of model structures. However, for complex system configurations it is challenging to further incorporate prior knowledge about the structure e.g. within a primal-dual optimization setting of support vector machine related models. Often the structure is also unknown and one wants to estimate it by making use of sparsity inducing penalty terms related to L1 or L0 regularization and nuclear norm regularization. The aim of the research is to advance in this area which is at the interface between nonlinear system identification and machine learning, by combining and integrating the best of both paradigms and employing both parametric and kernel-based approaches with suitable regularization schemes.

CUPPENS Kris (Starting date : July 2009)
Dection of epileptic seizures based on video and accelerometer recordings
Thesis advisors : Bart Vanrumste, Sabine Van Huffel
The goal of this doctoral project is to develop a reliable stand-alone system for the detection and classification of nocturnal epileptic seizures (convulsions) based on video images. The purpose is at the one hand to log the epileptic seizures in a database so the neurologist has an objective measure of the number of seizures, and at the other hand to alarm the parents of caregivers when a seizure occurs where care is needed afterwards.

DE CAIGNY Jan (Starting date : October 2005)
Control of slowly time-varying systems
Thesis advisors : Jan Swevers, Joris De Schutter
The aim of this research is to analyze and control the behaviour of linear slowly time-varying dynamical systems. Considering the bounds on the rate of variation of the time-varying parameters, less conservative and more general applicable design techniques for linear time-varying controllers are developed.

DE COOMAN Thomas (Starting date : 1 October 2013)
Online epileptic seizure detection in a home environment
Thesis advisor : Sabine Van Huffel
The goal of this PhD is to detect epileptic seizures online in a home environment. Seizure detection is typically done by analyzing the EEG, but acquisition of the EEG outside the hospital is impractical. Therefore the ECG, EMG and accelerometers will be processed in this PhD. Two multimodal approaches will be investigated. A first one puts extra emphasis on user-friendliness of the multimodal system by combining a good performance, low detection delay and a minimal amount of sensors to be worn. A second algorithm exploits the correlation between the different modalities (by for example online tensor decompositions) in order to get a further improved performance. An adaptive LS-SVM classifier that is able to evolve from a patient-independent classifier into a patient-specific classifier by using the feedback from the patient on earlier warnings will be defined.

DE GROOTE Friedl (Starting date : October 2005)
A convex optimization approach to dynamic musculoskeletal analysis
Thesis advisors : J. De Schutter, J. Swevers
Recently, at the division PMA of the K.U.Leuven a new inverse approach for dynamic musculoskeletal analysis was developed. This so-called physiological inverse approach takes into account the muscle physiology while preserving the numerical efficiency of inverse approaches. The goal of this Ph.D. is to further develop this approach: the problem structure can be exploited to further increase the numerical efficiency, the approach needs to be adapted to handle fast motions and there is a need for measures to handle the sensitivity for inaccuracies in the measured kinematics.

DE WEL Ofelie (Starting date : 5 October 2015)
Monitoring of neonatal brain maturation using tensor approaches
Thesis advisor : Sabine Van Huffel
About 10% of the births world-wide are premature and preterm birth is considered to be one of the main causes of neonatal morbidity. Although the survival rate of preterm infants has increased during the last decades due to advances in the NICU, neurodevelopmental impairment is still a major complication in these premature survivors. As a result, monitoring of brain function during the critical weeks of the preterm infant has gained attention. The goal of this thesis is to monitor the brain maturation of preterm infants by analysing the EEG using tensor approaches. Tensor decompositions will be used to extract features that describe brain maturation. These features will then be used to develop preterm EEG growth charts, which will assist the neonatalogists in assessing the current state of the infant, planning the treatment and predicting the outcome.

DEBALS Otto (Starting date : 15 September 2013)
Tensorisation with applications in blind source separation and blind system identification
Thesis advisor : Lieven De Lathauwer
Many powerful tensor techniques can be found in literature. In this project, these techniques will be applied directly onto matrixdata using tensorisation, i.e. the reformulation/restructuration of matrices into higher-order structures. In this way one avoids the need of tensordata itself. We focus on blind source separation and blind system identification (the convolutive case of the former with still a lot of potential for tensor-based methods), with elaborations to non-negative matrix factorization and problems concerning big data. We develop new methods based on various innovative points of view with expected improvements in speed, robustness and possibilities.

DOMANOV Ignat (started predoc student October 2008, PhD student since 20 April 2009)
Tensor-Based Signal Separation
Thesis advisor : Lieven De Lathauwer
Signal processing makes more and more use of techniques based on multilinear algebra. In my thesis, two tensor decompositions that are at the heart of these methods will be further studied, namely the Canonical / Parallel Factor Decomposition and the Block Term Decomposition. The work will include the study of uniqueness and the derivation of algorithms. The results will be used to develop new methods for blind source separation.

FERRANTI Micol (Starting date : March 2012)
Spectral properties of (perturbed) normal matrices and their applications
Thesis advisor : Raf Vandebril
In this research project computational and theoretical issues linked with spectral properties of the classes of normal and perturbed normal matrices will be studied. The overall research goal is the development of new, theoretically well-supported, fast, accurate and reliable algorithms for computing (approximate) spectral decompositions of (perturbed) normal matrices. The embracing goal encloses three weighty subgoals. We aim at notable progression in, and aspire towards significant contributions in theory, algorithms design and the development of competitive software.

GILLIS Joris (Starting date : November 2010)
Optimal design of robust periodic controllers for nonlinear mechanical systems
Thesis advisor : M. Diehl
The possibility to use efficient derivative based optimization tools promises to make the framework of periodic Lyapunov differential equations (PLDE) applicable to non-trivial real-world systems from the mechanical engineering domain. While an elegant mathematical framework already exists, some fundamental extensions are required to apply the ideas to these real-world systems. The purpose of this PhD is to propose new formulations, extensions and computational tools for the PLDE framework, such that it can really enter engineering practice.

GONG Xiao-Feng (Starting date : December 2014)
Coupled tensor decomposition for multi-set signal processing
Thesis advisor : Lieven De Lathauwer
Tensors have been widely using in signal processing. In particular, there are many applications that involve the joint processing of multi-set signals, such as multi-model / multi-subject data fusion, or BSS of convolutive signals transformed into multiple frequency bins. These multiple datasets are usually linked with one another via the relationship between latent source components, and could be formulized as multiple tensors coupled with commonly shared factors. Therefore, it is of high significance to develop coupled tensor decomposition tools for multi-set problems. By taking into consideration the coupling of tensors in the decomposition procedure, one can expect to obtain better identifiability results, and improved estimates of unknown parameters involved in the applications.

GOOVAERTS Griet (Starting date : September 2013)
ECG based risk stratification for sudden cardiac death
Thesis advisor : Sabine Van Huffel
Sudden cardiac death is the second most important cause of death in Belgium. In this PhD project, the goal is to develop methods to predict the risk on sudden cardiac death using the ECG signal. Tensor methods will be used since they allow us to look at multidimensional characteristics of the ECG signal. Since the signals are two-dimensional, they first need to be tensorized by adding extra dimensions to the signal. The tensors can then be decomposed in different components, and risk factors for sudden cardiac death can be derived from these components.

HALANDUR NAGARAJA Bharath (Starting date : April 2014)
Tensor based Blind Source Separation in Magnetic Resonance Spectroscopic Imaging
Thesis advisor : Sabine Van Huffel
MRS(I) signals contain information for estimating metabolite concentrations in-vivo. Along with useful components MRS(I) signals also contain unwanted components such as water, baseline etc. The damped exponential model is widely used for filtering unwanted components. We try to exploit this model and other sparse representations for removal of water artefact and automated quality control using tensor decomposition methods (CPD and BTD). We try to develop new class of algorithms based on BTD/CPD to improve automation and enable all-at-once quantification of the entire 2D/3D MRS(I) set. Another major application is brain tumour tissue typing. Here, each source represent the MRS signal from a pure tissue type (healthy, actively growing tumour, necrotic tissue…). Blind source separation (BSS) with prior knowledge can help tremendously in the absence of training data. So far only BSS approaches based on NMF or ICA have been used but they are restrictive since they only allow a linear combination of sources. Therefore we will try CPD/BTD based BSS to capture the spatial information and inherent non-linearity. We try to develop BTD/CPD based algorithms to improve automation and brain tumour heterogeneity characterization. No training set and hence no supervised classification is needed.

HE Zhuo-Heng (Starting date : September 2014)
Matrix equations, Matrix decompositions and their applications in engineering and control theory
Thesis advisor : Bart De Moor
It is known that matrix equations have been one of the main topics in matrix theory and its applications, and a large number of papers have presented several methods for solving several matrix equations. In mathematics, control theory, engineering and others, many problems can be transformed into some matrix equations. For instance, many problems in systems and control theory are concerned with the solution of the Sylvester matrix equation. The Sylvester matrix equation has found huge applications in optimal control, robust fault detection, pole/eigenstructure assignment design, and so on, and has been investigated by several researchers. The decompositions of matrices have always been at the heart of system theory and signal processing. We aim to analyze in detail the structure of some new decomp-ositions for some sets of matrices. As applications of these decompositions, we give some solvability conditions, general solutions, as well as the range of ranks of the general solutions to the system of k generalized associated Sylvester matrix equations.

HOUTHUYS Lynn (Starting date : September 2014)
Coupled Data-Driven Models
Thesis advisor : Johan Suykens
This research focuses on multitask learning with spectral regularization or nuclear norm regularization to obtain a low-rank solution when combining multiple submodels. This will mostly be done for kernel-based models like SVM and LS-SVM. Multiple forms of coupling and regularization will be researched, as well as the possibility of adding extra layers. This will have applications within various areas like electricity usage and weather forecasting.

HUMET Matthias (Starting date : October 2010)
Matrix computations and orthogonal functions
Thesis advisor : M. Van Barel
A lot of research has already been done in the field of matrix computations as well as in the field of orthogonal functions. However, the full potential of the interplay between these results has not been considered up to now. The aim of this PhD-project is to investigate how results of one of these domains can be applied into the other.

ION-MARGINEEANU Adrian (Starting date : 1 October 2013)
Classifiers for tumour tissue differentiation based on multimodal Magnetic Resonance data
Thesis advisor : Sabine Van Huffel
This doctoral project will develop classification methods that can cope with multi-modal MR data (in particular, anatomic MRI, spectroscopy, diffusion and perfusion MRI), and apply them to the diagnosis and follow-up of brain tumour and multiple sclerosis. One of the main goals is to solve some of the standing problems in brain tumour applications, e.g., non-invasive tumour subtyping and grading, prediction of tumour infiltration, or tumour recurrence after treatment. To this end, several data processing steps need to be designed and optimized: computation of relevant features from the data, matching the image resolutions among the different modalities, multi-modal data fusion and classification.

JEURIS Ben (Starting date : September 2011)
Reestablishing smoothness for matrix manifold optimization via resolution of singularities
Thesis advisor : Raf Vandebril
In this PhD project, Riemannian optimization techniques for matrix manifolds are examined. These techniques maintain and exploit the desired matrix structure throughout the algorithm, resulting in efficient optimization methods. Our aim is to enhance existing optimization techniques to cope with important classes of matrices, not necessarily leading to smooth manifolds (e.g. displacement rank matrices). Loss of smoothness implies major difficulties for many optimization algorithms. Exploiting well-developed mathematical theories for the resolution of these singularities in an algebraic/analytic context should enable us to expand the state-of-the-art. We will apply our new methods to two important unsolved matrix problems: computing the matrix geometric mean of particularly structured matrices and solving common matrix equations like Lyapunov and Riccati for classes of data-sparse structured matrices. Concerning this second problem, we will pay specific attention to matrices of low displacement rank.

JUMUTC Vilen (Starting date : March 2012)
Flexible software design and kernel-based learing in advanced data-driven modeling
Thesis advisor : Johan Suykens
Kernel-based methods like SVM and LS-SVM are among the most popular machine learning techniques for solving complex classification, regression or correlation analysis tasks. Currently there are many software packages available for the end user but the majority of such solutions are merely intended for the use of machine learning practitioners and cannot be adopted by the out-of-field scientists. One might consider the ultimate necessity for the simple self-explanatory software design and usage patterns where sophisticated machine learning methods are wrapped by the out-of-box tuning, cross-validation and evaluation procedures. Combining kernel-based methods with the advanced software design for elaborating scalable, robust and user-friendly black-box modelling library is one of the major objectives of my research.

KAREVAN Zahra (Starting date : June 2014)
Data-driven prediction of complex networks
Thesis advisor : Johan Suykens
The aim of the thesis is to study the feasibility of applying black-box data driven modelling to real-life complex systems for predictive modelling. Many real-life systems and networks are critically dependent on high quality predictions, e.g. weather and climate, environmental systems, financial systems, traffic systems, power grids and others. Often a white box physical modelling is taken. The challenge in this thesis however is to study the application of advanced black-box modelling techniques to obtain high quality predictions for systems that are characterized by high dimensional input spaces and for which large amounts of data are available.

KOOLEN Ninah (Starting date : October 2011)
Automated neonatal EEG background analysis for predition of neurological outcome
Thesis advisor : Sabine Van Huffel
This PhD project aims at quantifying and interpreting the neonatal EEG for specific pathologies in a novel way. The first and foremost important scientific objective of the project consists of quantifying the EEG-signal, i.e. both to detect and characterize seizure activity as well as the ongoing background activity (baseline pattern and variability over time). To this end, advanced signal analysis is used, as an objective tool, allowing neonatologists to better predict outcome as well as plan immediate treatment and monitor the effect of the started treatment. Three tasks are investigated here:
1. Automatic analysis of the background pattern, including quantification of the most relevant descriptive features.
2. New features in the background pattern, such as asynchrony, spontaneous brain activity (SAT), brain connectivity patterns etc., by using multichannel signal analysis.
3. Comparison and correlation of these descriptive background measures with the neurological outcome of the babies. In particular, these background patterns will be related to 8 grades, which are assigned to neonates according to the evolution of discontinuity and presence of sleep-wake cycles in their background EEG.

LATAFAT Puya (Starting date : 03 November 2014)
Operator splitting techniques and their application to embedded optimization problems
Promotor : Panos Patrinos
Operator splitting techniques are algorithms that work by splitting a problem into smaller subproblems that are easier to solve. They can be used to solve problems arising in PDEs, monotone inclusions, variational inequalities and optimization. These methods were introduced in the 50's and have recently gained popularity due to their simplicity and applicability to large scale optimization problems such as the ones encountered in machine learning, signal processing and control. Over the years, numerous splittings have been proposed but in most of the cases the relation between them is not very well understood. Furthermore, there is usually a lack of tight convergence rate estimates as well as an ambiguity in how to select the parameters of the algorithms. The aim of my research is to try to answer to some of these short comings.

LAUWERS Joost (Starting date : October 2010)
Identifiability and calibration of anaerobic digestion models
Thesis advisor : Jan Van Impe
The aim is to investigate the identifiability of an anaerobic digestion model. Identifiability encompasses two aspects, structural and practical identifiability. The former relates to the possibility of successfully estimating the parameters of the model in perfect conditions(noiseless measurements, continuous sampling, model is a perfect representation of the ongoing process). This question is related to the concepts of observability and can be answered using methods of differential algebra and geometry. Practical identifiability pertains to the question of parameter estimation in realistic conditions, i.e. presence of measurement noise, a limited sampling frequency and time frame and practical limitations on the input of the process.

LAUWERS Oliver (Starting date : October 2014)
Clustering Time Series : An information theoretical approach
Thesis advisor : Bart De Moor
Clustering time series is still very much an open problem. Though many algorithms, evaluation criteria and similarity measures exist, they often inadequately treat temporal data. This research project focuses on a novel similarity metric for time series data. In state-of-the-art metrics, the time series usually is vectorized, and then a vector-distance is chosen. Quite often, some transformations are done on the data in order to find the best fit for two time series, but ultimately, a lot of temporal information is discarded. We try to generalize the cepstral distance, as explored by R. Martin, K. De Cock and prof. De Moor, to a more general class of time series. this distance measure has a fundamental description in both statistics and information theory, and makes use of the full dynamical and temporal interpretation of the time series, rather than of a vector representation of it.The ultimate goal is to have an insightful way to cluster a broad class of time series, and apply this to real-life problems as smart grid segmentation and customer relation management in banks.

LEJON Annelies (Starting date : September 2011)
Thesis advisor : Giovanni Samaey
Systems evolving on different time scales are a topic of research for many years now, especially if one is only interested in the macroscopic evolution of the system. Applying multi scale methods yields a very efficient simulation. Moreover we want to develop an asymptotic preserving scheme such that the discretization of the kinetic equations describing the system corresponds with the macroscopic behavior.

LEJON Nele (Starting date : October 2012)
Analysis and applications of orthogonal polynomials in the complex plane
Thesis advisors : Daan Huybrechs and Arno Kuijlaars (Department of Mathematics)
The aim of this research is to find a numerically efficient algorithm to approximate highly oscillatory integrals. Classical Gaussian quadrature formula fail in this oscillatory setting. Therefore an alternative path of integration is chosen in the complex plane. As classical in these approximations zeroes of orthogonal polynomials are used as quadrature points. It is therefore essential to know these orthogonal polynomials in the complex plane.

LIU Yang (Starting date : November 2012)
Thesis advisor : Sabine Van Huffel
More information soon

LOPEZ Yaidel Reyes (Starting date : August 2009)
Multiscale particle based simulation software
Thesis advisor : Dirk Roose
The aim of the project is to develop methods and software for multiscale particle based simulation, with emphasis on efficient generic algorithmic components, and implementation on parallel high performance computers (HPC).

MALL Raghvendra (Starting date : March 2012)
Sparsity in parametric and kernel models
Thesis advisor : Johan Suykens
Kernel based methods like SVM and LS-SVM aim for learning the underlying model providing appropriate results for out-of-sample data points. Typically training kernel models implies a computational complexity which scales quadratically w.r.t to the number of training points and creating the kernel matrix requires memory which also scales quadratically w.r.t to the number of training points. We try to exploit the primal-dual formulation orginating from the theory of convex optimization which leads to parametric and kernel models respectively. We explore sparsity at two levels - (i) For the selection of an active subset of prototype vectors using quadratic Renyi Entropy (ii) Performing operations like L0-norm or selecting fewer but relevant prototype vectors leading to sparse kernel models. Sparsity can lead to more memory and computational efficient techniques, enables scaling to large scale datasets and allows better interpretation of data (in comparison to black-box methods). We plan to investigate the trade-off between model accuracy and the degree of sparsity introduced for problems ranging from classification, regression, kernel spectral clustering, kernel PCA etc. using a iterative sparsifying algorithm based on a convex optimization formulations. We also investigate sampling in large scale community detection and use the out-of-sample extension property of KSC to obtain communities from large scale networks.

MARTINEZ LOBETE Maria (Starting date: March 2012)
Studying the relation between microbial dynamics and novel food formulations
Thesis advisor : Jan Van Impe
Partial or total elimination of certain ingredients from a food product, such as fat or sugar, may affect the behavior of food concerning microorganisms and therefore, food safety and quality. Available predictive models do not consider the effect of novel food formulations; this research focuses on the growth dynamics of food pathogens and spoilage micro-organisms in presence of novel compounds. Obtained results will contribute to an optimal design of food safety assurance systems.

MATIC Vladimir (Starting date : March 2010)
Neonatal EEG signal processing
Thesis advisor : Sabine Van Huffel
The main objective will be to analyse the background activity of the neonatal EEG. It is important to describe and develop an automated method for detecting changes in EEG bursts in neonates with varying severity of encephalopathy. Apart from this, research should also include compressive sensing theory and possible future applications to neonatal EEG. The potential benefits consists in compressing the neonatal EEG data, improving seizure detection algorithms and artifact removal.

MEHRKANONN Siamak (Starting date : December 2011)
Incorporating prior knowledge into the learning task
Thesis advisor : Johan Suykens
The primary objective of my research is to explore the possibilities of incorporating the prior knowledge into the learning framework. In many real-life problems one usually encounters a huge amount of unlabeled data points whereas the portion of the labeled data points is few. This is due to the fact that the acquisition of labeled data often requires a skilled human agent or a physical experiment which are of course costly. Semi-supervised learning (SSL) is a framework in Machine Learning which aims at learning from both unlabeled and labeled data points. I have proposed a multi-class semi-supervised learning algorithm (MSS-KSC) and successfully applied it in classification and clustering. The core model is Kernel Spectral Clustering (KSC), a completely unsupervised algorithm. The MSS-KSC approach is formulated as an optimization problem in the primal and dual settings where the side-information (labeled data points) is incorporated to the core model (KSC) using a regularization term. In addition I have made the MSS-KSC approach applicable for dealing with large-scale and data stream such as in video segmentation.

MIJOVIC Bogdan (Starting date : October 2007)
Advanced signal processing techniques for data fusion of multi-modal biomedical information
Thesis advisor : Sabine Van Huffel
In medical image analysis and signal processing there is an ever growing demand for the integration of different heterogeneous sources since they provide us with complementary information. In this PhD project we will study techniques for the combination of EEG (electro-encephalography, 4 up to 20 channels), and other polygraphic signals (ECG, EOG, EMG, respiration, saturation). One of the main issues is the difference in dimensionality of the different heterogeneous sources we want to fuse. To this end we will have to develop new algorithms for : the multi-linear decomposition of multi-channel EEG signals, and the 3D canonical Correlation Analysis (CCA) with other polygraphic signals. The medical support and the data will be provided by the divisions of neonatology of the ERASMUS hospital in Rotterdam and UZ Leuven.

MILOSEVIC Milica (Starting date : July 2010)
Detection of nocturnal epileptic convulsions in pediatric patients based on accelerometers
Thesis advisor : Sabine Van Huffel
The overall objective of this PhD project is to investigate use of accelerometers (ACM) to detect convolutions (epileptic seizures with motor components such as: tonic manifestations, tonic-clonic manifestations, jerks or other seizures with motor components) in pediatric patients. The PhD research work is focused on developing a detection algorithm that use patient-specific prior-knowledge since the clinical manifestations of a convolution strongly vary between patients. The aims are to validate and refine the detection algorithm to reach a sensitivity of 90% in pediatric patients with convulsions to develop a wireless prototype system for real-time monitoring of nocturnal cconvolutionsat home and to investigate the therapeutic impact of having objective measure of the number of nocturnal seizures over a longer period of time.

MOEYERSONS Jonathan (Starting date : September 2016)
Thesis advisor : S. Van Huffel
More information soon.

MOHAMMADI Adeleh (Predoc starting date : 1 February 2013 - PhD start : March 2014))
Robust Linear Model Predictive Control
Thesis advisor : Moritz Diehl
In this research, we study robust MPC for constrained uncertain discrete-time linear systems in the form of ARX-models with uncertainty described as a covariance matrix on the parameters. we are interested in developing robust MPC techniques that are practically applicable for relatively large systems in industrial applications. This means they need to be computationally tractable for long horizons, and very time-consuming off-line computation are not allowed.

NIMMEGEERS Philippe (Starting date: October 2014)
Monitoring and diagnosis of processes with bunched quality measurements
Thesis advisor : Jan Van Impe
The objective of this PhD is to develop algorithms and tools for dynamic flux characterization in biochemical reaction network models for the description of transient phenomena in microbial dynamics.

OSORIO GARCIA Maria Isabel (Starting date : January 2008)
Advanced Signal Processing for Magnetic Resonance Spectroscopic Imaging
Thesis advisor : Sabine Van Huffel
The overall objective of the PhD project is to investigate and develop cutting-edge Magnetic Resonance Spectroscopic Imaging (MRSI) software – eMRSI – for quantification and imaging of metabolites. The PhD research work is focused on / Automatic Statistical Semi-Parametric Estimation /, encompassing novel, advance signal processing and experimental design for metabolite quantification, properly accounting for the nuisance signals of water, macromolecules, and lipids. In particular, two specific MRS quantification software packages will be integrated for acquiring improved results using prior knowledge of the metabolites and molecules that are present in healthy brain and brain tumours. Attention will be paid to optimizing of the baseline approach, robustifying semi-parametric estimation via regularization and quantifying parameter uncertainty. The research is done in a multidisciplinary environment in collaboration with the Department of Radiology in the University Hospitals of Leuven.

PIAMPONGSANT Supinya (Starting date : 1 november 2014)
Systems Physiology: the human body as a hierarchical control system
Promotor : Bart De Moor
This project focuses on the application of systems and control techniques to the domain of human physiology. This includes the simulation of biological pathways and organs as dynamical systems, the application of parameter estimation and systems identification techniques to characterize these systems, and the visualization and enumeration of control systems in the body.

PIPELEERS Goele (Starting date : October 2004)
Robust optimal repetitive control design
Thesis advisors : J. Swevers, J. De Schutter
Based on the Youla-parameterization a general frame-work is developed to design optimal controllers in the presence of periodic inputs. This framework covers the design of feedforward controllers, pseudo-feedforward controllers, as well as repetitive controllers; and many results from literature can be reproduced. Moreover, the optimal control problem is transformed into a convex optimization problem, guaranteeing numerical efficiency and allowing the computation of trade-off curves between conflicting design objectives. In a second stage, uncertainty on the period-time of the periodic input is accounted for.

PUERTAS Gervasio (Starting date : March 2012)
Prior knowledge incorporation in data-driven modelling
Thesis advisor : Johan Suykens
LSSVMS and SVMs are pervasive tools in the machine learning community. Unfortunately though, there remains to be a gap between practitioners and developers when moderately more advanced models are considered. Such models may arise often due to the presence of "prior" knowledge about the problem at hand. Therefore, the aim of this line of work is the development of the necessary tools to allow practitioners to easily and efficiently incorporate different sources of prior information when available.

QUIRYNEN Rien (Starting date : October 2012)
Efficient simulation and optimization of differential algebraic equations on embedded hardware for control
Thesis advisor : Moritz Diehl
The use of dynamic optimization for real-time estimation of the state of a system or to control the same system, is nowadays already widely used in control engineering. This is due to the ability to incorporate constraints and the best available system knowledge into the estimation and control algorithm. For fast time scales in the order of milli- or even microseconds such as in mechatronics, its high computational burden is the major bottleneck. Both estimation and control involve a similar structure for the optimization problem that needs to be solved at each time step. The simulation and derivative generation of the system generally takes a great, often even dominating, fraction of the computation time per step. The goal of this project is therefore to develop more efficient, embedded algorithms for this specific type of simulation in the context of dynamic optimization. In addition to the use of state of the art methods, these implementations shall also exploit the specific requirements of the online algorithm and the principles of automatic code generation to result in extremely fast custom code. This will widen the area of application problems for these advanced control techniques, by being able to cope with much smaller sampling times. The effectiveness of these new implementations will be tested on academic as well as industrial real-world problems from mechatronics and aerospace.

SAUWEN Nicolas (Starting date : October 2012)
Multi-modal Magnetic Resonance Data Processing and Classification Methods for Diagnosis and Therapy Assessment of Tumour Patients
Thesis advisor : Sabine Van Huffel
Malignant gliomas are aggressive brain tumours with poor prognosis. The main objective of this PhD project is the design of computational methods for the analysis of MRSI signals and multi-modal classification techniques, in particular metabolic feature extraction with 3D spatial prior knowledge, and multi-modal approaches to 3D nosologic imaging. These new methods are meant to assist clinicians in the following areas:
(a) establish an accurate diagnosis,
(b) make an early prognosis on success of therapy,
(c) identify areas of microscopic tumour infiltration,
(d) identify mechanisms that contribute to success and failure of (new) therapeutic interventions.

SMET Cindy (Starting date : October 2012)
Cold Atmospheric Plasma for food decontamination
Thesis advisor : Jan Van Impe
This project studies the efficiency of Cold Atmospheric Plasma to inactivate pathogens and spoilage microorganisms in food. The main goal is to study the effect of food properties on the efficiency of the technique. Results will be used to build a predictive model.

SMOLDER Kris (Starting date : January 2004)
Tracking control design for nonlinear systems
Thesis advisors : J. Swevers, P. Sas
This research deals with accurate tracking control for nonlinear multivariable systems. A parametric grey-box state space model structure is develloped together with an identification procedure to identify its parameters for nonlinear time invariant systems. For accurate tracking a nonlinear iterative learning procedure is develloped in which at each iteration an update of the control input is given as a function of the previous control input and the difference of the initial input and the model inverse of the current measured output. As nonlinear system inversion is a critical point in this procedure an algorithm is developped to invert the model structure for a given desired periodic model output.

SOPASAKIS Pantelis (Starting date : October 2016)
Thesis advisor : P. Patrinos
More information soon.

STALLAERT Bert (Starting date : September 2004)
Active noise source control applied to gearbox noise
Thesis advisors : P. Sas, J. Swevers
The idea of this research is to reduce the noise radiated by rotating machinery and specifically gearboxes, by acting on the structure as close as possible to the noise source and in this way cancelling the vibration as soon as possible in the transfer path. An experimental set-up is built and a dedicated active bearing with incorporated piezoelectric acuators is developed. The applicability of control strategies like repetitive control is investigated for this application where the controlled plant typically has a high modal density.

STAMATI Ioanna (Starting date : October 2009)
Optimal experiment design for parameter estimation and model discrimination
Thesis advisors : Jan Van Impe, Filip Logist and Moritz Diehl
The research is about optimal experiment design for parameter estimation and model discrimination. The work is divided in two parts. In the first part optimal experiment design for parameter estimation has been explored through an industrial project. The second part deals with optimal experiment design for model discrimination with applications in predictive microbiology.

STELLA Lorenzo (Starting date: 12 February 2013)
Algorithms for nonsmooth structured optimization: global and local convergence
Promotor : Panos Patrinos
Nonsmooth structured optimization problems arise in all sorts of application areas, such as optimal control, statistics, machine learning, signal and image processing. Methods for solving these are usually based on splitting algorithms, such as the proximal gradient method or the alternating direction method of multipliers (ADMM): being first order methods, though, these algorithms are usually efficient for finding low accuracy solutions, and may become very inefficient when the problem is ill conditioned. My research focuses on accelerating this type of algorithms using second order information on the problem. This can be achieved by reformulating it as an equivalent smooth unconstrained minimization. This way one is able to devise Newton and quasi-Newton methods which possess global complexity estimates but, despite the lack of differentiability of the original problem, converge much faster asymptotically. Our algorithms are implemented in a generic MATLAB framework called ForBES, which is available on Github for download.

SURYANARAYANA Gowri (Starting date : December 2011)
Quasi Monte Carlo algorithms for the Schrödinger equation
Thesis advisers : Dirk Nuyens and Ronald Cools
The aim of this project is to investigate approximations for multivariate integrals based on low-discrepancy point sets for intermediate to high-dimensional problems from computational physics. More specifically this project will target the electronic Schrödinger equation, solving which is very demanding and challenging due to the high-dimensionality of the problem. The aim is to develop new algorithms to determine the probability density of N electrons in three-dimensional space around a collection of K nuclei. Recent results have shown that the number of basis functions required to solve this problem doesn’t grow exponentially in the number of electrons. We analyse the computational complexity of the problems based on such results and then develop new algorithms.

TACK Ignace (Starting date : October 2011)
Multi-scale modeling of microbial colony dynamics
Thesis advisors : Prof. Jan Van Impe
As a general rule low-complexity models are applied in industry to predict and control microbial dynamics. However, extrapolation of these models to colony growth in structured food media is not accurate. For this reason mesoscopic phenomena at the colony level and microscopic knowledge about the microbial cells should be taken into account. In this research project the different scales of interest (cell, colony, global population) are linked to each other. Individual-based Models (IbM) bridge the gap between the microscopic cell processes and the mesoscopic colony behaviour, while the connection between the colony dynamics and global population evolution is modeled by a population balance equation.

THEMELIS Andreas (Starting date : 04 November 2013)
Fast globally convergent methods for convex and nonconvex optimization
Promotor : Panos Patrinos
With the constantly increasing amount of data that modern problems need to face with it is of primary importance to develop efficient algorithms that ensure satisfactory results in reasonable time. I address such issue from two different perspectives, developing methods suited for optimization problems with numerous applications such as control, signal processing, machine learning and image analysis, to name a few. On the one hand I am developing a method with fast asymptotic convergence rate, which is the result of a suitable adaptation of Quasi-Newton methods to nonsmooth optimization. The extension to the nonconvex case is also being considered. On the other, I am developing a "stochastic" algorithm that has same convergence rate guarantees of "full" counterparts, meant for big data-oriented applications in which the size of the problem is prohibitive and cheap computations are to be enforced.

VAN BEEUMEN Roel (Starting date : October 2011)
Low-order approximations of large-scale nonlinear dynamical systems: theory, algorithms and applications
Thesis advisors : Wim Michiels, Karl Meerbergen
The aim of the PhD project it to develop, analyze and apply new numerical methods for the analysis and control of large-scale nonlinear dynamical systems, which are explicitly or implicitly based on low-order approximations. The approach is at the intersection of numerical linear algebra and optimization, and relies on embedding nonlinear problems into infinite-dimensional linear problems, projection techniques (in particular rational Krylov methods) and solving large-scale Lyapunov equations. Applications are mainly envisaged in the area of systems and control. Particular attention will be paid to distance problems in linear algebra, which lie at the basis of robust control.

VAN CRAEN Robin (Starting date : 01 October 2015)
Subspace identification of multidimensional dynamic systems
Promotor: Bart De Moor
This PhD project is about subspace identification of multideminsional linear shift-invariant dynamic systems. For the case of one-dimensional dymamic systems (with for instance time as independent variable), there are already a wide variety of subspace techniques/algorithms available. However in many cases one can take into account also spatial independent variables (e.g. heat distribution in a room) which lead to multidimensional systems. For these systems, the subspace identification problem still yields a lot of unsolved problems. In these PhD we will tackle a few of these problems.

VANDECAPPELLE Michiel (Starting date : September 2016)
Thesis advisor : L. De Lathauwer
More information soon.

VANDECASTEELE Kaat (Starting date : September 2016)
Thesis advisor : S. Van Huffel
More information soon.

VAN DE MOORTEL Nina (Starting date : October 2012)
Stress induced bioflocculation in membrane bioreactor (MBR)
Thesis advisor : Jan Van Impe
The activated sludge process is by far the most widely applied method to treat wastewater. While numerous configurations exist, process efficiency is always dictated by the separation of activated sludge (AS) and purified water. A prerequisite for efficient AS/water separation is a well-flocculated AS consortium. To stimulate bioflocculation, an activated sludge process is operated at low food-to-microorganisms ratios (F/M) such that the microbial community is at all times exposed to substrate stress. The aim of this PhD research is to develop methods to make bioflocculation possible at high F/M-ratios.

VAN DE STAEY Glenn (Starting date : October 2012)
Thesis advisor : Ilse Smets
This PhD focuses on the main drawback of membrane bioreactors as an alternative for wastewater treatment, namely membrane fouling. The goal is, by in situ visualization of the fouling layer build-up, to get a better understanding of how activated sludge properties affect membrane filtration.

VAN EEGHEM Frederik (Starting date : October 2014)
Tensor-based independent component analysis: from instantaneous to convolutive mixtures
Thesis advisor : Lieven De Lathauwer
Blind system identification (BSI) denotes the identification of a system using solely its outputs. An important class of techniques to solve this problem assumes that the inputs are statistically independent. In this Ph.D., tensor-based variants of these techniques will be developed. Tensors are higher-order generalizations of vectors (first order) and matrices (second order). This tensor-based approach is motivated by the properties of tensor decompositions, which make them proper tools for blind system identification.

VAN EYNDHOVEN Simon (Starting date : 1 October 2015)
Tensor based blind sourse separation in simultaneous EEG-fMRI integration
Promotor : Sabine Van Huffel
We focus on the integration of simultaneously acquired EEG-fMRI recordings because of its importance in brain studies. Several matrix based Blind Source Separation (BSS) approaches, unmixing both modalities together, have been proposed such as Parallel ICA, Joint ICA, multiset-CCA, matrix-CPD. Coupled Tensor Decomposition (TD) approaches are emerging, their development is our main challenge here. Coupled Canonical Polyadic Decompositions as well as Block term Tensor decompositions will be investigated and developed. Relevant coupling constraints will be imposed. Novel task-related validation criteria will be set up that better measure the quality of the extracted components. Attention is focused on visual/auditive recognition studies and to epileptic onset zone localization. We expect that our novel coupled TD approaches, applied to EEG-fMRI integration, extract components that are more closely related to the imposed task. These are expected to better reveal the spatiotemporal brain pathways in cognitive brain functioning and better model trial-to-trial fluctuation. In epileptic patients, the techniques are expected to improve spatiotemporal resolution in defining the irritative zone.

VAN HOORDE Kirsten (Starting date : August 2010)
Updating and evaluation of polytomous logistic regression models
Thesis advisor : Sabine Van Huffel
The main purpose of this research is to generalize the updating and performance evaluation methods of dichotomous logistic regression models to polytomous logistic regression models, and more specifically to logistic regression models with a nominal outcome. Those methods facilitate the development of polytomous decision support systems for routine use by clinicians.

VANDEKERCKHOVE Steven (Starting date : October 2011)
Simulation of elastodynamic and electromagnetic wave propagation in nonlinear media
Thesis advisors : Herbert Degersem, Stefan Vandewalle, Koen Van Den Abeele
More information soon.

VANDERMEERSCH Antoine (Starting date : 1 November 2013)
A fast solver for problems with multivariate polynomials: exploiting sparsity and structure in the polynomial numerical linear algebra framework
Thesis advisor : Bart De Moor
Polynomial Numerical Linear Algebra, or PNLA, allows solving systems of multivariate polynomial equations, commonly done using algebraic geometry or homotopy methods, now using numerical linear algebra instead. Data structures and theoretical considerations in the PNLA framework need to be translated into efficient and robust algorithms in order to construct fast solvers and present itself as a competitive alternative. The combinatorial explosion that arises for such large systems with many variables when formulated as an eigenvalue problem can be counteracted by exploiting the sparsity pattern and structural replication of coefficients found among the supplied polynomials. The aim of this thesis is to provide such a translation using both algorithmic improvements and novel implementations as well as current computer architectures bundled into a working software package.

VANNIEUWENHOVEN Nick (Starting date : September 2010)
The tensor rank decomposition: truncation and identifiability
Thesis advisor : R. Vandebril, K. Meerbergen
Hitchcock's rank decompositions of tensors of order at least three have found application in a variety of fields such as algebraic statistics, chemometrics, and signal processing. The aim of this research is to further the theoretical understanding of the tensor rank decomposition, as well as develop practical algorithms for computing it. The research is focused on explaining metamorphoses in the properties of this decomposition, such as trunctability and identifiability, when transitioning from the case of second order tensors, i.e., matrices, to higher orders.

VARON PEREZ Jenny Carolina (Starting date : April 2011)
Decision Support Systems for Cardiovascular Diseases
Thesis advisor : Sabine Van Huffel
The goal of this research is to develop algorithms to detect acute situations such as epileptic seizures and sleep apnea events. Currently, the focus is on sleep apnea detection using single lead ECG. The respiratory signal and different features are extracted from the ECG and they can be used to detect breathing disorders during sleep. All the algorithms used to extract features, derive signals from the ECG and classify the different events must be adapted to work on a completely automated way. In addition, most of the well-known algorithms used to analyze ECG signals, are highly sensitive to outliers and artifacts, and hence the robustness and accuracy of the algorithms must be improved. This system should be used to monitor other kinds of disorders that can be detected using ECG signals, for example epileptic seizures.

VAULET Thibaut (Starting date : October 2016)
Thesis advisor : B. De Moor
More information soon.

VERCAMMEN Dominique (Starting date : October 2010)
Flux balance analysis for microbial dynamics
Thesis advisors : Jan Van Impe, Filip Logist
The objective of the PhD is the selection of objective functions which can be applied to describe a microbial growth curve. These objective functions will be implemented in static and/or dynamic FBA models which link macroscopic variables with microscopic information.

VERGAUWEN Bob (Starting date : October 2016)
Thesis advisor : B. De Moor
More information soon.

VERSCHEURE Diederik (Starting date : September 2005)
Modelling and identification of contact dynamics
Thesis advisors : J. Swevers, J. De Schutter
Robots are increasingly used to perform complex tasks, which often involve interaction and contact with unstructured environments. By taking the dynamic behavior of the environment into account, the accuracy, robustness and autonomy of intelligent robot systems can be considerably improved. The aim of this research is to develop a methodology to generate and update dynamic contact models. This involves the development of methods to generate, reduce, and select model structures, to design and optimize experiments, and to estimate the parameters of these models on-line.

VERSCHUURE Myriam (Starting date : January 2005)
Convex optimization applied to counterweight balancing of mechanisms
Thesis advisors : J. De Schutter, J. Swevers
This research focuses on improving the dynamic behavior of linkages through counterweight addition. This is an optimization problem in which the counterweight mass parameters are the optimization variables. It has previously been shown at the PMA division of K.U.Leuven that a basic variant of this problem (involving only rigid-body dynamics of planar four-bar mechanisms) can be reformulated as a convex program. The purpose of the current research is to extend these results to planar/spatial mechanisms with arbitrary complexity and to take into account additional dynamic criteria, such as the vibration of the supporting machine frame.

VERVLIET Nico (Starting date : 15 September 2013)
Compressive sampling based signal separation
Thesis advisor : Lieven De Lathauwer
Tensors (loosely called 'higher-dimensional matrices') provide a richer and more natural representation for many large data sets than 'flat' matrices. Often, these tensors are structured and this fact can be exploited to decompose these tensors into compact and sophisticated multilinear models. Using these models, large data sets can be handled efficiently. The computation of a decomposition of a very large data set (so called Big Data), however, is hard due to the curse of dimensionality. In this PhD research we will develop new compressed sensing type algorithms to alleviate and even remove this burden, and we will develop new applications for large and computationally challenging data sets in data analysis and signal processing.

VOLCKAERT Marnix (Starting date : January 2008)
ILC for highly-nonlinear systems
Thesis advisors : J. Swevers, M. Diehl, J. Schoukens
Details to follow soon.

VUKOV Milan (Starting date : December 2010)
Embedded Optimization Algorithms for Nonlinear MPC and Nonlinear MHE
Thesis advisor : Moritz Diehl
The research focuses on the development of efficient optimization algorithms for nonlinear model predictive control (NMPC) and nonlinear moving horizon estimation (NMHE). These algorithms need to be tailored to applications with high sampling rates (in the kHz range) and shall be verified on real-world benchmark problems. The optimization methods will most likely be based on the well-established real-time iteration (RTI) scheme and might make use of ideas from sequential convex programming.

WIDJAJA Devy (Starting date : September 2010)
Cardiorespiratory dynamics : algorithms and applications
Thesis advisor : Sabine Van Huffel
The interactions between several systems of our human body are complex processes that require a lot of research. Two important and strongly related systems are the cardiovascular and the respiratory system. A well-known example of these interactions is found in respiratory sinus arrhythmia (RSA), which is the modulation of the heart rate due to respiration. The main objectives of this research are: the derivation of a respiratory signal from the ECG (ECG-derived respiration); the composition of a surrogate ECG signal without the respiratory component, and; the identification of the common dynamics between respiration and the ECG.

WILLEMEN Tim (Starting date : 11 February 2013)
Biomechanics based modeling of sleep
Thesis advisors : Jos Vandersloten, Sabine Van Huffel, Bart Haex
Sleep is a complex process influenced by a large number of parameters. Smart bedding systems try to measure and actuate some of these parameters in order to achieve optimal sleep conditions. Hereby, sleep quality estimation is an important aspect. Smart bedding systems try to estimate sleep quality based on off-body detectable physiological signals, such as heart rate, respiration and movement. The goal of this project is to measure these parameters off-body in a mechanical way, use them to estimate sleep quality and investigate how this information could be used to improve sleep.

WITTERS Maarten (Starting date : September 2005)
Distributed control of a semi-active and an active suspension system of a passenger car
Thesis advisors : J. Swevers, P.Sas
The aim of the study is to develop a distributed model free controller for a semi-active or an active suspension system of a passenger car. The model free approach allows on-line tuning of the controller based on subjective road tests. For the distributed implementation, 'intelligent' shock absorbers have to be developed which exchange information using a communication network.

WUYTS Sam (Starting date : September 2014)
Monitoring and diagnosis of processes with bunched quality measurements
Thesis advisors : Jan Van Impe, Geert Gins
This research project investigates how quality measurements that are only available as low-frequency averages over longer time periods when compared to typical process measurements (e.g., daily quality averages vs. second-by-second online sensors) can be exploited for quality prediction and online process monitoring of continuous processes. Hereto, novel multivariate techniques will be developed to specifically deal with this multi-rate nature. The developed procedures will be validated on an industrial production process.

WYNANTS Laure (Starting date : October 2011)
Clinical prediction models based on multicentric studies: methods for clustered data
Thesis advisor : Sabine Van Huffel
Prediction models, with the purpose of helping clinicians make a diagnosis or prognosis, are increasingly being built based on datasets to which multiple hospitals have contributed. This has the advantage that the resulting model is applicable in a broad range of settings. It is nonetheless problematic that the current methodology doesn’t take into account the clustered nature of the data, which is typical for multicentric studies with patients clustered in hospitals. The purpose of this research is to develop coherent methods for the development, validation and application of prediction models based on multicentric data. It will be investigated how sample size planning, variable selection and model validation has to be performed in multicentre studies, as well as how an existing multicentre model can be applied to a new and unrelated centre.

ZHANG Huili (Starting date : September 2012)
The development of powder circulation systems for solar energy capture and storage
Thesis advisor : Jan Degreve and Jan Baeyens
Energy storage technologies are a strategic and necessary component for the efficient utilization of renewable energy sources and energy conservation.To provide a durable and widespread primary energy source, solar energy must be captured, stored and used in a cost-effective fashion. Solar energy is of unsteady nature, both within the day (day-night, clouds) and within the year (winter-summer). The storage of solar energy is critical if a significant portion of the total energy needs to be provided by solar energy. The main objective of this PhD thesis is to develop the powder circulation system which can be used in solar energy capture and storage and high temperature heat recovery in industry.

ZINK Rob (Starting date : 3 September 2013)
Mobile EEG neurofeedback
Thesis advisor : Sabine Van Huffel
Development of new methods that allow for mobile EEG recordings is needed in order to be able to measure outside the ‘traditional’ laboratory/hospital setup. Detection of artefacts and feature selection in large (repeatedly measured) datasets are the main topics in this project. Tensor based techniques will be explored to fully exploit the multidimensional aspects of the EEG. This will be validated on patients in a vegetative or minimally conscious state (revival from coma). Based upon repeated mobile EEG recordings at the patients’ homes we aim to detect the level of consciousness. This would be deemed successful if we outperform existing behavioural measures. The (tensor) methods should not be restricted to this problem, but be applicable in many forms of neurofeedback.

Post-Docs

AGUDELO Oscar Mauricio (Starting date : October 2005, Postdoc since November 2009)
Implementation of a nonlinear model predictive controller for the river Demer
Host : Bart De Moor
Flooding of rivers is a worldwide problem with severe consequences. For the particularly case of Belgium, the river Demer has caused several floods during periods of heavy rainfall. It is clear that the current control system, a rule-based three position controller whose rules are derived from expert knowledge, is not good enough. This is due to the fact that the control actions are only based on local information, and therefore they are not necessarily optimal for the entire river system. Furthermore, rainfall predictions are not taken into account. Model Predictive Control (MPC) does not suffer from these limitations and previous simulation studies have shown that this advanced control strategy leads to a much better flood control. The goal of this project is to design and implement a nonlinear Model Predictive Controller for the Demer. In this industrial project, several partners are involved, namely, Antea group, IDMC, IPCOS NV, Cofley Fabricom, KU Leuven ESAT, KU Leuven Hydraulica, and VMM (Vlaamse Milieu Maatschappij).

ALAIZ Carlos M. (starting date : 1 September 2015)
Extensions on the Relationship between lasso and SVMs
Promotor : Johan Suykens
Regularized models have gained a renewed interest nowadays due to the apparition of big data problems. These models are generally described as those minimizing an error term, which measures how well the model fits the data, and a regularization term, to avoid over-fitting and enforce desirable properties. Two important linear representatives of regularized models are the Lasso for regression, based on an L1 regularization, and the linear SVM for classification, based on the maximization of the margin. The training problems of some instances of these two models have been proved to be equivalent recently. The main goal of this research is to deepen into this equivalence, transferring the advances into one field to the other, in order to define new models and algorithms inspired by this relationship.

CAICEDO DORADO Alexander (Starting date : 1 February 2009 – Postdoc since 7 June 2013)
Development of a multivariable nonlinear framework for cerebral hemodynamic and brain function monitoring in neonates
Host : Sabine Van Huffel
The aim of this project is to provide a framework for the use of near Infrared Spectroscopy (NIRS) and amplitude EEG (aEEG) measurements for online monitoring of cerebral hemodynamic (CH) and brain function in neonates. Studies of CH using NIRS have applied linear univariate approaches leading to inconclusive result. In addition, the relation between aEEG and NIRS has not been explored using quantitative analysis. Improvement of the current methods can be achieved through the use of nonlinear multivariate methods. However, the interpretation of the results given by nonlinear methodologies, in terms of the physiology of the underlying processes involved in CH regulation, is challenging.

CORVELEYN Samuel
Development and analysis of algorithms for solving tensor-structured systems of ordinary differential equations in a low-rank tensor format
Host : K. Meerbergen
The semi-discretization of time-dependent partial differential equations by the method of lines leads to a system of ordinary differential equations. If the equation needs to be solved over a range of different values of the parameters in the differential equation or initial/boundary conditions, then the additional discretization of the parameter space leads to a possibly very high-dimensional tensor-structured system of ordinary differential equations. Due to the so-called curse of dimensionality, this becomes prohibitively expensive to solve for increasing dimension. In this work, we aim to develop and analyse different algorithms for solving the tensor-structured system approximately in a data-sparse low-rank tensor format. The low-rank tensor formats that we will mainly focus on are the hierarchical Tucker tensor format and the tensor train tensor format.

CROITOR Anca (Starting date : August 2007, postdoc since April 2012)
Advanced signal processing and classification techniques for data fusion of multimodal information with clinical applications
Host : Sabine Van Huffel
This study aims to improve and facilitate the clinical application of Magnetic Resonance Spectroscopy (MRS) by the development of advanced classifiers with adaptive learning abilities combining segmentation, signal processing and pattern recognition and using multimodal information such as Magnetic Resonance Imaging (MRI), MRS, MRS Imaging and high resolution ex vivo MRS.

DE ROECK Wim
Aero-acoustics, 1D modelling, numerical modelling

DE VOS Maarten (Starting date : October 2005 – postdoc since December 2009)
Multilinear decomposition methods and applications in neuroscience
Host : Sabine Van Huffel
The goal is to improve existing multilinear decomposition methods and to develop new ones. The methods will also be applied to applications in neuroscience, e.g. analysis of ElectroEncephaloGram (EEG) for the localisation of epileptic foci, and analysis of functional Magnetic Resonance Imaging data (fMRI).

DOMANOV Ignat (started predoc student October 2008, PhD student since 20 April 2009 – Postdoc since 19 September 2013)
Tensor-Based Signal Separation
Host : Lieven De Lathauwer
Signal processing makes more and more use of techniques based on multilinear algebra. In my thesis, two tensor decompositions that are at the heart of these methods will be further studied, namely the Canonical / Parallel Factor Decomposition and the Block Term Decomposition. The work will include the study of uniqueness and the derivation of algorithms. The results will be used to develop new methods for blind source separation.

FANUEL Michael (Postdoc, starting date 1 March 2015)
Physics inspired methods for Support Vector Machines and complex networks
Promotor: Johan Suykens
The analogies between data analysis methods and physics are often striking. The cross-fertilization between both fields is of great potential interest. On the one hand, methods from quantum physics (such as coherent states and wavelets) can contribute to formalize extensions of Support Vector Machines. As an example of this interplay, the role of transforms in the primal-dual framework of LS-SVM was recently emphasized. On the other hand, many datasets are actually network data or can be organized in graphs or networks. Networks are interesting instances of discrete geometrical objects, whose topology and geometry are relevant in practice. In this context, we consider theories such combinatorial Hodge theory which uses discrete differential operators in order to study the topology of graphs. Similarly, we study the relevance of dynamical processes on networks in connection with their community structure.

FENG Yunlong (Postdoc, starting date : 1 October 2013)
Nonlinear Learning Machines: Theoretical Understandings and Applications
Host : Johan Suykens
Nonlinear learning machines, e.g., kernel machines, have attracted particular attentions for decades. Various core models have also been proposed. By employing nonlinear machines, nonlinear information may be exploited from data. However, this also leads to barriers in theoretical understandings. For example, we are still far from fully understanding the kernel machines. Our studies concentrate on proposing and understanding various concrete learning algorithms, including, but not limited to, information-theoretic learning algorithms and adaptive sparse learning algorithms. We are concerned with their theoretical foundations, sparseness, computational issues and also real-world applications.

FERNANDEZ PASCUAL Angela (Postdoc, starting date : 1 April 2016)
Host : Johan Suykens
More information soon.

FRANDI Emanuele (Starting date : 1 January 2014)
Advanced optimization tools for Kernel learning and data-driven modelling
Host : Johan Suykens
The broad scope of this research is to investigate, both from a theoretical and a practical point of view, high-performance optimization methods for large-scale data-driven modelling applications. An important algorithmic tool in this context is provided by first-order solvers of Conditional Gradient type (specifically, Frank-Wolfe and Generalized Frank-Wolfe), whose applicability encompasses traditional SVM classification as well as other important problems in data-driven modelling, such as Lasso regression and matrix completion. We also aim to provide open-source software tools implementing these solvers on a variety of problems, and possibly explore the use of modern computational resources (such as parallel and GPU computing) to maximize their performance.

FRIEDHOFF Stephanie
Robust optimization with parabolic PDEs
Host : S. Vandewalle
Optimization with partial differential equations (PDEs) is of interest in many fields, such as in applications within the Optimization in Engineering Center (OPTEC). In this area of research, PDEs are used to model the problems mathematically, and the aim is to find control functions that minimize cost functions for the PDE-systems. Often, model parameters are not exactly known and, thus, efficient algorithms for robust optimization of PDE-constrained systems taking uncertainty into account are needed. In this research project, the design of such algorithms to control parabolic systems is considered. A special focus lies on applying multigrid and parallel-in-time-integration techniques in the framework of optimal control to develop efficient and fast numerical methods.

GAUTHIER Bertrand (starting date : 1 March 2015)
Promotor : Johan Suykens
More information soon.

GAWAD Jerzy
Multiscale modeling and simulation

GINS Geert (Starting date : January 2003, post-doc since 01/01/2008)
Host : Prof. Jan Van Impe
In his PhD research, data-driven techniques have been exploited to model and predict chemical and biochemical processes. This line of research is continued in his postdoc.

HANG Hanyuan (Starting date: April 7th, 2015)
Statistical Learning Theory
Promotor: Johan Suykens
Statistical Learning Theory (SLT) is a mathematical framework that deals with how machines or mechanisms predict results through a process of learning which is thought to be an alteration of behavior as a result of individual experience. When an organism can perceive and change its behavior, it is said to learn. In the framework of SLT, input data is analyzed in accordance to a desired application result.The corresponding factors that are thought to have an effect on the final outcome of the values are then considered, drawing out a potential relationship. The relationship between these factors will then form the basis of this algorithm in which the model "learns" and predicts data -- the desired end result of this "learning".

HELSEN Jan
Multibody simulation

HEIRMAN Gert
Multibody simulation, model reduction

HUANG Xiaolin (Starting date : April 2012)
Nonlinear classification, identification, and optimization via continuous piecewise linear analysis
Host : J. Suykens
A continuous piecewise linear function equals a linear or affine function in each of subregions which tessellate the domain. By continuous piecewise linear identification, one nonlinear system can be approached by a set of linear systems and then linear techniques are applicable. We are going to do research on CPWL classification and related optimization technique, including piecewise linear feature mapping, piecewise linear loss function, and some applications.

HUNYADI Borbala (Starting date : 17 September 2009 – postdoc since 2 June 2014)
Epileptic seizure detection
Host : Sabine Van Huffel
This research aims at developing automated analysis techniques which can support the presurgical evaluation of refractory epilepsy patients. Electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) measure the changes of brain activity in time over different locations of the brain. As such, they provide valuable information on the nature, the timing and the spatial origin of the epileptic activity. Unfortunately, both techniques record activity of different brain and artefact sources as well. Hence, EEG and fMRI signals are characterised by low signal to noise ratio. Data quality and the vast amount of recordings make the interpretation of these signals challenging. To overcome these difficulties, tensor-based blind source separation techniques are proposed which can exploit the characteristic spatio-temporal structure underlying the inherently multidimensional epileptic brain signals.

LANGONE Rocco (Starting date 15 February 2010 – postdoc since 2 July 2014)
Kernel spectral clustering for community detection in complex networks
Host : Johan Suykens
My research deals with clustering and community detection in complex networks. In the framework of static clustering, I introduced a soft kernel spectral clustering (SKSC) algorithm, which can better deal with overlapping clusters and provides more interpretable outcomes. Moreover, a whole strategy based upon KSC for community detection of static networks is proposed, where the extraction of a high quality training sub-graph, the choice of the kernel function, the model selection and the applicability to large-scale data are key aspects. Regarding dynamic clustering, a novel clustering algorithm for the analysis of evolving networks called kernel spectral clustering with memory effect (MKSC) has been developed. In this algorithm the temporal smoothness between clustering results in successive time steps is incorporated at the level of the primal optimization problem, by properly modifying the KSC formulation. Also, a new algorithm called incremental kernel spectral clustering (IKSC) for online learning of non-stationary data has been proposed. This ambitious challenge is faced by taking advantage of the out-of-sample property of kernel spectral clustering to adapt the initial model, in order to tackle merging, splitting or drifting of clusters across time. Real-world applications of the aforementioned algorithms, developed during my PhD, include image segmentation, time-series clustering, community detection of static and evolving networks, predictive maintenance in industrial machines.

LANGONE Rocco (starting date 15 February 2010 – postdoc since 2 July 2014)
Clustering evolving data using kernel-based methods
Promotor : Johan Suykens
My research deals with clustering and community detection in complex networks. In the framework of static clustering, I introduced a soft kernel spectral clustering (SKSC) algorithm, which can better deal with overlapping clusters and provides more interpretable outcomes. Moreover, a whole strategy based upon KSC for community detection of static networks is proposed, where the extraction of a high quality training sub-graph, the choice of the kernel function, the model selection and the applicability to large-scale data are key aspects. Regarding dynamic clustering, a novel clustering algorithm for the analysis of evolving networks called kernel spectral clustering with memory effect (MKSC) has been developed. In this algorithm the temporal smoothness between clustering results in successive time steps is incorporated at the level of the primal optimization problem, by properly modifying the KSC formulation. Also, a new algorithm called incremental kernel spectral clustering (IKSC) for online learning of non-stationary data has been proposed. This ambitious challenge is faced by taking advantage of the out-of-sample property of kernel spectral clustering to adapt the initial model, in order to tackle merging, splitting or drifting of clusters across time. Real-world applications of the aforementioned algorithms, developed during my PhD, include image segmentation, time-series clustering, community detection of static and evolving networks, predictive maintenance in industrial machines.

LEE Geunseop (Starting date : 1 October 2014)
Efficient Update and downdate of tensor decompositions
Host : Lieven De Lathauwer
In many data processing applications, it is necessary to continuously update data tensors with a new tensor. However, due to its large number of elements, computation of a tensor decomposition from scratch at every update is prohibitively expensive. Our research aims to develop efficient adaptive subspace tracking algorithms to update or downdate a tensor decomposition by reusing its previous decomposition.

LIU Yipeng (Starting date : December 2011)
Compressive sensing for biomedical signals
Host : Sabine Van Huffel
Sparsity decreases computations, memory usage, and data communications. It is required by increasingly sophisticated information processing devices. The increase in biomedical measurement techniques (EEG, fMRI, EMG, ECG, Magnetic Resonance Spectroscopy (MRS), NIRS, …) for diagnosis and follow-up of human diseases, strongly requires compression in order to keep the dataflow tractable and save battery power in wireless applications. To extract the relevant information, the biomedical data sets should be represented sparsely in a context-dependent basis. In general however, this basis is not known. The goal of this project is:

• To identify appropriate bases or learn dictionaries to represent each of the different biomedical signals sparsely and compute the coefficients in this representation
• To study how nearly-sparse signals can be accurately approximated by sparse signals as preprocessing for compressive signals
• To optimize compressive sensing techniques for multichannel applications

MACH Thomas (Starting date : September 2012)
Exploiting unconventional QR-algorithms for fast and accurate computations of roots of polynomials
Host : Raf Vandebril
The embracing goal of this research project is the development of new unconventional (generalized) QR-algorithms for accurately, reliably and fastly computing roots of polynomials built upon a solid theoretical foundation. This includes the investigation of deflation and shift techniques as well as the implementation of blocked versions.

MATIC Vladimir (starting date 24 March 2010 – Postdoc March 26, 2015)
Neonatal EEG signal processing
Promotor : Sabine Van Huffel
The main objective will be to analyse the background activity of the neonatal EEG. It is important to describe and develop an automated method for detecting changes in EEG bursts in neonates with varying severity of encephalopathy. Apart from this, research should also include compressive sensing theory and possible future applications to neonatal EEG. The potential benefits consists in compressing the neonatal EEG data, improving seizure detection algorithms and artifact removal.

MEHRKANONN Siamak (PhD starting date : 01 Dec 2011 – Postdoc since July 2, 2015)
Incorporating prior knowledge into the learning task
Promotor : J. Suykens
The primary objective of my research is to explore the possibilities of incorporating the prior knowledge into the learning framework. In many real-life problems one usually encounters a huge amount of unlabeled data points whereas the portion of the labeled data points is few. This is due to the fact that the acquisition of labeled data often requires a skilled human agent or a physical experiment which are of course costly. Semi-supervised learning (SSL) is a framework in Machine Learning which aims at learning from both unlabeled and labeled data points. I have proposed a multi-class semi-supervised learning algorithm (MSS-KSC) and successfully applied it in classification and clustering. The core model is Kernel Spectral Clustering (KSC), a completely unsupervised algorithm. The MSS-KSC approach is formulated as an optimization problem in the primal and dual settings where the side-information (labeled data points) is incorporated to the core model (KSC) using a regularization term. In addition I have made the MSS-KSC approach applicable for dealing with large-scale and data stream such as in video segmentation.

SIGNORETTO Marco (Starting date : February 2007 & 01 October 2012 – postdoc since December 2011)
Optimization modelling using kernel methods and tensors
Host : Johan Suykens
The goal of this study is to combine the algebraic know-how developed along with tensor-based methods with kernels, convex optimization, sparsity and statistical learning principles. This encompasses both theoretical and algorithmical studies and is expected to impact significantly in all those contexts where data are structured and the number of available observations is limited. Two main objectives are then: A Develop and analyze a systematic kernel-based framework to tensorial data analysis. B Based on convex optimization, design new algorithms that combine tensors and kernels.

SIMA Diana (Starting date : March 2002, Postdoc since May 2006)
Advanced methods for in vivo metabolite quantification based on Magnetic Resonance Spectroscopy
Host : Sabine Van Huffel
This research focuses on automatic and robust methods for quantifying metabolite concentrations and for tissue differentiation in the human brain using magnetic resonance spectroscopy (MRS). We consider spatial prior knowledge in 2D/3D MRS imaging, as well as multi-modal data fusion of MRS with other imaging modalities acquired from the same patients. Our approach involves automatic signal (pre)processing, nonlinear optimization, and blind source separation methods. The main application in collaboration with the division of Radiology, U.Z.Leuven, is a longitudinal study of brain glioma patients treated with vaccination therapy.

SORENSEN Michael (Starting date : June 2010)
Structured Tensors with application in signal processing
Host : Lieven De Lathauwer
Tensor tools are increasingly used in signal processing. In many cases, such as system identification or equalization, the involved tensor decompositions are structured. Hence, the development of numerical methods to deal with structured tensors is very important. By taking the structure of the tensor into consideration one can expect to obtain better identifiability results, and more efficient and robust estimates of the parameters of the structured tensor.

TELEN Dries (Starting date : October 2010)
Advanced numerical methods for optimal experiment design: from theory to implementation in the (bio)chemical process industry
Thesis advisors : Jan Van Impe, Filip Logist, Moritz Diehl
This research focuses on the development of advanced numerical techniques and tools that are required to allow a practical implementation of Optimal Experiment Design techniques in the chemical and biochemical process industry.

VALLERIO Mattia (Starting date : September 2010)
Advanced numerical techniques for computer aided decision making in the (bio)chemical industry
Thesis advisors : Jan Van Impe, Filip Logist
In this research advanced numerical techniques and software are developed to support real-time decision making for the design and optimization of large-scale dynamic (bio)chemical processes.

VAN BELLE Vanya (Starting date : October 2006, postdoc since December 2010)
Interpretable clinical decision support systems
Host : Sabine Van Huffel
To decide upon the optimal treatment strategy for patients, accurate diagnostic and prognostic tools are crucial to assist clinicians. Unfortunately the complexity of the mathematics of most clinical decision support tools prohibits their use in practice. Different simplification methods have been proposed but all remain with large disadvantages. Simple models are interpretable but lack optimal performance properties. Advanced models generally perform better but are not interpretable, resulting in legal issues. This project will combine the best of these worlds and define a unifying framework for score models inhibiting clinical relevance. This umbrella model will include medical statistics, machine learning methods and optimization problems in combination with clinical aspects. The model will naturally extend the clinical work flow and allow application in clinical practice and home-monitoring settings. Methodological aspects include variable selection, regularization, convexity, duality, penalized likelihood and computational efficiency. The framework will be general and will be able to deal with (multi-class) classification, prognostic and multi-level data. The methods will be validated on different clinical problems within oncology and gynaecology.

VAN CALSTER Ben (Starting date : September 2004 – postdoc since March 2008)
Linear and nonlinear predictive models for medical classification problems using patient data and expert knowledge
Host : S. Van Huffel
The aim is to both develop and evaluate linear and nonlinear predictive models combining patient data and expert knowledge in medical classification problems. These models are based on generalized linear models, artifical neural networks, support vector machines and their integration. Attention will be given to their statistical characteristics and clinical importance.

VAN HERPE Tom (Starting date : October 2003, Postdoc since April 2008)
Clinical validation and filing for regulatory approval of a blood glucose regulator for the Intensive Care Unit
Hosts : Bart De Moor, G. Van den Berghe, D. Mesotten
Critical illness typically causes elevated blood glucose concentrations, which have been associated with increased mortality. Strictly normalising these blood glucose levels by intensive insulin therapy, coined tight glycaemic control (TGC), decreased morbidity and mortality in three Leuven randomised clinical trials. In contrast, a recent, large multi-centre study showed an increased mortality risk in patients treated with TGC. Despite these contrasting results, the intensive care community is persuaded that controlling blood glucose levels is imperative as a preventative strategy against complications and death in the intensive care unit (ICU). At the same time it has become crystal clear that TGC is a very complex intervention, making implementation in clinical ICU-practice particularly difficult. Notably, the invariably higher incidence of severe hypoglycaemia is worrisome.
To simplify the implementation of TGC, a computerised algorithm to control blood glucose levels in critically ill patients, named LOGIC-Insulin, was developed in collaboration between the Dept. of Intensive Care Medicine and the Dept. of Electrical Engineering of the KULeuven. Proof-of-concept studies showed that LOGIC-Insulin improved TGC and lowered the incidence of severe hypoglycaemia, compared to standard TGC performed by the nursing staff. This project represents the late stage, clinical validation of the LOGIC-Insulin algorithm. The ultimate utilisation aim is to successfully implement TGC in general clinical ICU-practice with the aid of the LOGIC-Insulin algorithm. A single-centre, as well as a multi-centre study will compare the impact of LOGIC-Insulin on clinically relevant outcome measures, such as severe hypoglycaemia, in large patient groups, in comparison with standard, nurse-directed TGC. Further, the clinically validated software will be filed for CE marking.

VARON PEREZ Jenny Carolina (starting date 15 April 2011 – postdoc since April 30, 2015)
Decision Support Systems for Cardiovascular Diseases
Promotor : Sabine Van Huffel
The goal of this research is to develop algorithms to detect acute situations such as epileptic seizures and sleep apnea events. Currently, the focus is on sleep apnea detection using single lead ECG. The respiratory signal and different features are extracted from the ECG and they can be used to detect breathing disorders during sleep. All the algorithms used to extract features, derive signals from the ECG and classify the different events must be adapted to work on a completely automated way. In addition, most of the well-known algorithms used to analyze ECG signals, are highly sensitive to outliers and artifacts, and hence the robustness and accuracy of the algorithms must be improved. This system should be used to monitor other kinds of disorders that can be detected using ECG signals, for example epileptic seizures.

WAGNER Andrew (Starting date : September 2011)
Sensors and Optimal State Estimation for Carousel-Launched Power Generating Kites
Host : M. Diehl
As part of the ERC HIGHWIND project on Simulation, Optimization and Control of High-Altitude Wind Power Generators, Dr. Andrew Wagner is investigating sensor and actuator configurations for power producing kites, as well as fully automated launching and landing sequences. The light software is undergoing extensive testing on an indoor prototype, and will be later verified on a larger outdoor prototype.

YANG Yuning (Starting date : 1 October 2013)
Tensor-based optimization and learning algorithms: theoretical and algorithmic aspects
Host : Johan Suykens
In many real world applications, matrix-based models are usually outperformed by higher-order tensor-based models benefiting from the abundant information concealed in the structures of high-order tensors. A central focus in tensor-based optimization and learning problems is information extraction from the tensorial data set. Starting from tensor-based optimization algorithms, we are going to propose various regularized learning algorithms to achieve this goal. Correspondingly, theoretical analysis will be presented and algorithmic aspect will be also covered.

YZELMAN Albert-Jan (Starting date : September 2011)
Intel ExaScience Lab
Hosts : Dirk Roose and Karl Meerbergen
Development of high-performant and highly-scalable algorithms in support of numerical codes, or in support of the parallelisation thereof. Special focus is on the sparse matrix--vector multiplication on highly non-uniform shared-memory architectures. Future kernels include the matrix powers kernel, and the incorporation of these kernels within parallel iterative solvers. These will be leveraged to full distributed-memory systems.

ZIELINSKI Przemyslaw
Accelerated micro/macro Monte Carlo simulation of stochastic differential equations
Host : G. Samaey
The main aim of my research is to analyse a micro/macro acceleration technique for the Monte Carlo simulation of dilute solutions of polymers modelled as a stochastic differential equation. In the model there is a separation between the (fast) time-scale, on which individual trajectories of the stochastic process need to be simulated, and the (slow) time-scale on which we want to observe the (macroscopic) function of interest. The algorithm combines short bursts of microscopic simulation with a macroscopic extrapolation step. During the extrapolation the microscopic ensemble needs to be matched onto the extrapolated macroscopic state. This is inference problem which calls for adopting the 'best guess' for the microscopic probability distribution. My research project involves the analysis of inference principles based on minimization of non-metric distances between probabilities, the applications to the matching problem and the study of convergence of proposed micro/macro acceleration method.

Useful links : UCL | KUL | UGent | VUB | ULg | UMons | UNamur | Stanford | Princeton | MIT