1 

Solving Stochastic PDEs with Approximate Gaussian Markov Random Fields using Different Programming Environments
This thesis is a study on the implementation of the Gaussian Markov Random Field (GMRF) for random sample generation and also the Multilevel Monte Carlo (MLMC) method to reduce the computational costs involved with doing uncertainty quantification studies. The GMRF method is implemented in different programming environments in order to evaluate the potential performance enhancements given varying levels of language abstraction. It is seen that the GMRF method can be used to generate Gaussian Fields with a Mat{\'e}rn type covariance function and reduces the computational requirements for large scale problems. Speedups of as much as 1000 can be observed when compared to the standard Cholesky Decomposition sample generation method, even for a relatively small problem size. The MLMC method was shown to be at least 6 times faster than the standard Monte Carlo method and the speedup increases with grid size. It is also seen that in any Monte Carlo type methods, a Krylov subspace type solver is almost always recommended together with a suitable preconditioner for robust sampling.
This thesis also studies the ease of implementation of these methods in varying levels of programming abstraction. The methods are implemented in different languages ranging from the most common language used by mathematicians (MATLAB), to the more performance oriented language (C++PETSc/MPI), and ends with one of the newest programming concept (ExaStencils). The GMRF method featured in this thesis also is one of the earliest application to be implemented in ExaStencils.

[PDF]
[Abstract]

2 

An Online Patient Scoring System
At this moment when a clinician needs to conduct a study he will send allhis patients at various moments a questionnaire form. After retrieving theseforms the clinician will manually process all the data by hand. This is a time consuming job and could be automatized. During this project we created an application that automates this process. When using the developed application the only thing a clinician has to do is create forms, add patients and tell the system when a patient has to receive a form.

[PDF]
[Abstract]

3 

Improvement and Adaption of Bearing Monitoring System
With the ongoing developments in the offshore wind turbine industry, weight reduction is an important aspect. With these weight reduction, manufacturers are trying to increase the effciency and size of their turbines. An example of such a technique is a single bearing drive train. These designs use a special bearing which is extra stiff due to the support of axial raceways. Manufacturers are still proving the robustness of this promising design. At the moment, some manufacturers already constructed a prototype. An example of this prototype is the XD115 of XEMC Darwind. This turbine is permanently monitored with the use of a health monitoring system. This system mainly consists of accelerometers which are located throughout the whole turbine. This monitoring system can be extended by performing spectral analysis on the measured generator current. The purpose of this thesis is to investigate if it is possible to measure mechanical vibrations in the current output of a generator with full converter.
The bearing is one of the most complex elements in this single bearing drive train. There fore, the focus is put on monitoring the bearing. In order to investigate in this topic, two types of tests are performed, lab tests and full scale tests. During the lab tests, a small setup is made and some bearings are tested. The full scale tests are performed on the XD115 prototype. During the tests in the lab most of the characteristic bearing frequen cies could be measured. An increase in the amplitude of these frequencies was measured after damaging the bearing. In these lab tests, an increase in friction was also measurable. The characteristic bearing frequencies could be identied much easier in the frequency spectrum of the current than in the frequency spectrum of the accelerometer. The re sults of the tests in the full scale turbine were less promising. Neither in the frequency spectrum of the accelerometer and current, a characteristic bearing frequency could be identied. However, when the rotational speed synchronisation algorithm is used, some mechanical vibrations could be identied. The shaft speed and each time a blade passes the tower were identied in the frequency spectrum of the current. This means that a mechanical vibration was measured in the generator current.

[PDF]
[Abstract]

4 

Unit Root Testing for AR(1) Processes
The purpose of this study is to investigate the asymptotics of a first order auto regressive unit root process, AR(1). The goal is to determine which tests could be used to test for the presence of a unit root in a first order auto regressive process. A unit root is present when the root of the characteristic equation of this process equals unity. In order to test for the presence of a unit root, we developed an understanding of the characteristics of the AR(1) process, such that the difference between a trend stationary process and a unit root process is clear.
The first test that will be examined is the DickeyFuller test. The estimator of this test is based on Ordinary Least Square Regression and a ttest statistic, which is why we have computed an ordinary least square estimator and the test statistic to test for the presence of unit root in the first order auto regressive process. Furthermore we examined the consistency of this estimator and its asymptotic properties. The limiting distribution of the test statistic is known as the DickeyFuller distribution. With a Monte Carlo approach, we implemented the DickeyFuller test statistic in Matlab and computed the (asymptotic) power of this test. Under the assumption of Gaussian innovations (or shocks) the limiting distribution of the unit root process is the same as without the normality assumption been made. When there is a reason to assume Gaussianity of the innovations, the Likelihood Ratio test can be used to test for a unit root.
The asymptotic power envelope is obtained with help of the Likelihood Ratio test, since the NeymanPearson lemma states that the Likelihood Ratio test is the point optimal test for simple hypotheses. By calculating the likelihood functions the test statistic was obtained, such that an explicit formula for the power envelope was found. Since each fixed alternative results in a different critical value and thus in a different unit root test, there is no uniform most powerful test available. Instead we are interested in asymptotically point optimal tests and we will analyze which of these point optimal tests is the overall best performing test. By comparing the asymptotic powercurve to the asymptotic power envelope for each fixed alternative we could draw a conclusion on which fixed alternative results in the overall best performing test.
On the basis of the results of this research, it can be concluded that there does not exist a uniform most powerful test, nonetheless we can define an overall best performing test.

[PDF]
[Abstract]

5 

Scalable JoinPattern Matching for Observables
As more and more of our systems become reactive, and have to scale to evergrowing demands, it becomes clear that the democratization of asynchronous programming is an important research subject. We present a novel programming construct that enables declarative and scalable coordination of asynchronous streams of events. Our model is integrated with Scala’s native patternmatching. The coordination is driven by a novel nonblocking Joincalculus implementation with unique support for bounded buffers leveraging finegrained concurrency. We use Scala’s experimental, static meta programming facilities to make our implementation as flexible as a library, but stat ically optimizing like a compiler. Our microbenchmarks show an overall, average increase in throughput of 30% when compared to similar stateoftheart coordination mechanisms.

[PDF]
[Abstract]

6 

An Algorithmic Readout Approach for Thermal Conductivity Based CO2 Sensors
This thesis presents a new approach to reading out thermalconductivitybased gas sensors. This method is intended for the readout of a CMOS compatible resistive thermalconductivity transducer for indoor CO2 sensing applications, without requiring precision offchip components. Instead of accurately regulating the power dissipated in the transducer and measuring its temperature, the temperature and power dissipation are both measured using an algorithmic approach. A highresolution ADC digitizes the voltage drop across the transducer and the current through it, measured using an onchip reference resistor. Moreover, by digitizing several baseemitter voltages of an onchip bipolar transistor, a precision bandgap voltage reference is constructed in the digital domain, and accurate information about the ambient temperature is obtained, which is used to temperature compensate the voltage reference and the reference resistor. Thus, all necessary ingredients are obtained to calculate the power dissipation and temperature of the transducer, from which the thermal conductivity of the surrounding air, and hence CO2 concentration, can be obtained.
A prototype integrated circuit implementing this readout approach has been realized in 0.16um CMOS. It has been tested in a climate chamber in combination with a platinum resistor mimicking the transducer. The digitallyconstructed voltage reference has a temperature coefficient of 9ppm/°C, while ambient temperature is sensed with accuracy of ±0.2°C, with a resolution of 0.15°C. The resistance readings have an inaccuracy ranging between 1mOhms to 4mOhms on a nominal resistance of about 100Ohms (10ppm  40ppm) with a resolution of around 2mOhms in the temperature range from 10°C to 40°C; for the power measurements, the circuit achieved an accuracy between 0.03% and 0.06%, with an 800nW of resolution (in the same temperature range) which is one order of magnitude better than results presented in previous work.
Although no CO2 measurements have been performed, an estimated thermal resistance accuracy of around 2862ppm with a resolution of 155.64[K/W] should be possible, which would enable detection of the CO2 levels in the air with an accuracy of around 0.72% and a resolution of 7705ppm.

file embargo until: 20151128
[Abstract]

7 

Optimal boundary point control for linear elliptic equations
This master thesis is concerned with optimal boundary point and function control problems for linear elliptic equations subject to control constraints. The elliptic partial differential equation with Robin boundary condition is considered. The control is chosen as a linear combination of the Dirac delta functions in the point control problem. The weak formulation and optimality conditions are obtained for the function control problem. The main goal is to examine the existence of the weak solution and to derive optimality conditions of boundary point control. Introducing sufficient discretization methods such as a finite volume and a finite element methods, we obtain finite dimensional problem.
We apply efficient numerical methods including primaldual active set strategy, projected gradient and conjugate gradient methods. The test examples are presented clarifying the performance of numerical methods mentioned earlier. The conjugate gradient method is compared to the unconstrained matlab function QUADPROG and primaldual active set strategy is compared to the constrained QUADPROG. The projected conjugate gradient method is applied to improve the projected gradient method. Due to its importance, the sparse point and function control problems are studied. We apply primaldual active set strategy where the sparse parameter f is chosen differently. All of the results are presented in both point and function control problems.

[PDF]
[Abstract]

8 

Estimating the extreme value index for imprecise data
In extreme value theory the focus is on the tails of the distribution. The main focus is to estimate the tail distribution for a rounded data set. To estimate this tail distribution the extreme value index should be estimated, but due to the rounded data this extreme value index oscillates heavily. Therefore a correct estimate can not be obtained. By adding a small uniform stochast the rounded data can be smoothend out and in this way the oscillation can be cancelled.

[PDF]
[Abstract]

9 

Investigation of Different Solvers for Radiotherapy Treatment Planning Problems
Radiotherapy treatment planning involves solving inequality constrained minimization problems. The currently used interior point solver performs well, but is considered relatively slow. In this thesis we investigate two different solvers based on the logarithmic barrier method and Sequential Quadratic Programming (SQP) respectively. We argue that the behaviour of the logarithmic barrier solver is uncertain, thereby making it generally unreliable in this context. In addition we substantiate that the performance of the SQP solver is solid, but lacks efficiency in computing the minimizers of its related quadratic subproblems.
We conclude that without serious improvements, none of the solvers investigated are faster than the currently used interior point optimizer.

[PDF]
[Abstract]

10 

Determining Performance Boundaries and Automatic Loop Optimization of HighLevel System Specifications
Designers are confronted with high timetomarket pressure and an increasing demand for computational power. As a result, they are required to identify as early as possible the quality of a specification for an intended technology. The designer needs to know if this specification can be improved, and at what cost. Specification tradeoffs are often based on the experience and intuition of a designer, which in itself is not enough to make design decisions given the complexity of modern designs. Therefore, we need to identify the performance boundaries for the execution of a specification on an intended technology.
The degree of parallelism, required resources, scheduling constraints, and possible optimizations, etc. are essential in determining design tradeoffs (e.g., power consumption, execution time, etc). However, existing tools lack the capability of determining relevant performance parameters and the option to automatically optimize highlevel specifications to make meaningful design tradeoffs.
To address these problems, we present in this thesis a new profiler tool, cprof. The Clang compiler frontend is used in this tool to parse highlevel specifications, and to produce instrumented source code for the purpose of profiling. This tool automatically determines, from highlevel specifications, the degree of parallelism of a given source code, specified in C and C++ programming languages. Furthermore, cprof estimates the number of clock cycles necessary to complete a program, it automatically applies loop optimization techniques, it determines the lower and upper bound on throughput capacity, and finally, it generates hardware execution traces. The tool assumes that the specification is executed on a parallel Model of Computation (MoC), referred to as a Polyhedral Process Network (PPN).
The proposed tool adds new functionality to existing technologies: the estimated performance by cprof of PolyBench/C benchmarks, as compared to realistic implementations in FieldProgrammable Gate Arrays (FPGA) platforms, showed to be almost identical. Cprof is capable of estimating the lower and upper bound on throughput capacity, making it possible for the designer to make performance tradeoffs based on real design points. As a result, only the highlevel specification is used by cprof to assist in Design Space Exploration (DSE) and to improve
design quality.

[PDF]
[Abstract]

11 

Anomaly detection for internet banking using supervised learning on high dimensional data
Nowadays, a high number of transactions are performed via internet banking. Rabobank processes more than 10 million transactions per day. Most of these transactions are (part of) normal behaviour. On the other hand, some transactions are considered to be out of the ordinary. These anomalous events occur relatively infrequently (less than 10 per day). Employees, that try to find these anomalous events, combine the transactions data, historical knowledge of the anomalous events and their expertise to detect and quantify them. Several types of anomalies are considered to be interesting and so they are labelled. These anomalies need to be detected, so they can be prevented in the future. The employees try to find events similar to known anomalies. Characteristics of anomalies change over time and employees also need to detect this slightly changed, but similar, behaviour. It is not our goal to detect completely new types of anomalies. In this thesis, the focus lies on finding events similar to the known anomalies. In order to assist these employees, a model that uses the transaction data and incorporates known anomalous events is built. Our model is able to score new incoming transactions and use these to update the model parameters. The scores can be returned to the employees to assist them in finding transactions that are similar to a particular type of anomaly. The AdaGrad algorithm with diagonal matrices is used. Also, l1regularization is used on the parameter to create a more sparse solution.

[PDF]
[Abstract]

12 

Precedence Probability Penalised Nonparametric Maximum Likelihood Estimation with IntervalCensored Data
Many algorithms have been developed over the years to compute the nonparametric maximum likelihood estimate (NPMLE) for the cumulative distribution function of event times under interval censoring. The effective support of the NPMLE can be reduced to a set of disjoint socalled maximal intersection intervals. One feature of the NPMLE that is sometimes described as unsatisfactory is that, with most data sets, several of the estimated probabilities are equal to zero, implicitly inducing an ordering between overlapping intervals and leaving the user with a rather coarse estimate of the CDF.
We attempt to overcome this problem by focusing on the natural interval order the censored data describe. A Precedence Probability Matrix (PPM) for the data can be constructed, an object that uses information about all the possible orderings within the interval order. By estimating the true PPM – in itself a nontrivial task – and penalising the nonparametric loglikelihood for deviations from the data’s true PPM, smoother estimates of the CDF that better represent the possible ordering of the data can be acquired.

[PDF]
[Abstract]

13 

A Stochastic Approach for Selective Search Algorithms
This thesis introduces a new algorithm that generates candidate proposals for an object detection pipeline. We introduce Stochastic Selective Search (SSS), a segmentation based selective search method, which differs from previous work in two ways. First and most importantly, SSS is much faster than current stateoftheart algorithms while maintaining comparable accuracy. This is a result of our efficient stochastic segment merging process. Other work requires the computation of features to determine the order in which segments are merged. We show that currently used features from other work does not improve the results of SSS significantly and are therefore omitted. This makes our algorithm nearly twice as fast as the fastest prior selective search algorithms. Secondly, due to the stochastic merging process of SSS, it is not critically affected when two wrong segments are merged during the merging process, which leads to object proposals of higher quality. We show that SSS outperforms existing deterministic selective search methods while generating the same amount of proposals in less time. Additionally, we demonstrate the performance of our SSS algorithm in a stateoftheart object detection pipeline based on convolutional networks.

[PDF]
[PDF]
[Abstract]

14 

Robust Solutions for the ResourceConstrained Project Scheduling Problem: Understanding and Improving Robustness in Partial Order Schedules produced by the Chaining algorithm
Robustness is essential for schedules if they are being executed under uncertain conditions. In this thesis we research robustness in Partial Order Schedules, which represent sets of solutions for instances of the ResourceConstrained Project Scheduling Problem. They can be generated using a greedy procedure called Chaining, which can be easily adapted to use various heuristics. We use an empirical method to gain understanding of robustness and how the chaining procedure adds to this.
Based on the findings of an exploratory study we develop three models, each capturing aspects of robustness on a different level. The first model describes how a single activity is affected by various disturbances. The second model predicts how structural properties of Partial Order Schedules can reduce the effect of these disturbances. The third model describes how heuristics for the chaining procedure can influence these properties.
Using experimental evaluation, we found that the model is not complete. Experimental results did conform to the expectations set by the third model, but not of the second model. We therefore suspect that it is too simplistic for accurate predictions, but since it does match earlier observations we believe it is a good starting point for further understanding.

[PDF]
[Abstract]

15 

Modelling and Transient Analysis of a 10 MW Superconducting Wind Turbine Generator
The number of installed wind turbines is growing every year as the need for energy increases in an exponential manner. As the wind turbines are trying to fulfill these requirements, the demand for output power from each individual turbine is increasing as well. Although this is great, implementation of new technologies such as direct drive superconducting generators are required as they have a great potential for becoming a great contender for satisfying the increasing demand of individual turbine power output.
The efficiency and the mass reduce by using this technology but other issues have to be checked regarding its transient performance. Due to the low subtransient, transient and synchronous reactance, compared to a conventional machine, the transient properties such as shortcircuit current, field current and electromagnetic torque is much higher. This may cause problems as high values may damage the windings and turbine properties and even cause the superconducting field windings to lose its superconducting property. The thesis report will try to model a superconducting generator for a 10 MW wind turbine and analyse its transient properties. Also several important parameter calculations will be made taking into account the saturation due to iron in the generator. This simulation will be made for three different topologies of generator while taking into account their own individual generator geometry.

[PDF]
[Abstract]

16 

Evaluation of different storage systems for Apache Spark and Apache Hadoop
Big Data systems have been used for multiple years to solve problems that require scale. A framework takes care of scalability and resiliency issues, and allows the user to focus on relevant computation, in the form of map and reduce functions. In these Big Data systems, we currently see a shift from the traditional use of the Hard Disk Drive (HDD) towards inmemory computation. In this thesis, we have theoretically evaluated these two generations of Big Data systems, as well as two implementations, being Apache Hadoop and Apache Spark, in combination with Flash technology. We have also evaluated the possible use of Flash technology in these Big Data sys tems, by performing two experiments. Our first experiment exam ined the performance of Apache Spark versus Apache Hadoop for a representative iterative algorithm, and the performance degradation of Apache Spark under memory constrains. We have found that, for the chosen algorithm, Apache Spark performs equalorbetter compared to Apache Hadoop when data has to be loaded from the HDD, such as is the case of the initialization phase of a program. For the iterative part of our program, we have seen an overall speedup of 30, and a speedup of 100 for the map and reduce phases. In our second experiment, we evaluated two ways of using Flash, in particular using the IBM FlashSystem 840 connected to a Power8, in Apache Spark. We have first evaluated Flash technology with a mounted file system, and used this setup to replace the HDD as default spill loca tion. We have found that this was not valuable, as the possible performance improvements were negligible compared to the overhead generated by data aggregation and system calls. We then shifted our focus to CAPI connected Flash, and modified Apache Spark to spill intermediate data directly to the FlashSystem using key value pairs. In our experiment, while limiting the memory to a fixed amount to force spilling, we were able to remove 70% of the overhead caused by spilling. This was mainly overhead of Operating System (OS) involvement. In our future work, we will address the overhead caused by data aggregation, as we can write smaller amounts of data, because we are writing in key value pairs. We believe that, once this overhead is removed, Big Data systems can benefit from Flash technology, and especially CAPI Flash technology, as one can use a system with a limited amount of expensive DRAM and a large Flash backend, to solve larger problem sets while maintaining a performance equal to inmemory computation.

[PDF]
[Abstract]

17 

The choices of weights in the iterative convex minorant algorithm
In statistics one often encounters the problem of estimating a function based on a given dataset. Sometimes shape properties such as monotonicity of the function are known. This property can be used in a nonparametric regression model. The iterative convex minorant(ICM) algorithm can be used to compute an estimate of a convex regression function. In the ICM algorithm positive weights can be chosen arbitrarily. In this thesis we describe the (solution of the) isotonic regression problem, explain the ICM algorithm, describe the convex regression problem and present a simulation study to assess the effect of the choice of weights.

[PDF]
[Abstract]

18 

The stabilization and destabilization of Brouwer's Rotating Vessel
We beschouwen Brouwers roterende lichaam, een voorbeeld van een mechanisch systeem met het volgende paradoxale gedrag; zonder demping is er stabiliteit, met demping is er sprake van instabiliteit. Met modulatie van de hoeksnelheid wordt er een model afgeleid waarbij we pogen de instabiele gevallen te stabiliseren, hierbij is het meenemen van de hoekversnelling cruciaal. De differentiaalvergelijkingen die gevonden worden bevatten niet constante termen die langzaam in de tijd variëren met parameter Ɛ. We gebruiken de methode van middelen om een benadering te krijgen van de oplossing die op lange tijd geldig is. Met O(Ɛ) demping treedt er instabiliteit op ten gevolgen van demping voor alle resonanties. Met O(Ɛ²) demping kan er zowel stabiliteit en instabiliteit optreden afhankelijk van de amplitude van de modulatie voor één specifiek combinatie resonantie. Voor alle andere resonanties treedt er alleen instabiliteit ten gevolgen van O(Ɛ²) demping op. Brouwers roterende lichaam is dus zowel te stabiliseren als te destabiliseren.

[PDF]
[Abstract]

19 

Arbitragefree methods to price European options under the SABR model
In this thesis we discuss several methods to price European options under the SABR model. In general, methods given in literature are not free of arbitrage and/or inaccurate for long maturities. This led to the development of a new pricing approach. We extend the BCOS method from one dimension to two dimensions. This extension is necessary for application of a simplification of the BCOS method, the DCOS method, to the SABR model. In this pricing method we use the characteristic function of the discrete forward process and the Fourierbased COS method. It is possible to price European options under the SABR model for multiple strikes in one computation with the DCOS method. Besides valuing European options, we can also price Bermudan and discretely monitored barrier options with this pricing approach.

[PDF]
[Abstract]

20 

IMS messaging gateway in the Cloud
For mobile messaging service providers to endure in a competitive and dynamic market, it is vital to be flexible, which involves keeping up to date to technological developments, and to be cost effective. In order for service providers to provide highly available service, the use of Cloud computing technology is a wellsuited solution. Cloud computing allows resources to be provided as general utilities, which users can lease and release in an ondemand fashion through the Internet. In this thesis an HTTP based mobile messaging solution is designed, implemented and described and deployed in the Amazon Cloud, with the main goal of determining to what extent the availability of mobile messaging services improves when deployed in the Amazon Cloud. Three testing environments are proposed for deploying mobile messaging services, i.e. using a proprietary server, Auto Scaling service and the Elastic Load Balancing service provided by Amazon. The architectures for all three scenarios are described and illustrated, and a description is provided of their implementation. Based on performance tests executed in all three scenarios the improvement in availability is determined. Also stress tests are executed in all scenarios with the purpose to compare the performance, i.e. the average response time required to process a subscriber’s request for service, of each scenario with each other. A comparison of the test results provides insight into the availability of the tested messaging service, and the relationship that exists between each scenario’s performance.

[PDF]
[Abstract]
