MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Small and variable‐length encryption GUIDANCE {atul.luykx,bart.mennink}@esat.kuleuven.be For more information, please contact: Bart Preneel, Vincent Rijmen Promotor: Atul Luykx, Bart Mennink Supervisor: CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part) Block ciphers form the most basic building blocks in cryptography. A block cipher is defined as a function E : {0,1}k × {0, 1}n → {0, 1}n , where {0, 1}k denotes the key space and {0, 1}n the plaintext and ciphertext space. For any secret key K, E(K, ·) maps a plaintext m ∈ {0, 1}n to a unique ciphertext c ∈ {0, 1}n. However, block ciphers are inherently just fixed objects, and should be used in a certain mode to offer flexibility. One example is cipher block chaining, where an encryption scheme on arbitrary input length messages is constructed. The focus in this master thesis project is on a problem of a different kind: length doubling. Given above block cipher E, consider the problem of constructing a block cipher E* with plaintext and ciphertext space {0, 1}[n..2n-1], hence all strings of length between n and 2n − 1. Length doubling appears to be the right way to handle incomplete blocks in (authenticated) encryption, but it is not at all trivial to construct efficient and secure length doublers. The first length doubler that appeared in literature was XLS by Ristenpart and Rogaway (2007), but a recent attack by Nandi (2014) showed that XLS is completely insecure. Another known construction of a length doubler is HEM by Zhang (2012).
GOAL XLS found a broad range of applications, and the attack of Nandi has demonstrated the relevance of a length doubler that is not only efficient but also secure. The goal of this master thesis project is to analyze the existing schemes to get a solid understanding of them, and to improve upon these solutions, both from an efficiency as from a security point of view.
RESULT The thesis consists of a broad literature study to get familiar with the state of the art on encryption and length doubling, existing problems, and their theoretical background. Next, the student will invent new designs that improve upon existing solutions.
PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
50% literature, 50% theoretical work 1 of 2 studenten
1
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Mystery of bits and bytes GUIDANCE For more information, please contact: Promotor:
[email protected] Vincent Rijmen
Begul Bilgin Supervisor: CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part) Symmetric-key cryptography is a great asset to provide confidentiality, integrity and data authenticity. It gathers block and stream cipher, authenticated encryption and hashing algorithms under the umbrella of single or no secret keyed schemes. These algorithms typically take field elements of characteristic two, hence bits and bytes, and transform them using a set of permutations and vectorial Boolean functions. The strength of these algorithms against cryptanalysis relies on the combination of these mathematical operations. Cryptanalysis of symmetric-key algorithms is not the only method to breach their security. A bad implementation can reveal information about the secret to an attacker performing so called side-channel analysis. This type of analysis use information such as the power consumption or timing of the device. A common countermeasure to such attacks is masking and uses ideas from secret sharing and multi-party computation. The underlying Boolean functions of the cryptographic algorithms have great impact on the success and efficiency of such countermeasures, sometimes for an unknown reason.
GOAL The goal of this thesis project is to examine the vectorial Boolean functions used in symmetric-key algorithms in combination with the aforementioned countermeasures. There will not be one single question provided to the student. Instead, the student will be free to choose the exact question(s) of interest from a list of open problems following a field study and discussion. Interest in discrete algebra, number theory, combinatorics alike are advantageous.
RESULT Depends on the student’s interest. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
40% literature, 30% theoretical work, 30% software 1 of 2 studenten 1 or 2
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Anonymous transaction propagation of cryptocurrencies GUIDANCE {ren.zhang,Michael.herrmann}@esat.kuleuven.be For more information, please contact: Bart Preneel Promotor: Ren Zhang, Michael Herrmann Supervisor: CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part) A cryptocurrency is no more than a public ledger maintained by a peer-to-peer (P2P) network. Yet there are more than seven hundred of them operating nowadays and the market capitalization of Bitcoin alone is more than four billion dollars. They have value because people believe so. However, such belief is shaken drastically every time an attack happens. An attack that cannot be mitigated will affect the future of cryptocurrencies. From the aspect of privacy there are two deanonymization attacks. First, although a user is able to do her different transactions with different pseudonyms, an attacker may be able to correlate all her transactions in the public ledger and defeat the protection of the pseudonyms by graph and transaction flow analysis. Second, by collecting the traffic and connectivity information from the P2P network, an attacker can link a user’s transactions to her IP address and thus effectively identifying the respective real-world person. Many new cryptocurrencies and services are proposed from the industry for mitigating both attacks, however academia has mainly focused on defending against the first attack so far.
GOAL The first part of this thesis is to understand the designs of existing self-claimed privacy-preserving cryptocurrencies with regard to their resistance to P2P deanonymization attack. The student will review several most prominent ones by reading their technical documents and running experiments with a modified version of the original source code. In the second part, the designs are evaluated to identify the privacy risks for users. Finally, the student can propose a new network architecture to fix the vulnerabilities of an existing system. The student’s design should be verified with a suitable mechanism, such as simulations and/or a proof-of-concept implementation.
RESULT PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
30% literature, 20% theoretical work, 50% software 1 of 2 studenten 1 or 2
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Node assignment functionality for anonymous communications via multi‐party computation GUIDANCE {fateme.shirazi,elena.andreeva}@esat.kuleuven.be For more information, please contact: Claudia Diaz, Bart Preneel Promotor: Fateme Shirazi, Elena Andreeva Supervisor: CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part) Anonymous communication (AC) networks (e.g., Tor) enable users to communicate and access information anonymously over the Internet. ACNs are often realized by a set of nodes that relay the communication in order to disguise the connection between the sender and receiver of a communication. Some AC networks such as cMix [1] need to change the layout of the nodes frequently. A secure way to make such changes without the need to trust a third party is involving all nodes using Multi-Party Computation (MPC) protocols.
GOAL The aim of this work is to implement a node assignment functionality for AC networks using MPC.
RESULT PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera)) 30% literature, 10% theoretical work, 60% programming
The student should have good programming skills in either Python or C and a good mathematical understanding.
1 of 2 studenten
1
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Learning with errors GUIDANCE For more information, please contact: {Wouter.castryck,ilia.iliashenko}@esat.kuleuven.be Bart Preneel Promotor: Wouter Castryck, Ilia Iliashenko Supervisor: CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part) Public-key cryptography builds on notoriously hard mathematical problems, which (ideally spoken) an attacker should solve in order to crack the corresponding code. The main current examples of such problems are integer factorization (as in RSA), discrete logarithm computation in Fq \ {0}, · (with Fq a finite field), discrete logarithm computation in E(Fq),+ (with E an elliptic curve over a finite field Fq). About 10 years ago a new hard problem was proposed by Regev: learning with errors (LWE). It came along with a proof that breaking LWE would amount to breaking certain shortest vector problems (although not the NP-hard ones), thereby yielding a promising hardness statement. Moreover LWE has turned out to be extremely versatile for use in cryptography, enabling applications that were impossible before, such as homomorphic encryption. Loosely spoken, the problem is to solve an approximate linear system of equations As=b over a given finite prime field Fp for the unknown secret s. Here approximate means that each component bi carries a very small error term, say contained in {−r,...,r} ⊂ Fp with r the residue mod p of a very small integer. The system is understood to be slightly overdetermined, i.e. m > n. In fact, in Regev’s model the attacker is allowed to ask for new equations (i.e. let m grow) indefinitely in the hope of gradually unveiling the secret: this is where the terminology learning with errors comes from. All standard linear algebra methods seem useless in attacking this problem: for instance when carrying out Gaussian elimination the errors increase to a size where they look like random elements of Fp, and the system of equations becomes indistinguishable from an inconsistent one.
GOAL This project would mainly be a study of the existing literature on LWE and its applications, with a particular focus on the recently proposed Ring-LWE problem [LPR], which is a ring-based variant allowing the user to generate n equations at once, thereby turning the resulting systems much more efficient, at the expense of having a weaker hardness statement. This involves notions from algebraic number theory. Depending on the student’s preferences, the thesis could also have a programming component, implementing aspects of the foregoing in Magma.
RESULT PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
80% literature and theory, 20% implementation 1 of 2 studenten 1
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Efficient arithmetic on elliptic curves GUIDANCE For more information, please contact: Promotor: Supervisor:
[email protected] Bart Preneel Wouter Castryck, Carl Bootland, Simon Friedberger
CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part)
An increasingly important part of cryptography builds on the hardness of the elliptic curve discrete logarithm problem. For this one considers a large finite field Fq along with an equation of the form y2 = x3 + Ax + B, where A, B ∈ Fq are field elements. When equipped with the point at infinity ∞ := (0 : 1 : 0) its set of solutions E(Fq):={(x0,y0)|y02=x03+Ax0+B}∪{∞} can be turned into an abelian group E(Fq),+ using so-called tangent-and-chord arithmetic. The challenge for an attacker is then: given P and a multiple Q=nP reveal n. This is well-stated modulo the order of P, and for well-chosen parameters it is believed to be extremely hard. But the person setting up the system should of course be able to compute nP in the first place, preferably in a very efficient way. A first approach would be to use repeated doubling and adding, based on the binary expansion of n (similar to the modular exponentiation used in e.g. RSA), each time explicitly performing tangent-and-chord arithmetic. This works reasonably well, but further speed-ups are desirable in order to make the system competitive. Over the past years, many ideas have been launched, both to • improve upon the repeated doubling and adding method, in order to process the digits of n in a faster way: using non-adjacent forms, curve endomorphisms, . . . • improve upon the tangent-and-chord arithmetic: switching to projective coordinates, using alternative curve models, . . .
GOAL This project would be about studying these methods, several of which involve beautiful mathemat- ics. Depending on the student’s preferences, the focus could be on the underlying theory of elliptic curves, or on the more practical side.
RESULT PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
50% literature and theoretical work, 50/% implementation 1 of 2 studenten 1
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Fast matrix multiplications GUIDANCE Lieven De Lathauwer For more information, please contact: Lieven De Lathauwer Promotor: Marc Van Barel, P.‐A. Absil (UCL) Copromotor: Ignat Domanov Supervisor: CONTEXT (e.g. Information about problem definition, research project of which the master’s thesis is part) The straightforward way to multiply two (2×2) matrices requires 8 scalar multiplications. The arithmetic operations can however be grouped to reduce the work to 7 multiplications. By working recursively, we can reduce the multiplication cost for (N×N) matrices from the straightforward O(N3) to O(N2.8074) operations. More substantial computational savings may be obtained by starting from the reduction that can be achieved for (3×3) or (4×4) matrices, for instance. The reduction of the complexity may actually become so significant that a new architecture for large matrix multiplication emerges. Essential is first that we find inexpensive schemes for the multiplication of relatively small matrices. The latter problem amounts to finding decompositions of associated "multiplication tensors". GOAL The multiplication of two matrices A and B can be viewed as the bilinear map (A,B) → A*B. By duality, we can also view the multiplication as the trilinear form M: (A,B,C) → trace(A*B*C). The trilinear form is naturally associated with a third‐order tensor. The decomposition of this tensor in rank‐1 terms is the key to a reduction of the complexity of matrix multiplication. The purpose of this project is to investigate decompositions of the multiplication tensor in order to speed up matrix multiplication. A suggested approach relies on exploiting the invariance trace(A*B*C) = trace((X *A* Y‐1)(Y *B* Z‐1)(Z *C* X‐1). If the student does not progress sufficiently well in this topic, an alternative topic would be to investigate tensor block term decompositions from a Grassmann manifold perspective, as in Ishteva's http://dx.doi.org/10.1137/090764827. RESULT The required know‐how w.r.t. tensor methods is acquired through discussions, attending a few lectures and a limited, guided study of the literature. A concise course text is available. For optimization on manifolds, we use the book P.‐A. Absil, R. Mahony, R. Sepulchre, Optimization Algorithms on Matrix Manifolds, Princeton Univ. Press, 2008. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
An interest in numerical optimization is required. Algorithms are implemented in Matlab. 1 or 2 students are allowed
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Sets of polynomial equations, decompositions of higher‐order tensors and multidimensional harmonic retrieval: connections and algorithms GUIDANCE Lieven De Lathauwer For more information, please contact: Lieven De Lathauwer Promotor: Bart De Moor Copromotor: Mikael Sorensen, Frederik Van Eeghem Supervisor: CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part)
Sets of polynomial equations can in principle be solved by constructing the so‐called Macaulay matrix of equation coefficients and by exploiting shift‐invariance structure in its null space via an eigenvalue decomposition. Canonical Polyadic Decomposition (CPD) of a higher‐order tensor is its decomposition in a minimal number of rank‐1 terms. Higher‐order tensors are “matrices in more than two dimensions". Tensors are becoming increasingly important in signal processing, data mining, machine learning, scientific computing and many other fields, where they open up remarkable new opportunities. Harmonic Retrieval (HR) is the estimation of exponential (sinusoidal) components of an observed signal. Multidimensional HR involves components that behave exponentially along more than one axis. These are fundamental problems with countless applications, e.g. in array processing, telecommunication, biomedical data analysis, … GOAL The size of the Macaulay matrix limits its use to small‐scale problems. However, the Macaulay matrix is very sparse and has a block‐Toeplitz structure, such that we can envisage handling much larger sets of equations if we compute the null space by means of iterative methods. Moreover, the null space has the structure of a higher‐order tensor of which the CPD reveals the unknown roots. As a matter of fact, the tensor has the same structure as in a multidimensional HR problem. For multidimensional HR, we have recently derived the most relaxed uniqueness conditions to date, and we have developed eigenvalue decomposition based solvers that exploit the multidimensional exponential structure in an efficient manner. The goal of the thesis is to combine these points of view 1) to obtain insight in the properties and the uniqueness of the solution of sets of polynomial equations and 2) to obtain algorithms for larger‐scale sets of equations that reduce the problem to a matrix eigenvalue decomposition. RESULT First the required know‐how is acquired through discussions, attending a few lectures on tensor methods and a limited, guided study of the literature. For tensor methods a concise course text is available. Then algebraic connections are explored and new algorithms are developed with attention for numerical aspects. The algorithms are validated on practical examples. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
Depending on the interest, the emphasis can be put less or more on algebraic foundations, numerical aspects or implementation. The algorithms are implemented in Matlab. 1 or 2 students are allowed
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Tensor decomposition updating and tracking GUIDANCE Lieven De Lathauwer For more information, please contact: Lieven De Lathauwer Promotor: Raf Vandebril Copromotor: Nico Vervliet, Martijn Boussé Supervisor: CONTEXT (e.g. Information about problem definition, research project of which the master’s thesis is part) An important research trend is the transition from vector and matrix based mathematical engineering to generalizations that make use of higher‐order tensors. Higher‐order tensors are “matrices in more than two dimensions". This trend manifests itself across many disciplines, such as signal processing, data mining, machine learning, scientific computing, … Tensor methods indeed offer remarkable conceptual possibilities beyond matrix techniques. Tensorlab is an in‐ house developed Matlab toolbox that comprises numerically reliable and efficient algorithms for the computation of tensor decompositions (www.tensorlab.net). GOAL Tensor applications typically involve large data sets. As a matter of fact, tensor decompositions are promising new tools for the analysis of big data, and big data are nowadays everywhere. One strategy to obtain decompositions of large data matrices or tensors, is to compute them recursively, i.e. new entries are taken into account by updating the decomposition. Combined with downdating, this moreover allows us to track decompositions of nonstationary data. Tensorlab does not yet provide updating/tracking algorithms. The development of such methods is the goal of this thesis. They have high application potential, also in multimodal data fusion (think for instance of the simultaneous processing of EEG, ECG, MRI, … as new samples arrive). RESULT The required know‐how w.r.t. tensor methods is acquired through discussions, attending a few lectures and a limited, guided study of the literature. A concise course text is available. New algorithms are developed with attention for numerical aspects. The algorithms are validated on synthetic examples and real‐life data. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
Depending on the interest, the emphasis can be put less or more on numerical aspects, implementation or applications. The algorithms are implemented in Matlab. 1 or 2 students are allowed
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Generalized cost functions in tensor optimization GUIDANCE Lieven De Lathauwer For more information, please contact: Lieven De Lathauwer Promotor: Stefan Vandewalle Copromotor: Martijn Boussé Supervisor: CONTEXT (e.g. Information about problem definition, research project of which the master’s thesis is part) An important research trend is the transition from vector and matrix based mathematical engineering to generalizations that make use of higher‐order tensors. Higher‐order tensors are “matrices in more than two dimensions". This trend manifests itself across many disciplines, such as signal processing, data mining, machine learning, scientific computing, … Tensorlab is an in‐house developed Matlab toolbox that comprises numerically reliable and efficient algorithms for the computation of tensor decompositions (www.tensorlab.net). The high performance is the result of a careful handling of multilinear (rank‐1) structure that is inherent in tensor decompositions. GOAL Tensorlab assumes that the tensors that need to be decomposed, are given. However, the tensor itself could only be implicitly available, for instance as the solution of a linear system. This leads to solving sets of linear equations under low‐rank (tensor) constraints on the solution. Such problems are fundamental and open up significant new applications. For instance, recognition tasks (image, video) are naturally formulated in this way. There are also intimate links with (large‐scale) system identification, compressed sensing, array processing and wireless communication, to give just a few examples. The aim of the thesis is the extension of the algorithms in Tensorlab to cost functions that encompass such constrained set of linear equations. A second possible aim is the extension from least‐squares optimization to the minimization of non‐quadratic cost functions that are a better match for pixel intensities, audio, spectra, concentrations, … RESULT The required know‐how w.r.t. tensor methods is acquired through discussions, attending a few lectures and a limited, guided study of the literature. A concise course text is available. New algorithms are developed with attention for numerical aspects. The algorithms are validated on synthetic examples and real‐life data. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
Depending on the interest, the emphasis can be put less or more on numerical aspects, implementation, development of a software tool or applications. The algorithms are implemented in Matlab. 1 or 2 students are allowed
MASTER WISKUNDIGE INGENIEURSTECHNIEKEN TITEL Inferentie van hiërarchische structuren in menselijke anatomie via temporele informatie BEGELEIDING
[email protected] Voor meer informatie, contacteer: Prof. Dr. Bart De Moor Begeleider: Ms. Supinya Piampongsant Dagelijkse begeleider(s): CONTEXT Studies hebben aangetoond dat er een correlatie bestaat tussen structurele, functionele en temporele hiërarchieën in de menselijke hersenen, waarbij de laagste niveaus verantwoordelijk zijn voor snelle signaalverwerking, terwijl de hoogste niveaus trage contextuele veranderingen verwerken. In deze thesis wordt het verband tussen andere delen van de menselijke anatomie bestudeerd. We zoeken naar een automatische inferentie van de temporele hiërarchie in het lichaam op basis van tijdsdata. Het resultaat moet vergeleken worden met de structurele en functionele structuur van het lichaam. Dit zal gedaan worden met behulp van de Anatomage tafel, een biomedisch leermiddel dat een volledig correcte observatie en virtuele dissectie van een menselijk lichaam mogelijk maakt (zie figuren). Deze virtuele dissectietafel wordt ter beschikking gesteld zodat de student inzicht kan verwerven over het menselijk lichaam vanuit het perspectief van anatomische experts.
Meer context kan teruggevonden worden in de volgende publicaties: Aburatani, S., Goto, K., Saito, S., Toh, H., & Horimoto, K. (2005). ASIAN: A web server for inferring a regulatory network framework from gene expression profiles. Nucleic Acids Research, 33(Web Server). doi:10.1093/nar/gki446 Anatomage ‐ Medical Home. (n.d.). Retrieved April 14, 2016, from http://medical.anatomage.com/ Friston, K. (2008). Hierarchical Models in the Brain. PLoS Comput Biol PLoS Computational Biology, 4(11). doi:10.1371/journal.pcbi.1000211 Kiebel, S. J., Daunizeau, J., & Friston, K. J. (2008). A Hierarchy of Time‐Scales and the Brain. PLoS Comput Biol PLoS Computational Biology, 4(11). doi:10.1371/journal.pcbi.1000209 DOEL Het doel van deze thesis is om een hiërarchische structuur van het menselijk lichaam te infereren via tijdsdata. De Anatomage tafel zal gebruikt worden als referentie voor de structuur van het menselijk lichaam zoals beschreven door anatomische experts. De student zal data krijgen over het dynamisch gedrag van ieder lichaamsdeel, inclusief de tijdschaal waarin deze werken. Waar toepasselijk, zal tijdschaal‐informatie geïnfereerd
worden via dynamische modellen uit de BioModels database. Een tijdsgebaseerde hiërarchische structuur van het menselijk lichaam wordt dan afgeleid en vergeleken met bestaande standaarden (zie voorgaande referenties voor meer informatie over de methodologie). De student dient dan de resultaten voor te stellen op een informatieve en visueel aantrekkelijke manier. Deze thesis kan uitgebreid worden om ook andere eigenschappen in de hiërarchische structuur in te passen, zoals energieverbruik, spatiale informatie en biologische complexiteit. WERKVERDELING Literatuur en analyse (50%) / programmeren en simuleren (50%) PROFIEL - Goede kennis van biologische systemen, systeemtheorie en regeltechniek. - Noties van optimalisatie, parameterschatting en MATLAB/Simulink. - Sterke interesse in het onderwerp. - Bereidwillig om initiatief te nemen. De student krijgt in dit project veel vrijheid om het project zelf vorm te geven. Aantal studenten
1
MASTERPROEF MASTER WISKUNDIGE INGENIEURSTECHNIEKEN TITEL Multi‐dimensionele systeemtheorie: ontwikkelen van een MATLAB toolbox BEGELEIDING
[email protected] Voor meer informatie contacteer: Prof. Dr. Bart De Moor Promotor: Robin Van Craen Begeleider(s): CONTEXT (Info over het probleem, ev. info over het onderzoeksproject waarin deze masterproef kadert) Dynamische systemen zijn alom vertegenwoordigd in ons dagelijkse leven. Daarom is het modelleren en analyseren van deze systemen van cruciaal belang in onze samenleving. In de laatste decennia is er enorm veel aandacht en onderzoek geweest omtrent één‐dimensionele lineaire tijdsinvariante systemen (LTI), welke variëren in één variabele (meestal tijd). Echter zijn er in de realiteit dikwijls ook andere onafhankelijke variabelen die een belangrijke rol spelen in het systeem, zoals ruimtelijke coördinaten (denk bijvoorbeeld maar aan de temperatuursverdeling in een kamer enz.). De dynamica van zo’n soort systemen wordt doorgaans beschreven via partiële differentiaal/differentie vergelijkingen en een manier om deze te modelleren is met behulp van multi‐dimensionele shift‐invariante systemen (nD systemen). De analyse en identificatie van deze systemen krijgen de laatste jaren veel aandacht (met een initialisatie zelfs al eind jaren 70’). Men zou kunnen hopen dat alle gekende technieken voor 1D systemen gemakkelijk generaliseren naar hogere dimensies, maar dit is echter niet het geval. Er zijn intrinsieke verschillen tussen de twee (zo zal bijvoorbeeld een toestand in een nD systeem niet meer ‘alle nodige informatie bezitten’) en wiskundige technieken zullen niet altijd even eenvoudig generaliseren (bv. de fundamentele stelling van de algebra is niet meer geldig in nD). Dit leidt er bv. ook toe dat er verschillende toestandsbeschrijvingen mogelijk zijn (waar dit voor 1D LTI systemen natuurlijk niet zo is). Het is dan ook niet onlogisch dat voor nD systemen zaken die ‘gekend’ zijn in 1D nog geen tegenhanger hebben (of slechts gedeeltelijk), en bijgevolg nog niet extensief gebruikt worden in industriële toepassingen, waar dit over een aantal jaar waarschijnlijk wel zo zal zijn. Een voorbeeld hiervan is systeemidentificatie voor nD systemen, waar onze focus ook gedeeltelijk naar uit zal gaan. Op dit moment bestaat er een MATLAB toolbox die het gebruik van nD systemen toelaat, maar deze is niet gericht op toestandsbeschrijvingen, laat staan dat gekende identificatie methoden geïmplementeerd zijn. Voor later gebruik zou zo een toolbox zeer handig zijn, zowel voor verder onderzoek naar nD systemen (wat nog volop aan de gang is), maar ook voor allerhande toepassingen in verscheidene domeinen zoals de klimatologie (de temperatuur in de atmosfeer kan gezien worden als een nD systeem), massaspectrometrie (bv bij het bestuderen van de hersenen van muizen), signaalanalyse en beeldverwerking (verbetering van röntgenfoto’s voor medische toepassingen of verbetering en analyse van luchtfoto’s voor bv detectie van oogstschade of bosbranden) en vele andere industriële toepassingen. DOEL Het doel van deze thesis is het ontwikkelen van een MATLAB toolbox voor analyse van nD systemen. De focus hiervoor zal meer liggen op de verschillende toestandsbeschrijvingen en gekende technieken betreffende systeemidentificatie voor nD systemen.
De toolbox zou zeker de basisfunctionaliteiten voor het werken met nD shift invariante systemen moeten bevatten. Dit houdt o.a. volgende dingen in: nD systemen creëren via transfer functie, toestandsbeschrijving enz. Het converteren van verschillende toestandsbeschrijvingen (indien dit mogelijk is). Het berekenen van de impulsresponsie en meer algemeen de responsie op een willekeurige door de gebruiker gegeven input. Het berekenen van de polen en nulpunten van een gegeven transfer functie. Implementatie van enkele gekende identificatie algoritmes. … De toolbox zou zo gemaakt moeten worden dat toekomstige uitbreidingen eenvoudig zijn. Aangezien nD systemen nog volop aan onderzoek onderworpen zijn, is het niet ondenkbaar dat in de recente toekomst een aantal nieuwe technieken en algoritmes hun intrede zullen maken. Zo is het wenselijk dat in de toekomst deze technieken en algoritmes toegevoegd kunnen worden. Op het einde van de thesis zou de student een functionerende MATLAB toolbox met de basisfunctionaliteiten moeten afgeleverd hebben, die klaar is voor gebruik. UITWERKING Een grondige literatuurstudie omtrent nD systemen plus implementatie van de relevante dingen in MATLAB. PROFIEL (bv. eerder theoretisch/eerder implementatie, vereiste voorkennis (bv. gevolgde of te volgen vakken, methodieken, programmeertalen, ... ))
Uitgebreide kennis van MATLAB, dynamische lineaire 1D systemen, systeemidentificatie en een basiskennis en gezonde interesse in lineaire algebra. 1 of 2 Studenten 1
MASTERPROEF MASTER WISKUNDIGE INGENIEURSTECHNIEKEN TITEL Online clustering van tijdreeksen: een snel, innovatief en accuraat algoritme BEGELEIDING
[email protected] Voor meer informatie contacteer: Prof. Dr. Ir. Bart De Moor Promotor Oliver Lauwers Begeleider(s): CONTEXT (Info over het probleem, ev. info over het onderzoeksproject waarin deze masterproef kadert) De analyse van gegevens doorheen de tijd is een belangrijk onderdeel van vele wetenschappelijke, industriële en commerciële industrieën. Zowel in de financiële wereld, de procesindustrie als in biomedische toepassingen bieden de temporele correlaties doorheen gegevens een schat aan extra informatie, die niet uit de statische data zelf te halen valt. Een belangrijke uitdaging daarbij is het clusteren van tijdreeksen. Hierbij is een degelijke notie van afstand of gelijkaardigheid tussen tijdreeksen onontbeerlijk. In dit project kijken we naar een algoritme, gebaseerd op een innovatieve afstandsmaat die gegarandeerd elke afhankelijkheid tussen tijdreeksen in rekening brengt, voor bepaalde modelklasses. Een proof‐of‐conceptalgoritme werd reeds uitgewerkt en toegepast op beursgegevens, verkoopsgegevens van AB Inbev en data uit de procesindustrie. DOEL In dit thesisproject trachten we een innovatief algoritme voor het clusteren van tijdreeksen te verbeteren, zowel op conceptueel als op computationeel niveau. De proof‐of‐concept werkt momenteel voor lineaire tijdreeksen gegenereerd door (of modelleerbaar als) stochastische SISO ARMA‐modellen. De doelstelling is dit algoritme uit te breiden naar algemenere modelklassen, zoals modellen met deterministische input, of modellen met meerdere in‐ en uitgangen, alsook exotischere zaken als niet‐lineaire systemen. Verder is ook het vastleggen van parameters een niet‐triviaal conceptueel probleem. Anderzijds is het ook van belang op computationele wijze de efficiëntie en werking van het algoritme te verbeteren. Zeker voor grote datasets, zoals die typisch voorkomen in reële toepassingen, is het cruciaal dat de methode snel en paralleliseerbaar is. Het uiteindelijke doel is deze methodes dan uit te testen en toe te passen op een reëel probleem, met gegevens van AB Inbev, bedrijven uit de procesindustrie in samenwerking met spin‐offbedrijf Dsquare, en elektriciteitsgegevens in samenwerking met Eandis. UITWERKING Het algoritme verwacht een lijst tijdreeksen als input. Vervolgens wordt een stochasisch model opgesteld van de eerste tijdreeks, waarna alle andere tijdreeksen door dit inverse model gehaald worden. Wanneer de zo verkregen uitgang voldoende op witte ruis lijkt, voegen we de reeks toe aan de cluster rond de eerste tijdreeks. We nemen dan de lijst van ongeclusterde reeksen, en herhalen het algoritme. Op theoretisch vlak werd, voor bepaalde modelklassen, een hele systeem‐ en informatietheoretische achtergrond uitgewerkt. We trachten dit raamwerk uit te breiden. PROFIEL (bv. eerder theoretisch/eerder implementatie, vereiste voorkennis (bv. gevolgde of te volgen vakken, methodieken,programmeertalen, ... ))
De thesisstudent(e) vindt de intellectuele uitdaging van een sterke theoretische component een meerwaarde, en wil zich verdiepen in een wiskundig uitgewerkt raamwerk rond een afstandsmaat voor tijdreeksen. Daarnaast brengt de student(e) deze theoretische grondslag graag in de praktijk door middel van een uitbreiding van de implementatie van een bestaand algoritme. De proof‐of‐ concept is uitgewerkt in Matlab, maar kennis van andere talen, zoals Python of R, is ook voldoende. Kennis van talen die makkelijk parallelisatie toelaten, zoals Hadoop of Spark, zijn een pluspunt. Interesse in concrete toepassingsgebieden, zoals de financiële wereld, de elektriciteitssector of de procesindustrie wordt ook aangemoedigd. 1 of 2 studenten 1 student
MASTERPROEF MASTER WISKUNDIGE INGENIEURSTECHNIEKEN TITEL Ontwikkeling van een Javascript bibliotheek voor analyse en ontwerp van regelsystemen met terugkoppeling BEGELEIDING Voor meer informatie contacteer: Promotor: Begeleider(s):
[email protected] Prof. Dr. Bart De Moor Dr. Oscar Mauricio Agudelo Msc. Supinya Piampongsant
CONTEXT (Info over het probleem, ev. info over het onderzoeksproject waarin deze masterproef kadert) Massive Open Online Courses (MOOCs) zijn de recentste vooruitgang in het onderwijs. Het zijn opleidingen die van op een afstand gevolgd kunnen worden. Een MOOC levert online studiemateriaal aan elke persoon die het betreffende vak wilt volgen, met geen limiet op het aantal personen. Deze MOOCs gebruiken meestal traditioneel studiemateriaal zoals video’s, cursussen, oefeningen, maar ook interactieve fora voor een vlottere communicatie tussen studenten, professoren en assistenten. Voor het specifieke geval van een MOOC voor systeemtheorie en regeltechniek is het cruciaal om web‐gebaseerde werktuigen te hebben die de analyse, ontwerp en simulatie van regelsystemen met terugkoppeling vergemakkelijken. Naast HTML en CSS is JavaScript een van de drie essentiële technologieën die hiervoor wereldwijd gebruikt wordt. Het is high‐level taal die gebruikt wordt in het overgrote deel van de websites en wordt ondersteund door alle moderne web browsers zonder plug‐ins. Hoewel er duizenden JavaScript bibliotheken zijn die de ontwikkeling van web‐gebaseerde applicaties gemakkelijker maken, is er toch geen enkele bibliotheek voor het implementeren van eenvoudige operaties voor analyse en ontwerp van regelsystemen met terugkoppeling. Daarom is het ontwikkelen van demo’s, en meer algemeen het ontwikkelen van web‐gebaseerde applicaties omtrent systeemtheorie en regeltechniek zeer tijdrovend en lastig. DOEL Het doel van dit project is het ontwikkelen van een JavaScript bibliotheek voor analyse, design en simulatie van regelsystemen met terugkoppeling. Math.js Numeric.js
JavaScript bibliotheek voor systeemtheorie en regeltechniek
HighCharts.js
Third‐party bibliotheken
Web‐gebaseerde applicaties voor systeemtheorie en regeltechniek
De bibliotheek zou de basisfunctionaliteiten voor het werken met SISO (Single‐Input Single‐Output) LTI (lineair tijdsinvariante) systemen moeten bevatten, zowel voor het continue als voor het discrete tijdsdomein. De gebruiker zou ook o.a. volgende dingen moeten kunnen doen: modellen creëren via transfer functies, nulpunten‐polen‐versterking en toestandsbeschrijving, in zowel continue als discrete tijd. Lineaire modellen op verschillende manieren verbinden (serie, parallel, terugkoppeling, …).
Het pool‐nulpunt diagram genereren en het berekenen van de polen en nulpunten van een gegeven transfer functie. Het berekenen en visualiseren van de stapresponsie, de impulsresponsie, de beginvoorwaarden (van de toestandsbeschrijving), en de responsie op een input die gegeven wordt door de gebruiker zelf. Het berekenen en visualiseren van de regimeoplossing van het systeem via het Bode diagram en de Nyquist kromme. Het bepalen van de faseverschuiving en versterking. De wortellijnen van een system genereren. De transfer functie en toestandsbeschrijving weergeven, gebruik makend van mathjax Het converteren van continue tijdsystemen naar equivalente discrete tijdsystemen en omgekeerd, gebruik makend van verschillende gekende methoden (zero‐order‐hold, Tustin regel, achterwaartse differentiatie regel , enz.) Het creëren PID controllers …
De functies zouden zo veel mogelijk in lijn moeten zijn met de beschikbare dingen in de MATLAB control system toolbox. De bibliotheek zou zo gemaakt moeten worden dat toekomstige uitbreidingen eenvoudig zijn. Zo is het bijvoorbeeld wenselijk dat voor MIMO systemen en systemen gegeven door de toestandsbeschrijving, er in de toekomst technieken voor ontwerp en regeling kunnen toegevoegd worden. De student mag elke vrije (voor niet‐commerciële doeleinden) third‐party bibliotheek gebruiken die hij/zij nuttig acht zoals bv. Math.js (een uitgebreide wiskunde bibliotheek voor JavaScript), Numeric.js (voor doorgedreven wiskundige berekeningen), HighCharts.js (voor grafieken en visualisatie), MathJax enz. Op het einde van het project zou de student een website (gehost door de STADIUS/SMC servers) gecreëerd moeten hebben waar zowel de bibliotheek, de handleiding en ander ondersteunend materiaal zoals demo’s beschikbaar zouden moeten zijn. UITWERKING Literatuurstudie: 30% / Programmeren: 70% PROFIEL (bv. eerder theoretisch/eerder implementatie, vereiste voorkennis (bv. gevolgde of te volgen vakken, methodieken, programmeertalen, ... ))
Uitgebreide kennis in de ontwikkeling van web applicaties (Javascript, HTML, enz.), regeltechniek en dynamische systemen, MATLAB, en een basis kennis in numerieke analyse en lineaire algebra. 1 of 2 Studenten 1
MASTERPROEF MASTER WISKUNDIGE INGENIEURSTECHNIEKEN TITEL Tijdreeksen en complexe netwerktheorie: een duaal problem? BEGELEIDING
[email protected] Voor meer informatie contacteer: Prof. Dr. Ir. Bart De Moor Promotor Oliver Lauwers Begeleider(s): CONTEXT (Info over het probleem, ev. info over het onderzoeksproject waarin deze masterproef kadert) De analyse van gegevens doorheen de tijd is een belangrijk onderdeel van vele wetenschappelijke, industriële en commerciële industrieën. Zowel in de financiële wereld, de procesindustrie als in biomedische toepassingen bieden de temporele correlaties doorheen gegevens een schat aan extra informatie, die niet uit de statische data zelf te halen valt. Vaak bevatten deze gegevens niet‐ lineariteiten, die notoir moeilijk te verwerken zijn. Ook het bestuderen van complexe netwerken is een interessante uitdaging in het verwerken van gegevens. Sociale netwerken, verkeersinformatie, maar ook data rond besmetting en epidemieën kunnen gemodelleerd worden in de vorm van een netwerk. Recent werden een aantal transformaties uitgewerkt die deze twee types data in elkaar omvormen. Op die manier ontstaat er een dualiteit tussen tijdreeksen en complexe netwerken. DOEL Het doel van deze thesis is deze dualiteit te onderzoeken. Verschillende transformaties veranderen dezelfde tijdreeks in verschillende netwerken. Belangrijk hierbij is te onderzoeken welke informatie behouden blijft, en op welke manier deze in het netwerk wordt opgeslagen. Bij sommige transformaties is dit analytisch uitgewerkt, bij andere in het geheel niet. We trachten deze lacune op te vullen. Op die manier wordt het arsenaal aan analysetechnieken voor zowel het bestuderen van complexe netwerken als het verwerken van temporele gegevens uitgebreid. UITWERKING Op basis van een transformatie waarvoor verschillende verbanden analytisch uitgewerkt zijn, trachten we deze overgangen te begrijpen. Verder hebben we ook numeriek gegenereerde informatie over de samenhang tussen bepaalde tijdreeks‐ en netwerkkarakteristieken. We combineren deze twee informatiebronnen om ook voor andere transformaties deze verbanden analytisch uit te werken. Tenslotte zoeken we experimentele bevestiging van deze resultaten. PROFIEL (bv. eerder theoretisch/eerder implementatie, vereiste voorkennis (bv. gevolgde of te volgen vakken, methodieken,programmeertalen, ... ))
De student(e) heeft een grote interesse in een theoretische, analytische benadering van een interessante, nieuwe ontwikkeling in data science. Hij/zij is geïnteresseerd in gegevensanalyse, en hoe verschillende representaties van dezelfde data samenhangen en de verwerking ervan kunnen vergemakkelijken. Een goede kennis van programmeertalen is niet strikt noodzakelijk, maar een basiskennis van Matlab, Python of R is wel vereist voor de experimentele bevestiging van de analytische resultaten. 1 of 2 studenten 1 of 2 studenten
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Fast function approximation with lattice sets on arbitrarily shaped domains in Julia GUIDANCE Daan Huybrechs For more information, please contact:
Promotor:
Daan Huybrechs, Ronald Cools
Roel Matthysen, Daan Huybrechs Supervisor: CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part)
There is a growing list of modern numerical software packages dedicated to function approximation, in combination with equation solving in a variety of problems. A prime example is the Chebfun package. Recent research focuses mostly on the difficulties of multivariate problems. Two such packages have been written at KULeuven: FrameFuns and ChebInt. Both packages have advanced the state of the art for multivariate problems. FrameFuns is based on classical Fourier series, but operates on domains that can have an arbitrary shape (including fractals). It uses a basis on a bounding box, which becomes a redundant set on the embedded domain ‐‐ this redundant set is called a frame. ChebInt has pioneered the use of so‐called lattice‐based Fourier series, and it requires drastically fewer degrees of freedom for higher‐dimensional problems. However, it is restricted to box‐shaped domains. In this thesis we will combine the strengths of both packages. FrameFuns is the more recent package. It is written in Julia, and available on GitHub. Julia is a modern open‐source language for technical computing, that originated at MIT and is quickly gaining widespread adoption at companies and universities. It has a simple Matlab‐like syntax, and it is a fully dynamic language like python, but it offers execution speed that is comparable to C code. Julia code is typically very short, yet amazingly fast. GOAL The goal of this thesis is to investigate the use of lattice point sets in frame approximations on arbitrary domains. The starting point is a Matlab implementation of lattice‐based approximations, and a Julia package for computing with frames on complicated domains. The implementation goal is to combine the two. The theoretical goal is to extend an existing fast approximation algorithm for Fourier based frames to lattice based Fourier series. This algorithm was only recently discovered and described by Roel Matthysen in the Nines research group in Leuven. RESULT The main result of the thesis is a fast implementation of lattice‐based Fourier series for the approximation of functions on arbitrary domains. The efficiency of the implementation is based on the use of Julia, as well as on the application of a novel fast algorithm for frame approximations. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
No prior knowledge of Julia or lattice sets is required. You have an interest in programming, but you don't want to be limited by the speed of Matlab and you don't appreciate the daunting error messages of C++ template programming. 1 of 2 studenten 1 or 2 students
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Implementation of asymptotic expansions for Hermite–type orthogonal polynomials GUIDANCE Daan Huybrechs, Peter Opsomer For more information, please contact:
Promotor:
Daan Huybrechs
Peter Opsomer Supervisor: CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part) Orthogonal polynomials play a crucial rule in many algorithms in numerical analysis and scientific computing. Computations of orthogonal polynomials are typically based on their three‐term recurrence relation. For that reason, even the simple problem of evaluating a polynomial of degree n requires an O(n) algorithm. The computation of Gaussian quadrature rules is O(n^2). Asymptotic expansions are game‐changing in this context: orthogonal polynomials of large degree can be approximated very well with an explicit formula. This makes polynomial evaluation an O(1) algorithm, and the construction of Gaussian quadrature rules O(n). A front page article in SIAM News (March 2015) by Alex Townsend (MIT) highlights the enormous impact of asymptotic expansions on the computation of Gauss‐Legendre quadrature rules. It is somehow surprising that substantial progress can still be made in such an elementary algorithm in the field of numerical analysis. Asymptotic expansions for Jacobi‐ and Laguerre‐type polynomials have been constructed in the department recently. However, analogous descriptions of expansions for generalized Hermite polynomials can also be useful for quadrature on the whole real line. Those polynomials are orthogonal on (‐∞, ∞) with respect to the weight function w(x) = e−Q(x) , where Q denotes a polynomial increasing at infinity, or possibly a general function. GOAL To construct and implement asymptotic expansions for generalized Hermite polynomials, based on asymptotic analysis in mathematical literature. The result may readily be incorporated in a software package for Gaussian quadrature rules developed by Alex Townsend. RESULT All asymptotic information is available in principle in mathematical literature [Deift, P., Kriecherbauer, T., McLaughlin K., Venakides, S. and Zhou, X. Strong asymptotics of orthogonal polynomials with respect to exponential weights, Commun. Pur. Appl. Math., LII, 1491‐1552 (1999)]. This is based on the so‐ called Riemann–Hilbert formulation and the Deift–Zhou nonlinear steepest descent method. Understanding this rather rigorous mathematical derivation can be made easier by comparing to the known Jacobi‐ and Laguerre‐type cases. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
Initially rather theoretical and mathematically challenging, gradually more implementation oriented. Some knowledge of symbolical and numerical mathematical software is recommended. Acquaintance with complex analysis such as branch cuts and contour integration is advised 1 of 2 studenten
1 student
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Nonstandard numerical methods for PDE's in wave propagation problems GUIDANCE Daan Huybrechs For more information, please contact:
Promotor:
Daan Huybrechs
Daan Huybrechs Supervisor: CONTEXT Numerical methods for problems with wave characteristics are plagued with many difficulties. First and foremost, the number of degrees of freedom needs to be high in order to represent an oscillatory solution. Second, classical finite‐element basis functions (e.g. piecewise polynomials) are not well suited to approximate oscillatory functions. Finally, condition numbers and other properties of the discretization typically deteriorate rapidly with increasing frequency of the problem at hand. For those reasons, several research groups around the world are studying nonstandard methods, in the sense that they discretize PDE's using oscillatory basis functions rather than piecewise polynomials. Example basis functions include plane waves, the Green's function of the underlying PDE, Hankel functions and other special functions. Invariably, these discretizations are extremely ill‐ conditioned, yet they provide accurate answers with a modest number of degrees of freedom. Ill‐conditioning of approximation problems arises, in ongoing research in Leuven, when a function is approximated using a redundant set rather than a basis. The redundancy means that the basis functions can be (close to) linearly dependent, hence the ill‐conditioning. However, it turns out that this ill‐conditioning is mostly harmless. For the approximation problem, this phenomenon is well understood by now and theoretical insight has also led to fast approximation algorithms. The question is whether this phenomenon also underlies nonstandard methods for wave problems, and whether the algorithms currently in use can be improved by taking this point of view. As a simple model problem, we will focus on the Helmholtz equation in acoustics. We will be using plane waves in a number of different directions as basis functions, and investigate the ill‐ conditioning of the discretization of the problem. The method is straightforward and rather popular. GOAL We want to assess, with numerical experiments and with reasoning based on available literature, whether the approximation of the solution to a wave problem using plane waves rather than more standard finite elements, is comparable to the approximation of a function from a redundant set. RESULT A lot of knowledge is available in the research group about the approximation of functions with redundant sets. The insights suggest implementation techniques, and also lead to fast algorithms. It is likely that numerical methods for wave problems may benefit substantially from this knowledge. A similar study of the so‐called Wave Based Method for problems in acoustics was performed last academic year (2015‐2016) and was very successful: the Wave Based Method could be significantly improved by adapting the techniques used for function approximation. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
Combination of theory and practice. Knowledge of PDEs and numerical methods is helpful. 1 of 2 studenten 1 or 2
MASTERPROEF MASTER WISKUNDIGE INGENIEURSTECHNIEKEN Titel van de masterproef
Quadratic eigenvalue problems from flowing plasma states Begeleiding: Rony Keppens, Karl Meerbergen • Voor meer informatie contacteer: Karl Meerbergen en Rony Keppens • Promotor idem • Begeleider(s): Context (Info over het probleem, ev. info over het onderzoeksproject waarin deze masterproef kadert) The theory to describe all linear waves and instabilities, given an exact stationary, gravitating, magnetized plasma equilibrium, in any dimensionality, has been known since 1960, and is the Frieman-Rotenberg equation. The equation governs the spatio-temporal variation of a Lagrangian fluid element displacement. When one is interested in eigenfrequencies, the temporal variation can be fixed to be exponential, and eigenfrequencies are found for all solutions obeying the Frieman Rotenberg equation and suitable boundary conditions. In the case of a static (no flow) equilibrium, the eigenvalue determination is governed by a self-adjoint operator in the Hilbert space of solutions. Due to the self-adjointness, only real eigenvalues can occur, so that stable waves or instabilities result. The full (mathematical) power of spectral theory governing physical eigenmode determination comes into play when using the Frieman-Rotenberg equation for moving plasma states. DOEL In that case, a generalized operator involving flow terms, together with a Doppler-Coriolis operator appear in the quadratic eigenvalue problem. Both operators are self-adjoint, but allow the possibility of intrinsically complex eigenvalues, with overstable modes, e.g., driven by shear flow (Kelvin-Helmholtz). Unlike the case of static plasmas, where the transition to instability in the complex eigenfrequency plane can only happen through the marginal zero frequency, the eigenmodes of stationary plasmas may enter the complex eigenfrequency plane from different locations along the real oscillation frequency-axis. The master thesis research will use existing codes exploiting finite-element discretizations to numerically translate the eigenvalue-eigenfunction determination, to be solved with modern numerical linear algebra techniques. The determination of eigenvalues-eigenfunctions for accretion disk configurations will then be confronted with recent theory, allowing to classify the eigenmodes in the complex eigenfrequency plane. UITWERKING Relevant papers and information can be found in the following literature. J.P. Goedbloed, R. Keppens, and S. Poedts, Advanced MHD, Cambridge University Press, Cambridge (2010). F. Tisseur and K. Meerbergen, SIAM review 43(2), 235-286 (2001) PROFIEL (bv. eerder theoretisch/eerder implementatie, vereiste voorkennis: (bv. gevolgde of te volgen vakken, methodieken,programmeertalen, ... ) Interest in numerical linear algebra, as well as numerical fluid dynamics. 1 of 2 studenten 1 student
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING PARAMETER ESTIMATION AND MODEL ORDER REDUCTION OF PDE‐MODELS FOR BIOSTATISTICS GUIDANCE Karl Meerbergen, Dirk Nuyens For more information, please contact:
Promoter:
Karl Meerbergen, Dirk Nuyens
Supervisors:
Pieter Lietaert, Xavier Woot de Trixhe (Janssen pharma)
CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part) The pharmaceutical industry uses parameter estimation to match mathematical models with measurements. The goal is a better understanding of the mechanism of diseases and medication, in contrast to machine learning techniques that do not provide the same detail. The parameter estimation algorithms are time consuming: a global optimization of the likelihood function needs to be performed, where evaluating the function and its gradient require solving a PDE for sample points in the often high dimensional stochastic parameter space. GOAL In this thesis, we will develop a numerical algorithm for maximizing the likelihood function. The focus is on fast computation of the high dimensional integrals over the output of the PDE. In order to reduce the complexity, we will use model order reduction techniques, i.e., the discretized PDE is first reduced to a small system of ODEs by using algebraic reduction techniques. The reduced model allows for fast evaluation of the solution of the PDE and its gradient in the sample points needed for computing the likelihood by numerical integration. Various sampling and optimization methods (local and global), including lattice rules and tensors, will be compared with the state of the art in biostatistics. METHODOLOGY The thesis will use results of previous master’s theses and a research project with intel and Janssen farmaceutica. The benchmark problem is a 3D reaction‐diffusion equation with reaction and diffusion coefficients whose probabilistic distribution, or the low order moments of the distribution, need to be determined. The focus of this work is on the interplay between model order reduction and sampling methods for numerical integration and numerical optimization.
PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera)) The thesis is rather practical, i.e., the roadmap is clear: it is a combination of known concepts. For students interested in a theoretical topic, it is possible to focus on clever combinations of model order reduction with lattice rules or tensors (without the parameter estimation itself). The choice of programming language is left to the student. 1 of 2 studenten 1
Fast methods for solving matrix distance problems Begeleiding: Wim Michiels Voor meer informatie contacteer: Wim Michiels, Karl Meerbergen Promotor Wim Michiels, Francesco Borgioli Begeleider(s): Context Distance problems receive an increasing interest in several fields of science and engineering, driven by the importance from both a theoretical and application point of view. Mathematical models for dynamical systems are namely characterized by uncertainty on parameters. An important question is then how large the uncertainty on the mathematical model can be without losing a desired property such as stability (called distance to instability), a level of performance or robustness. In this thesis we focus on eigenvalue problems with uncertainty, which arise in the frequency domain description of linear systems. Existing methods for the associated distance problems are computationally demanding, limiting the applicability to small‐scale model problems. The thesis fits within a collaboration with Prof. Emre Mengi (Koc University, Istanbul). DOEL Recently, a fast iterative method has been proposed for computing the distance to instability, which relies on solving a so‐called optimization problem. In the set‐up, it is assumed that the uncertain parameters are complex valued. This simplifies the analysis significantly, but is not realistic from an application point of view. Combining recent advances in linear algebra and control theory indicates that it is possible to relax this stringent assumption. The goal of the thesis is two‐fold: ‐ to analyze this method, to implement it and compare with existing methods; ‐ to extend this method from complex valued to real valued uncertain parameters. Nonlinear eigenvalue problems are important in mechanical and civil engineering applications (noise and vibrations), systems and control (handling feedback delays) and quantum‐physics (computing bound states), just to mention a few. UITWERKING 1. Study of the literature, analyzing the method, implementation and benchmarking 2. Extension from complex to real valued uncertainty 3. Validation on an application from mechanical engineering. PROFIEL Both a theoretical and implementation component, the theoretical component is the most challenging. Software in MATLAB Prerequisite: basic matrix computations. It can be useful to take the course "Numerical linear algebra" in the 2nd master. 1 student 1
TITEL Spectral optimization for vibration control of mechanical systems BEGELEIDING Wim Michiels Voor meer informatie contacteer: Wim Michiels Promotor Dan Pilbauer, Wim Michiels Begeleider(s): CONTEXT Recently, promising methods for vibration control of mechanical structures have been proposed, which include direct feedback of accelerometers’ outputs in the control design. At the same time a novel of class of input shapers has been proposed. The goal of including input shapers in a control system is to adapt references, control inputs or feedback signals, in order not to excite part of the structure (for instance, once want to adjust the position of a crane in a fast way, without having cable and load swing too much). This thesis fits within a collaboration between the Numerical Analysis and Applied Mathematics Division of KU Leuven and the Department of Instrumental and Control Engineering of the Czech Technical University in Prague (CTU).
(a) vibration control concept (b) Experimental set‐up at CTU DOEL The goal of the thesis is to develop computational tools for simultaneous design of controller and shaper parameters. As the system is assumed linear, this objective can be turned into an eigenvalue optimization problem, where the available parameters need to be tuned to optimize the location and/ or sensitivity of the eigenvalues of the closed‐loop, controlled system. UITWERKING Literature study, experiments with existing codes Development of algorithms for spectrum computation and optimization Application / validation for the model corresponding to the experimental set‐up PROFIEL
A theoretical, algorithm development and implementation component. Software in MATLAB. Prerequisite: basic matrix computations. It can be useful to take the course "Numerical linear algebra" in the 2nd master. 1 student 1
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Quasi‐Newton operator splitting methods for large‐scale conic optimization
GUIDANCE •
For more information, please contact:
[email protected]
•
Promotor:
[email protected]
•
Supervisor:
[email protected]
CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part) The exponential increase in computing power as predicted by Moore's law in combination with the need for improved performance in engineering systems as well as the ever‐increasing volume of stored data have rendered optimization ubiquitous in almost every branch of engineering. From embedded control and signal processing to data mining and machine learning, optimization problems arise everywhere. Convex optimization is a broad class of problems for which there exist solvability guarantees in polynomial time. It turns out that many problems (more than people initially imagined) can be formulated or approximated as convex ones and in particular be described through a conic optimization formalism which is considered a standard formulation widely supported by modeling languages such as CVX and YALMIP. Although conic optimization problems with some hundreds or thousands of variables and constraints can now be reliably solved using interior point methods, the status for large‐scale problems (in the order of millions of variables) is not the same. In particular, interior points methods simply cannot be applied due to their high complexity per iteration and storage requirements. This calls for new algorithmic developments to address this challenges. GOAL The purpose of this thesis is to study, specifically tailor, implement and test a new in‐house algorithm for convex optimization that can scale‐up to huge problem sizes, yet it is reliable enough to be used as a general‐purpose conic optimization solver. The algorithm was recently conceived and mathematically justified by the team of the promotor and the first numerical results are very encouraging. It combines ideas from operator splitting and quasi‐Newton methods. The goal of the thesis will be to find out which formulation of conic problems fits better into the solver (there are many), how to appropriately scale the problem data and finally an efficient C/C++ implementation of the solver. The solver will be applied to a large variety of problems ranging from model predictive control for smart grids to medical applications such as radiation treatment planning. [1] Boyd, S., Parikh, N., Chu, E., Peleato, B., & Eckstein, J. (2011). Distributed optimization and statisti cal learning via the alternating method of multi pliers. Foundati ons and Trends® in Machine Learning, 3(1), 1-122. [2] Parikh, N., & Boyd, S. P. (2014). Proximal Algorithms. Foundati ons and Trends in optimization, 1(3), 127-239.
RESULT The main outcome of the thesis will be software code implementing the algorithm and a thorough comparison with the state‐of‐the‐art. Potentially the algorithm will be linked as a general purpose solver to CVX, a modeling language for convex optimization. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera)) Literature and study ‐ 20%, Implementation ‐ 40%, Experiments ‐ 40% 1 or 2 students
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Newton-type operator splitting methods for embedded model predictive control
GUIDANCE For more information, please contact:
[email protected]
Promotor:
[email protected]
Supervisor:
[email protected]
CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part)
Model Predictive Control (MPC) is a real-time, optimization-based control strategy where at every sampling period, new information about the evolution of the plant (e.g. state measurements) is used to set up an optimal control problem based on a dynamical model. The problem is solved in real-time to obtain a sequence of future optimal control moves, but only the first one is applied to the real plant, and at the next time instant the same procedure is repeated. The benefit of MPC over conventional control strategies is the ability of online re-planning by using a model to predict the future over a finite time horizon (e.g., think of a robot interacting with a human where based on measurements, the robot can predict human action and adjust its behavior accordingly in order to perform a collaborative task) and the ability to respect actuator and output constraints. The advances in numerical optimization algorithms and in the availability of computational power have made online optimization possible in domains, such as automotive, aerospace, mechatronics and robotics, for which its use was unthinkable only a decade ago. In particular, model predictive control (MPC) is already in the commercialization phase in industries with products where dynamics are fast and computing power is low, such as the automotive industry and in power electronics. However, real-time optimization for MPC is challenging due to the following reasons: 1. The algorithm should be simple enough so as to be able to execute on low-cost, lowpower embedded hardware devices with limited computational resources (microcontrollers). 2. It should be robust and reliable with worst-case performance guarantees in order to be able to apply it in safety critical applications (e.g., engine control of a car or spacecraft take-off and landing). 3. It should be fast enough to be able to calculate the control action in the range of milliseconds or even nanoseconds.
GOAL
The purpose of this thesis is to study, implement and test a new in-house algorithm for convex optimization, specifically tailoring the solution to a wide range of MPC problems. The algorithm was recently conceived and mathematically justified by the team of the promotor and the first numerical results are very encouraging. It combines ideas from operator splitting and Newton-type methods. The goal of the thesis will be to identify a sufficiently large family of MPC problems (with/without state constraints, soft/hard constraints, linear/nonlinear dynamics and so on) and formulate them in a way which is amenable to be solved by means of the algorithm. Then the student should devise an efficient implementation of the algorithmic steps for such problems. The solver must be realized in C/C++, so that all the operations performed (matrix factorizations, matrix-vector products, memory allocation, etc.) can be fine-tuned and rigidly controlled. Alternatively, a higher level framework can be realized (in MATLAB, Python, or other high-level languages for technical computing) that uses code generation to produce the C/C++ code to be run on embedded platforms.
RESULT The main outcome of the thesis will be a C/C++ code, or a code generation framework, which implements in an efficient way the algorithm for MPC problems. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
Literature and study - 20%, Implementation - 40%, Experiments - 40% 1 or 2 students
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Robust optimization by means of Powel‐Sabin splines GUIDANCE Goele Pipeleers For more information, please contact: Stefan Vandewalle and Goele Pipeleers Promotor: Erik Lambrechts Supervisor: CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part)
While many engineering problems are nowadays translated into a mathematical optimization problem, the associated numerical data are often inaccurate or uncertain. As this uncertainty of data may turn the numerical optimum into a poor or even unacceptable solution for the true problem, there is a great push for robust optimization techniques suited for engineering applications. Robust optimization seeks a solution that is feasible for all possible values of the numerical data and gives the best worst‐case performance. As the data is generally considered to vary in infinite sets, robust optimization problems are numerically intractable in general. Therefore most problems are currently relaxed ––that is: replaced by a tractable, yet conservative approximation–– by means of Pólya’s theorem or sum‐of‐squares certificates of positivity. As these approaches suffer from limited applicability or high computational load, we have recently developed a novel relaxation scheme based on tensor‐product B‐splines and their properties. This approach has already been show very powerful in solving robust control and optimal control problems. To facilitate the usage of the novel relaxations a software toolbox for efficient spline manipulations and optimization has been developed (C++ with Python and Matlab interfaces through SWIG). GOAL In this master thesis you will extend the spline based robust optimization approach with relaxations based on Powel‐Sabin splines. These splines are naturally defined on polyhedral sets and hence, for such uncertainty sets relaxations based on Powel‐Sabin splines are expected to outperform the current ones based on tensor‐product splines. The thesis involves practical as well as theoretical parts: extension of the current toolbox with an efficient implementation for Powel‐Sabin splines, analysis of the conservatism of the relaxation as a function of its complexity (number of constraints), derivation of appropriate refinement strategies to reduce conservatism, comparison with existing relaxation approaches on various robust optimization problems. RESULT A new methodology for deriving relaxations of robust optimization problems based on the properties of Powel‐Sabin splines that is (i) complemented with an efficient toolbox to supports its use in practice; (ii) thoroughly analyzed and compared to existing approaches; and (iii) evaluated on various robust optimization problems. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
The thesis focusses on practical software implementation, but also includes theoretical parts. 1 or 2 students 1 student
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITEL Analyse en synthese van 3D puntenwolken BEGELEIDING Christopher Leyssen: 0474 503 411, Dirk Roose Voor meer informatie contacteer: Dirk Roose Promotor Christopher Leyssen (Argon www.argon‐ms.com) Begeleider(s): CONTEXT Bij het produceren van producten (bv een brandstoftank) is er regelmatige controle nodig om de afwijking van het geproduceerde product t.o.v. het model te inspecteren. Dit kan bekomen worden door een 3D scan van het product te maken. Om zo een 3D scan te visualiseren maken we gebruik van een warmtebeeld die de afwijking t.o.v. het model weergeeft. Dit kan momenteel enkel gebruikt worden om momentopnames te visualiseren en geeft ons geen totaalbeeld over wat er mis loopt bij de productie. DOEL Aangezien er bij vele productieprocessen op regelmatige basis 3D scans van de producten worden genomen zou het mogelijk zijn om meer geavanceerde statistieken te berekenen over een set van vele 3D scans. Bv op basis van een 3D sigma plot zou men kunnen afleiden waar er onstabiliteit zitten in het productieproces. Hierdoor kan men voorspellen en verifiëren welke aanpassing een beter resultaat zou geven. Het gaat hier vaak over honderden scans die op hun beurt elk honderdduizend punten kunnen bevatten. Daarbovenop zou het handig zijn voor de gebruiker om te bepalen voor welke set van scans hij statistieken wil zien. Dit wil zeggen dat de berekeningen zeer snel moeten kunnen uitgevoerd worden op grote datasets. UITWERKING Onderzoeken hoe deze data best voor te stellen Onderzoeken hoe deze voorstelling op te slaan met het oog op performantie Onderzoeken hoe de berekeningen uit te voeren op deze sets (sigma, range, ppk, gemiddelde, …) De resultaten terug om zetten naar een standaard puntenwolk formaat zodat deze kan weergegeven worden als warmtebeeld PROFIEL (bv. eerder theoretisch/eerder implementatie, vereiste voorkennis (bv. gevolgde of te volgen vakken, methodieken,programmeertalen, ... ))
Eerder implementatie 1 of 2 studenten
1 student
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE: Participation in the development and testing of the iPic3D‐MLMD code GUIDANCE Maria Elena Innocenti, For more information, please contact:
[email protected] Giovanni Lapenta, Dirk Roose Promoter: Maria Elena Innocenti Daily supervisor: CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part) The Multi‐Level Multi‐Domain (MLMD) method is a semi‐implicit, adaptive, fully kinetic method for Particle In Cell plasma simulations [Innocenti et al., JCP, 2013; Beck et al., JCP, 2014; Innocenti et al., CPC, 2015; Innocenti et al., JCP, 2016]. The aim of the method is to reduce the huge computational cost of kinetic simulations of plasmas. We use, together, a semi‐implicit algorithm, the Implicit Moment Method (IMM), and an adaptive approach. With the IMM, the strict stability constraints of explicit methods are removed. With the adaptive approach, high resolution is used only in small parts of the total domain, again reducing the computational cost of the simulations. The MLMD method is currently implemented in a two dimensional code, Parsek‐MLMD. We are implementing it in the three dimensional code iPic3D [Markidis et al., 2010]. The master thesis student will participate in the testing of the different components of the iPic3D‐ MLMD code in their early stage of development. He/ she will gain precious experience with one of the leading codes for plasma simulations, iPic3D, and with the MLMD method, at the moment the only semi‐implicit adaptive method for plasma simulations. GOAL The student will participate in the definition and execution of suitable tests, which will include the basic functioning of the code (e.g., correct definition of the adaptive geometry), the correct simulation of physical processes (from basic simulations of thermal plasmas, to kinetic instability, to magnetic reconnection, the main process that we want to simulate with the iPic3D‐MLMD code) and possibly performance testing with appropriate profiling tools, e.g. Scalasca and DDT. EXPECTED RESULT At the end of the thesis, the student will have acquired direct practical knowledge of advanced kinetic method for plasma simulations. He/ she will have participated to the day‐to‐day issues of testing a developing code. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera)) Practical implementation, experience with C/C++, parallel programming (MPI), scripting languages (MATLAB or python), 1 or 2 students? 1 student
MASTERPROEF MASTER WISKUNDIGE INGENIEURSTECHNIEKEN TITEL Voorspellend model voor koelmal‐afstellingen BEGELEIDING François Justin: 0498 45 57 11, Dirk Roose Voor meer informatie contacteer: Dirk Roose Promotor François Justin (Argon www.argon‐ms.com) Begeleider(s): CONTEXT De grootste uitdaging bij de productie van een brandstoftank is om de vorm ervan binnen tolerantie te krijgen na het blaasproces (blow molding). Dit proces bestaat uit 2 stappen: het blaasproces zelf en het afkoelen van de tank in een waterbad. Het is tijdens het afkoelen in dit waterbad dat de finale vorm van de tank wordt vastgelegd. Tijdens het afkoelen vervormt de geblazen tank significant. Deze vervormingen kunnen op dat moment nog niet voorspeld worden, daarom wordt er gebruik gemaakt van een ‘koelmal’ om de tank in een bepaalde vorm te houden tijdens het koelproces. Na een 2tal minuten te hebben gelegen in een koud waterbad, wordt de tank uit de koelmal gehaald en neemt deze dan zijn finale vorm.
De afstelling van de koelmal bepaalt de finale vorm. Het is niet zo dat de koelmal afgesteld wordt volgens CAD of een berekenbare waarde. Bij de start van de productie wordt via trial en error de koelmal afgeregeld door een expert gebruiker. Een blok harder laten duwen tegen de tank aan de ene kant kan invloed hebben op een andere gebied van de tank. Op heden zijn er een 15‐tal iteraties nodig om de koelmal af te stelen, waarbij een iteratie = koelmal afstellen, tank produceren, tank opmeten en analyseren (3D scan) om een nieuwe afstelling van de koelmal te proberen (ervaring gebaseerd). Dit heel proces kan ongeveer 1 week in beslag nemen. Eens de koelmal afgesteld, kan de productie starten. Typisch wordt er ongeveer gedurende 1 week geproduceerd, waarna de machine omgebouwd wordt voor de productie van een ander model en wordt het afstelproces opnieuw gestart. Na een aantal weken wordt dezelfde tank opnieuw geproduceerd en moet het hele afstelproces opnieuw doorgelopen worden. De afstellingen van de koelmal kunnen niet overgenomen worden omdat de instellingen van het blaasproces onmogelijk herhaalbaar her‐ingesteld kunnen worden. DOEL Het doel van de masterproef is om het aantal iteratiestappen om de koelmal af te stellen sterk te reduceren. Met behulp van de 3D scan technologie kan het effect van de verschillende contour elementen nauwkeurig en volledig in kaart gebracht worden (3D puntenwolk) en gegoten worden in een slim voorspelend model. Het model zal niet enkel academisch getest worden, maar zal ook effectief getest worden op het echte productieproces bij TI Automotive. UITWERKING Volgende aspecten zullen aan bod komen: bewerkingen op puntenwolken, simulatie en optimalisatie. PROFIEL (bv. eerder theoretisch/eerder implementatie, vereiste voorkennis (bv. gevolgde of te volgen vakken, methodieken,programmeertalen, ... ))
Combinatie van theoretsche analyse en praktische uitwerking en implementatie 1 of 2 studenten 1 student
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE
Analysis of efficient simulation methods for stochastic processes GUIDANCE Giovanni Samaey For more information, please contact: Giovanni Samaey Promotor: Przemek Zielinski; Giovanni Samaey Supervisor: CONTEXT Stochastic processes model a broad spectrum of phenomena, from the motion of large particles in a solvent, over cells in a population of organisms to products on the financial markets and behavior of people. Often, one is not so much interested in the specific details of an individual simulation, but in a set of averaged, "macroscopic" quantities, such as the average speed at which particles spread, the density of bacteria or the expected price of a certain stock. Monte Carlo simulation then computes the evolution of these macroscopic quantities by averaging over a large number of realizations of the stochastic process. Often, stochastic processes exhibit some form of stiffness: an individual realization of the stochastic process needs to be simulated using very small time steps, whereas the behavior of interest manifests itself over very long time scales (so that many time steps need to be taken). In such cases, one may possess an approximate, coarser description of the same system, in which the fast degrees of freedom have been eliminated. Such a description can then be simulated much more efficiently than the original fine‐scale system due to two reasons: (1) since there are no fast time scales, larger time steps can be taken; and (2) the system contains less degrees of freedom, so there is less work per time step. These gains, however, come at the price of a loss in accuracy. GOAL We will analyze and test a new algorithm to accelerate such stochastic simulations. The algorithm combines Monte Carlo simulation of the full fine‐scale system (over short time intervals) with extrapolation over macroscopic time scales based on the approximate coarser model. A crucial step in the algorithm is to reinitialize the full fine‐scale system, based on the extrapolation at the coarse level. The general strategy is to create a fine‐scale probability density that is both consistent with the extrapolated coarse state and as close as possible to the fine‐scale state before extrapolation. RESULT The mathematical challenge in the above algorithm is to attach an appropriate meaning to the phrase "as close as possible". We will look into some of the following questions: what norm is appropriate for the distance between probability measures? Do some norms allow for a better conservation of dynamics of the underlying process? Do some norms naturally lead to a convenient algorithm? How do we create a random sample from the obtained probability distribution? PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
This master thesis will involve some theoretical work (approximation and numerical analysis), complemented with numerical experiments. The relative weight of these two components will depend partly on the preferences of the student. Some knowledge of stochastic processes is welcome. 1 of 2 studenten
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Asymptotic Preserving methods to address the discrete‐continuum duality GUIDANCE Giovanni Lapenta For more information, please contact: Giovanni Lapenta ‐ Giovanni Samaey Promotor: Giovanni Lapenta Supervisor: CONTEXT
The description of Nature presents to us a duality that has followed human thinking for centuries: should we describe the natural world as a collective of discrete agents or as a continuum described by partial differential equation? Cellular automata, lattice‐Boltzmann methods, molecular dynamics, particle methods belong to the atomistic tradition. Solving a PDE with mathematical methods implies assuming a continuum description and an infinitely divisible continuum. Practice and experiment tell us that in some circumstances one description is more practical than the other. When describing a DNA molecule, one needs to consider the constituting atoms and all interactions among them. When describing the climate of the Earth, the atmosphere needs to be considered as a continuum with local properties that vary continuously in space and in time. Modern science pushes both approaches to their limits. The continuum model is pushed to such small scales that particle effects start to come into play, ruining the viability of a purely continuum approach. The discrete approach is pushed to the limit of considering so many interacting agents that even the world’s fastest supercomputers cannot handle the problem anymore. Two approaches then become possible. We can start from a continuum model and introduce discrete effects, for example when introducing statistical perturbations that resemble the randomness introduced by the chaotic motion of particles. Or we can start from discrete methods and approximate many‐body interactions in the limit of infinitely many particles using continuum approximations, as is done to describe water molecules in some DNA simulations. GOAL Our goal is to design an Asymptotic‐ Preserving (AP) discrete method that naturally transitions to a continuum model in the limit of large scales. In the present thesis, we will focus on a system of electrons interacting via electrostatic forces only. In the discrete description, the Boltzmann equation provides a statistical knowledge of particle position ands velocities. In the continuum description, the system is described by fluid equations for the electrons. RESULT We will proceed with successive steps: 1. We consider how one can go from the discrete to the continuum with the moment method 2. We will see how the discrete and the continuum methods can be solved in practice 3. We will design asymptotic methods, capable of automatically transitioning between discrete and continuum level of description. PROFILE The profile is theoretical with a strong numerical analysis and high performance computing component. The work can be designed to be more purely theoretical (i.e. pencil and paper) or more practical applied computing (i.e. code developing and programming) 1 of 2 studenten The work can be done alone or in team of 2
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Modified stochastic processes for nuclear fusion simulations GUIDANCE
[email protected] For more information, please contact: or
[email protected] Giovanni Samaey Promotor: Bert Mortier Supervisor CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part)
Nuclear fusion uses a clean, cheap and inexhaustible fuel (hydrogen) and therefore has the potential to provide a sustainable answer to our energy needs. Unfortunately, there are still some obstacles to clear before nuclear fusion can be used commercially. One example is the efficient removal of the reaction product helium and the created energy from the reactor. Because of the high temperatures of the plasma inside the reactor, the working conditions of a fusion reactor are unreachable experimentally for testing. (No material resists these temperatures.) Therefore, numerical simulation is the only viable path to efficient reactor design. Unfortunately, these simulations themselves also reach the limits of what is currently achievable: the simulation times are excessive, and it is very difficult to reach the required level of accuracy. GOAL Nuclear fusion simulation currently makes use of a fluid description of the plasma (consisting of electrons and ions), with partial differential equations for mass, momentum and energy. The neutral particles that appear at the edges of the reactor need to be modeled with a more microscopic (kinetic) description that is capable of describing collisions between particles in sufficient detail, requiring also information on the neutral particle velocities. The resulting kinetic equation has an increased dimension, since it describes the fraction of particles at a given position with a given velocity. It is therefore usually discretized using a Monte Carlo method, using stochastic particles that move and collide. These Monte Carlo simulations can become quite cumbersome when the frequency of collisions increases (and the mean free path decreases correspondingly). The goal of this thesis is the construction and analysis of algorithms to overcome this computational bottleneck. RESULT In the limit of zero mean free path, the process corresponds to Brownian motion, implying simulations can be performed without explicitly simulating every collision. In this thesis, we will propose and study a modified stochastic process that automatically incorporates Brownian‐motion‐ like behavior whenever this allows a significant speedup without compromising accuracy. Nuclear fusion simulations require the computation of path‐averaged properties of the neutrals. We will thus analyze how the modified stochastic process affects those quantities. The proposed process additionally has the property that it corresponds to a deterministic fluid model in the Brownian limit, allowing it to be used for variance reduction. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
The focus lies with the analysis of the modified stochastic process, but can shift towards implementation depending on the interest of the student. Some affinity with stochastic processes and/or numerical methods for differential equations is beneficial. Specific knowledge on numerical methods for stochastic processes can be acquired via optional courses during the master thesis. 1 of 2 students
MASTERPROEF MASTER WISKUNDIGE INGENIEURSTECHNIEKEN Aangepaste stochastische processen voor kernfusiesimulaties BEGELEIDING
[email protected] Voor meer informatie contacteer: of
[email protected] Giovanni Samaey Promotor Bert Mortier Begeleider(s): CONTEXT (Info over het probleem, ev. info over het onderzoeksproject waarin deze masterproef kadert) Kernfusie heeft het potentieel een duurzame oplossing te bieden voor onze energiebehoeften. Om tot een commerciële uitbating te komen zijn er echter nog enkele obstakels te overwinnen. Een voorbeeld daarvan is het efficiënt verwijderen van de geproduceerde heliumassen. De extreme omstandigheden in de reactor bemoeilijken experimenten dermate dat numerieke simulatie de te volgen optie is om scenario’s te testen en reactorontwerpen op punt te stellen. Deze simulaties vereisen meer dan wat momenteel mogelijk is. De berekeningstijd is nog veel te groot en de vereiste nauwkeurigheid bereiken is erg moeilijk. DOEL Kernfusiesimulaties maken momenteel gebruik van een fluïdummodel voor het plasma (bestaande uit geladen deeltjes), met partiële differentiaalvergelijkingen die de balansen van massa, momentum en energie modelleren. De neutrale deeltjes, die voorkomen aan de randen van de reactor, vereisen meer detail in hun modellering. Voldoende detail wordt bereikt met behulp van een kinetische beschrijving die de botsingen van die deeltjes voldoende gedetailleerd weergeeft. Het effect van deze botsingen hangt namelijk ook af van de snelheid van de neutrale deeltjes. Deze kinetische beschrijving heeft daarom een verhoogde dimensie, doordat de neutrale deeltjesdichtheid, naast de positie, ook in functie van de snelheid wordt beschreven. Die verhoogde dimensie noopt tot het gebruik van een Monte Carlo methode, die gebruik maakt van deeltjes die bewegen en botsen. Deze simulatie kan te duur worden wanneer er erg vaak botsingen plaatsvinden (en de gemiddelde afstand tussen botsingen dus erg klein kan worden). Het doel van deze thesis is het opbouwen en analyseren van algoritmes om deze computationele problemen op te lossen. UITWERKING In de limiet naar oneindig korte paden tussen botsingen, stemt dit overeen met een Browniaans pad. Dit laat toe de deeltjes te simuleren zonder elke botsing expliciet te simuleren. In deze thesis zullen we een aangepast stochastisch proces voorstellen en bestuderen dat automatisch een Browniaans aandoende beweging aanneemt wanneer dit een aanzienlijke efficiëntiewinst betekent, zonder onaanvaardbare verliezen in nauwkeurigheid. Binnen kernfusiesimulaties zijn gemiddelde kenmerken over het gevolgde pad van de neutralen van belang en er zal moeten worden geanalyseerd of die eveneens kunnen worden gevonden met een aangepast stochastisch proces, dat kenmerken vertoont van een Browniaanse beweging. Het aangepaste stochastische proces heeft tevens het kenmerk dat het in de limiet voldoet aan een deterministisch fluïdummodel, wat kan worden benut voor variantiereductie. PROFIEL (bv. eerder theoretisch/eerder implementatie, vereiste voorkennis (bv. gevolgde of te volgen vakken, methodieken,programmeertalen, ... ))
De focus ligt in de eerste plaats op analyse en uitwerking van het aangepaste stochastische proces, maar kan bijgestuurd worden naar meer implementatie, afhankelijk van de interesse van de student. Affiniteit met stochastische processen en/of numerieke methoden voor differentiaalvergelijkingen is een voordeel. Specifieke kennis over numerieke methoden voor stochastische processen kan worden verkregen via optionele vakken gedurende de thesis. 1 of 2 studenten
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Study of out‐of‐equilibrium effects in kinetic models for traffic flow GUIDANCE
[email protected] For more information, please contact:
[email protected]; Promotor:
[email protected] [email protected] Supervisor: CONTEXT The mathematical structure of traffic flow models is comparable to that of other flow phenomena that exhibit waves, such as fluids in open channels, compressible fluids in tubes, etc. A particular type of traffic flow models, initially proposed by the Belgian Nobel prize winner Prigogine in the sixties and expanded afterwards, considers each car as a particle. Its equations of motion then represent the behavior of the driver, making modeling potentially more realistic. This results in so‐ called gas‐kinetic models, which are very similar in structure to models that describe fluid flow. In principle, gas‐kinetic models of traffic flow represent traffic as a large collection of individual particles that have a position and velocity in phase space. The evolution law then describes how the phase space density evolves as a function of time. Because this equation (of Boltzmann‐type) is high‐dimensional, one typically resorts to more macroscopic fluid approximations of the kinetic model. In these macroscopic models, one averages out the velocity directions to obtain a model that describes the evolution of car density and flow speed as a function of space and time. Unfortunately, to obtain a closed partial differential equation for these quantities, one needs to assume that the velocity distribution equilibrates quickly to a function of car density and flow speed. This is not always true, as can easily be seen by thinking about a long line of cars that is stuck behind a single slow driver that cannot be overtaken easily. GOAL The effect of the macroscopic approximation is that a lot of the richness of the original model is lost. Even some of the most crucial features, such as the spontaneous emergence of stop‐and‐go waves in dense traffic, which are present in the kinetic model, disappear when passing to the approximate macroscopic model. In this thesis, we will solve this by developing a Monte Carlo simulation of the detailed gas‐kinetic model for traffic flow. We will then analyze in more detail what the reasons are for the qualitative difference in solutions between the gas‐kinetic model and its macroscopic approximation, with the goal of developing strategies that can automatically decide in which parts of the domain the current traffic situation allows the use of the (computationally advantageous) macroscopic model or requires the (more detailed) Monte Carlo simulation to be used. RESULT The result of this thesis will be a numerical method for traffic flow that combines the computational efficiency of macroscopic models with the modeling detail of gas‐kinetic Monte Carlo simulation. PROFILE Some affinity with stochastic processes and/or numerical methods for differential equations is beneficial. Specific courses on numerical methods for stochastic processes can be acquired via optional courses during the master thesis. 1 of 2 studenten
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE
Variance-reduced time integration of metastable stochastic processes GUIDANCE Giovanni Samaey For more information, please contact: Giovanni Samaey Promotor: Ward Melis; Giovanni Samaey Supervisor: CONTEXT A particularly interesting class of stochastic problems exhibits so‐called metastable behavior: the system fluctuates around some average state for extended periods of time, and occasionally ``switches regime'' quickly and drastically. Examples arise in polymer physics (in which the regime switch is a conformation change of the polymer), gene regulatory networks (in which gene expression can suddenly switch) or ecological systems (with drastic changes in ecosystems). For such systems, it is not possible to predict exactly when a regime shift will occur, but one can use Monte Carlo simulation to estimate the probability of such a shift. Mathematically speaking, such regime switches correspond to rare events, in which a transition occurs between two metastable states. The events are called ``rare'' because the typical time between two such events is much larger than the time step required for an accurate simulation of the stochastic process. One therefore cannot really expect these rare events to be regularly observed during even the most time‐consuming simulations. Computation of statistical quantities (such as rare event probabilities or mean time between events) may therefore be computationally prohibitive. At a coarser level, one can consider the construction of a finite state Markov chain model. Then, only the transitions between metastable states are modeled, and all fine‐scale dynamics is discarded. Of course, one can usually only obtain the transition rates approximately. GOAL We will analyze a new algorithm to simulate such metastable stochastic simulations. The algorithm uses Monte Carlo simulation of the full fine‐scale stochastic process to estimate the transition probabilities of a coarser finite state Markov chain. One particular aspect that will be investigated is the effect of the fine‐scale noise on the estimations. RESULT One approach that we are currently working on is the use of correlated random numbers for variance reduction purposes. In this master thesis, we will test different strategies, analyze their convergence behavior and apply the method to some test problems. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
Depending on the interest of the student, the focus can be numerical, analytical or applied. Some affinity with stochastic processes and/or numerical methods for differential equations is beneficial. Specific knowledge on numerical methods for stochastic processes can be acquired via optional courses during the master thesis. 1 of 2 studenten
MASTERPROEF MASTER WISKUNDIGE INGENIEURSTECHNIEKEN TITEL Data mining technieken voor machine/voertuig fleet monitoring BEGELEIDING
[email protected] Voor meer informatie contacteer: Johan Suykens Promotor Bram Cornelis (Siemens Industry Software) Begeleider(s): CONTEXT (Info over het probleem, ev. info over het onderzoeksproject waarin deze masterproef kadert) Siemens Industry Software (SISW) is een provider van geavanceerde simulatie en test software/hardware tools, die de design teams binnen de automobiel, lucht‐ en ruimtevaart en algemene mechanische industrie bijstaan tijdens de productontwikkeling. Door opkomende technologie trends zoals embedded sensors en Internet‐Of‐Things is er tegenwoordig ook veel “operationele” sensor data beschikbaar, nl. data die verzameld wordt wanneer het afgewerkt product (een machine/voertuig/…) door de eindgebruiker gebruikt wordt. Een van de toepassingen voor dergelijke operationele data is zogenaamde “conditie‐monitoring”. Dit zijn systemen waarbij de conditie van machines en hun kritische onderdelen (bvb. een elektro‐mechanische aandrijflijn in productiemachines of in elektrische voertuigen) automatisch opgevolgd worden. Bij langdurige machine‐operatie treden er na verloop van tijd elektrische of mechanische fouten (bvb. een defecte lager) op. Het is belangrijk om deze fouten zo vroeg mogelijk te detecteren, zodat er tijdig een onderhoud kan voorzien worden om grote schade en lange machine “downtimes” te voorkomen. DOEL Bij conditie‐monitoring systemen worden vele sensoren permanent op en rond de machine geïnstalleerd (bvb. accelerometers om trillingen te meten, microfoons, etc.). Bij het optreden van voorgenoemde fouten, wordt aangenomen dat er afwijkende trends in de sensor signalen kunnen waargenomen worden. Het is echter niet eenvoudig om de grote hoeveelheid aan sensor‐data te verwerken en deze trends inderdaad terug te vinden. Het doel van de thesis is daarom om datamining technieken toe te passen om dit mogelijk te maken. Zogenaamde clustering methodes (bvb. k‐means of self‐organizing maps) kunnen bvb. toelaten om gelijkaardige data automatisch te groeperen, en zogenaamde anomalie‐detectie technieken kunnen toelaten om statistische “outliers” binnen een groep te vinden. UITWERKING Het wordt verwacht van de student dat hij/zij deze data‐mining algoritmes programmeert (op basis van beschikbare literatuur beschrijvingen), bij voorkeur in matlab/python of in gelijkaardige statistische / mathematische software (bvb. ‘R’). Er zullen enkele representatieve datasets (met echte meetdata) ter beschikking zijn van de student om de ontwikkelde algoritmes op uit te proberen. PROFIEL (bv. eerder theoretisch/eerder implementatie, vereiste voorkennis (bv. gevolgde of te volgen vakken, methodieken,programmeertalen, ... ))
‐ ‐
Het topic is gerelateerd aan de cursus “Data Mining and Neural Networks” (Suykens) De student(en) moeten ervaring hebben met een mathematische software omgeving zoals Matlab, Octave, Python, R of andere.
1 of 2 studenten
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Data‐driven scoring systems for medical decision support: beyond binary problems GUIDANCE
For more information, please contact:
Lieven Billiet
Promotor:
Johan Suykens, Sabine Van Huffel
Supervisor:
Lieven Billiet
CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part)
Clinicians are often overwhelmed by the amount of patient data they get. Their task is to formulate a diagnosis and/or prognosis based on relevant measures and values. Using their experience and expertise, they can take into account measures as diverse as age, medical record, pulse, heart rate, pain etc. One way to alleviate the information overload they are facing is by using a decision support system. This can guide them in selecting the most important measures and giving them a proper weight in their consideration. Many techniques of the machine learning community could be exploited to this end. However, although they are established, easily applicable and thoroughly investigated, they tend to yield black box models. Hence, clinicians are not willing to trust them, a characteristic that is vital for their acceptance by clinicians and patients alike. Contrarily, medical scoring systems have been used for decades in practice. Such systems consist of powerful rules of thumb, based on clinical expertise and experience. Unfortunately, they are sometimes rather ad hoc, though often validated empirically or statistically. GOAL The work aims to combine advantages of the mentioned systems by extracting a scoring system from data using machine learning techniques, discarding spurious information whilst still retaining the essentials. Some work has already been done. A system has been developed in the framework of Support Vector Machines and sparse optimization. Currently, scoring systems for binary (two‐class) problems can be derived from training data, and a risk/probability for the target class can be established. A first version of a toolbox is available for use in validation studies. The focus of this master thesis is the expansion of the current framework to include multiclass problems. For example, this proves useful to distinguish different types of tumours. RESULT
Reformulation of the optimization problem for multiclass, including the mapping to risks
Implementation of the solution of the optimization
Validation of the solution on public datasets or patient data obtained from UZ Leuven.
PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera)) The student should have some prior knowledge on (sparse) optimization and machine learning as a basis to acquire insight in the proposed framework. As an example, linear programming, support vector machines and elastic nets have been used up until now. Programming will be performed in Matlab. 1 student
MASTERTHESIS MASTER WISKUNDIGE INGENIEURSTECHNIEKEN MASTER MATHEMATICAL ENGINEERING Title: Solving parametrized polynomial eigenvalue problems Guidance: •
For more information contact:
Marc Van Barel or Karl Meerbergen
•
Promoters:
Marc Van Barel and Karl Meerbergen
•
Daily supervision:
Pieter Lietaert
CONTEXT
Consider a matrix, whose elements are polynomials dependent on a parameter. We want to find all or part of the eigenvalues and corresponding eigenvectors for values of the parameter belonging to a subset of the real line or complex plane. For a fixed value of the parameter, several solution methods exist for solving the corresponding polynomial eigenvalue problem, e.g., using contour integrals or using linearizations. We want to extend these methods generating the solution set depending on the parameter. GOAL
The aim of the thesis is developing the theory and corresponding algorithms to solve parametrized polynomial eigenvalue problems. METHODOLOGY
The research includes a study of the literature, the development of the theory and the corresponding numerical methods, the implementation and testing of these methods, and applying those to nontrivial examples. For the representation of the parametric dependencies there are several possible approaches. - for some problems, the parametric dependency is a rational function. - The dependencies may also be represented by discretisation into a grid. In this case, the eigenvalues and eigenvectors can be represented by a low rank tensor. The work consists of rewriting the eigenvalue problem using these representations. We will analyse properties of these representations and propose efficient numerical solution methods. PROFILE
Depending on the interest of the student, the focus can be either on theoretical aspects or merely on developing and implementation of algorithms. Attending the course on Numerical Linear Algebra is recommended. The student can use a programming language of choice. 1 or 2 students
1
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Two‐parameter eigenvalue problems GUIDANCE Marc Van Barel For more information, please contact: Marc Van Barel Promotor: Marc Van Barel Supervisor: CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part)
Several problems in science and engineering can be formulated as a two‐parameter eigenvalue problem: given square matrices , and , look for all vectors and and complex numbers and such that the following equations are satisfied: Consider, e.g., a system of polynomial equations , 0 , 0 in two variables and . This system can be linearized into a two‐parameter eigenvalue problem. References: M. E. Hochstenbach, T. Košir, B. Plestenjak: A Jacobi‐Davidson type method for the two‐parameter eigenvalue problem, SIAM J. Matrix Anal. Appl. 26 (2005) 477‐497 B. Plestenjak, M. E. Hochstenbach: Roots of bivariate polynomial systems via determinantal representations, accepted in SIAM J. Sci. Comput. , arXiv:1506.02291 GOAL The aim of this thesis is to investigate existing methods and to design new methods to solve the two‐parameter eigenvalue problem. RESULT The research includes a study of the literature, the development of the theory and the corresponding numerical methods, the implementation and testing of these methods in Matlab, and applying those to nontrivial examples. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
Depending on the interest of the student, the focus can be more on theoretical aspects or more on developing and implementation of algorithms. 1 or 2 students 1 student
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE: Fast and reliable computation of centralizers in large networks GUIDANCE Raf Vandebril For more information, please contact: Raf Vandebril Promotor: Thomas Mach Supervisor: CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part)
Data linked to applications is often modeled by networks and graphs, e.g., gene interactions, the internet, traffic flows, human interactions, electricity networks. Often it is desired to detect the most important or most central nodes in such networks. Computing these nodes relates to computing a matrix function of the adjacency matrix. In this thesis we will focus on computing the matrix exponential of the adjacency matrix. Unfortunately the networks are typically very large, and the computing time can become extremely large. In this thesis we will construct and test a new algorithm based on approximate rational Krylov methods to retrieve the most important nodes in such a graph. Experiments will be carried out on real application data.
GOAL We aim at developing a new algorithm, faster than the currently available methods. Depending on the student’s interest the topic can become more theoretical, more tailored towards high performance computing, more application oriented,… RESULT The student will first study some applications and the importance of centralizers. The link between graph theory and numerical linear algebra needs to be established, after which it will become clear that functions of matrices are an important ingredient in computing the centralizers. Also the link with approximate matrix functions via Krylov methods plays an important role. PROFILE The student should have a basic knowledge of numerical linear algebra, be aware of the important aspects in scientific computing, such as accuracy, memory, and computing time. The student should also be able to implement the algorithms in, e.g., Matlab. 1 of 2 studenten 1 or 2
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE: Transient analysis of the electro‐thermal evolution in power MOS transistors. GUIDANCE:
For more information, please contact:
Promoters
Wim Schoenmaker ‐ CTO MAGWEL NV
[email protected] Raf Vandebril , Karl Meerbergen
Daan Camps Supervisor: CONTEXT: PowerMOS transistors are 'heavy‐duty' microelectronic switches constructed by many transistor channels operating in parallel mode. Before actual fabrication, the designs of the transistors are intensively simulated in order to acquire knowledge about their reliability, switching capacities (delays) and avoidance of local hot spots that may lead to interconnect damage. Figure: Illustration of a powerMOS (left panel) , thermal profile at fixed cut plane (right panel) During the simulation, the local temperature and current densities are tracked at any time instance by solving energy balance equations and current‐continuity equations, using a finite‐element approach. The evolution in time (temporal changes) are computed usually by using backward‐ difference Euler formulas. Recently, it has become clear that an interesting alternative for BDF may be a more exact method using matrix exponentials. The system of discretized equations requires in this approach the solution of the system:
(1)
where A is a large sparse matrix and b is vector and X is a vector that needs to be computed. The dimension of X > 100.000. GOAL: Based on the system description it is desired to identify a suitable Krylov ‐ based solution of equation (1). An additional goal is to provide a C++ implementation of the proposed method as a library component in the linear algebra tool suite of MAGWEL. This can be based on software available by the promoters. METHODOLOGY: PROFILE (e.g. rather theoretical / rather practical implementation, foreground knowledge (courses, methods, computer language(s) et cetera))
Knowledge about Krylov solving and C++ programming is desirable. 1 of 2 studenten
1
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE: Computing the tensor geometric mean GUIDANCE Raf Vandebril & Ben Jeuris For more information, please contact: Raf Vandebril Promotor: Ben Jeuris Supervisor: CONTEXT In many applications the data contain noise or arise from various sources and need to be averaged in some way or another, e.g., in bio‐informatics connections between proteins, or in radar technology many sources need to be averaged. If the data are matrices or tensors, translating existing means of scalar numbers to a higher dimension is not necessarily trivial. The geometric mean of some positive numbers is defined as the n‐th root of the product of these numbers. Whereas the arithmetic mean averages some numbers, the geometric mean is more a measure to average ratios between numbers. The geometric mean is therefore a more natural mean for a whole variety of applications. Unfortunately, due to the noncommutativity of matrices, it is not straightforward to generalize this to matrices. However, in the domains mentioned in the beginning extending the geometric mean to matrices seems the natural thing to do. Nowadays there is consensus on how the matrix geometric mean should be defined. When going to tensors, however, things become even more complicated. A straightforward way to tackle the problem is to go to the flattening of the tensor, which is putting all data of the tensor in one big matrix. One possible solution in the matrix case is based on computing the centroid of a triangle as shown below. This is likely not optimal for all tensors, but by using the correct approximation of the tensor one can develop some algorithms which should approximate the ``tensor geometric mean’’.
GOAL The aim is to study manifold optimization techniques and/or basic linear algebra routines for tensors by considering the appropriate flattening. Depending on the student’s interest the topic can become more theoretical (e.g., matrix structures, tensors), more tailored towards high performance computing, more application oriented (radar technology, avereging biomedical data),… RESULT The student has to study the problem of computing the matrix geometric mean. Next we will see what sort of algorithm we want to implement for tensors and test on a specific application. PROFILE The student should have a basic knowledge of numerical linear algebra, be aware of the important aspects in scientific computing, such as accuracy, memory, and computing time. The student should also be able to implement the algorithms in, e.g., Matlab. 1 of 2 studenten 1
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Neural network based grapheme‐to‐phoneme conversion GUIDANCE For more information, please contact:
[email protected] Promotor: Marc Van Barel Co‐promotor
[email protected] Supervisor: CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part)
A speech recognizer traditionally uses an acoustic model, which determines the most likely phonemes based on the acoustics, and a language model, which determines the most likely word sequences. These two knowledge sources are combined via the lexicon that maps a written word to its pronunciation. Generating these pronunciations is done with the aid of grapheme‐to‐phoneme (G2P) conversion. ESAT's current G2P first looks up the word in a list with phonetic transcriptions, and passes the words that do not appear in the list through an induction decision tree system [1]. This system however is not always perfect and makes errors especially for names. Recently, neural network based G2P's have proven to be more effective: for example in [2] and [3], the authors use a bidirectional long short‐term memory (LSTM) network (a powerful type of recurrent neural network) to predict the pronunciation of a word. GOAL The purpose of this thesis is to implement and train a neural network based G2P: this involves investigating several model types and parameter settings and testing the generated pronunciations for our speech recognizer. Thanks to the existence of modular and easily adaptable libraries such as Tensorflow, Keras, Caffe, Torch..., the neural network will not have to built from scratch. [1] Demuynck, K., Laureys, T. and S. Gillis. 2002. "Automatic generation of phonetic transcriptions for large speech corpora". In Proc. ICSLP, Vol. 1. [2] Rao, K., Peng, F., Sak, H., Beaufays, F. 2015. "Grapheme‐to‐phoneme conversion using long short‐ term memory recurrent neural networks". In ICASSP. [3] Yao, K. and G. Zweig. 2015. "Sequence‐to‐Sequence Neural Net Models for Grapheme‐to‐ Phoneme Conversion". In Interspeech 2015. RESULT A possible application for this new G2P is a speech interface for KU Leuven's "who‐is‐who", which will be especially interesting since names are notoriously difficult for a G2P, or for the word list of the Dutch language (also known as "Het groene boekje"), which will enable the user to query for a name (who‐is‐who) or a word (Groene boekje) by voice. Optionally, it is possible to implement a web interface for this voice querying system as well, depending on the student's interests. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
Literature and study ‐ 20%, Implementation ‐ 40%, Experiments ‐ 40% 1 or 2 students
MASTERPROEF MASTER WISKUNDIGE INGENIEURSTECHNIEKEN TITEL Neuraal netwerk gebaseerde grafeem naar foneem omzetting BEGELEIDING Voor meer informatie contacteer:
[email protected] Promotor Marc Van Barel Co‐promotor
[email protected] Begeleider(s): CONTEXT (Info over het probleem, ev. info over het onderzoeksproject waarin deze masterproef kadert) Traditioneel gebruikt een spraakherkenner een akoestisch model dat de meest waarschijnlijke fonemen bepaalt gebaseerd op het akoestisch signaal, en een taalmodel, dat de meest waarschijnlijke woordsequenties bepaalt. Deze twee kennisbronnen worden gecombineerd via het lexicon dat een gesproken woord afbeeldt op zijn uitspraak. De generatie van deze uitspraak gebeurt aan de hand van grafeem naar foneemconversie (G2P). De huidige G2P convertor van ESAT zoekt een woord eerst op in een lijst met mogelijke uitspraken, en geeft de woorden die niet in de lijst voorkomen door aan een systeem gebaseerd op de inductie van beslissingsbomen [1]. Dit systeem is echter niet perfect en maakt fouten, vooral voor namen. Recent werd aangetoond dat G2P systemen gebaseerd op neurale netwerken effectiever zijn: in [2] en [3] bijvoorbeeld gebruiken de auteurs een bidirectioneel “long short‐term memory” (LSTM) netwerk (een krachtig soort recurrent neuraal netwerk) om de uitspraak van een woord te voorspellen. DOEL Het doel van dit eindwerk is de implementatie en training van een G2P systeem gebaseerd op een neuraal netwerk: dit behelst het onderzoek van verschillende soorten modellen en parameterinstellingen en het testen van de gegenereerde uitspraken voor onze spraakherkenner. Dank zij de aanwezigheid van modulaire en gemakkelijk aanpasbare bibliotheken zoals Tensorflow, Keras, Caffe, Torch, … moet het neuraal netwerk niet vanaf nul gebouwd worden. [1] Demuynck, K., Laureys, T. and S. Gillis. 2002. "Automatic generation of phonetic transcriptions for large speech corpora". In Proc. ICSLP, Vol. 1. [2] Rao, K., Peng, F., Sak, H., Beaufays, F. 2015. "Grapheme‐to‐phoneme conversion using long short‐ term memory recurrent neural networks". In ICASSP. [3] Yao, K. and G. Zweig. 2015. "Sequence‐to‐Sequence Neural Net Models for Grapheme‐to‐ Phoneme Conversion". In Interspeech 2015. UITWERKING Een mogelijke toepassing van dit nieuw G2P systeem is een spraakinterface voor de KU Leuven “wie‐ is‐wie”, wat bijzonder interssant is omdat namen extra moelijk zijn voor een G2P systeem, of een interface voor de Nederlandse woordenlijst (gekend als het “Groene Boekje”), die de gebruiker toelaten om een naam (wie‐is‐wie) of woord (Groene Boekje) op te zoeken met de stem. Eventueel is het ook mogelijk om een webinterface te implementeren voor deze stembevragingen, afhankelijk van de interesse van de student(en). PROFIEL (bv. eerder theoretisch/eerder implementatie, vereiste voorkennis (bv. gevolgde of te volgen vakken, methodieken,programmeertalen, ... ))
Literatuur en studie – 20%, Implementatie – 40%, Experimenten – 40% 1 of 2 studenten
MASTER’s THESIS MASTER of SCIENCE in MATHEMATICAL ENGINEERING TITLE Building a state‐of‐the‐art speech recogniser GUIDANCE For more information, please contact:
[email protected] Promotor: Johan Suykens Co‐promotor
[email protected] Supervisor: CONTEXT (e.g. Information about the problem definition, research project of which the master’s thesis is part)
In the last decade there has been a major shift in machine learning. In the past, machine learning relied heavily on specialized algorithms that were created by experts. These algorithms were very different depending on the field. Someone working on images had a very hard time understanding speech recognition and the other way around. However, in a human brain this is entirely different. We can see that although the brain has different regions each responsible for a certain task, all those regions work in a very similar way and rely on the same simple building block, the neuron. Today, almost all machine learning fields are being dominated by (artificial) neural networks. The idea of a neural network is very simple: create a huge network of interconnected very simple building blocks and let the network learn whatever task you want it to perform. Also in speech recognition there has been a shift to neural networks and companies like Google, Apple, Microsoft, Facebook and Amazon are using neural networks in many of their products. The evolution of the use of neural networks in speech recognition is happening inside out. This means that in the beginning the preprocessing of the speech (signal processing) and similarly the postprocessing (language modeling) still relied on algorithms created by humans. Now, neural networks are being used for bigger and bigger parts of the recognition and the machine learning algorithms are moving to end‐to‐end systems where everything is done with a single big neural network. GOAL The purpose of this thesis will be just that, create an end‐to‐end speech recognition system where as little as possible is designed by hand and as much as possible is learned by the neural network. This sort of networks maps an input sequence to an output sequence. This means that a spectogram is given as input to the system (or possibly some other feature representation) and text characters come out. RESULT Near the end of 2015 Google made the tools that they use to create their neural networks open source (Tensorflow). We have already gained experience with Tensorflow by using it to create a more traditional speech recogniser. These tools will be used to design the end‐to‐end speech recognition system. We will start with a simple task of phoneme recognition in which an input spectrogram is mapped to a sequence of symbolic representations of the speech sounds it contains. Later, a model for word order can be added to map to word sequences. PROFILE (e.g. rather theoretical / rather practical implementation, foreknowledge (courses, methods, computer language(s) et cetera))
Literature and study ‐ 20%, Implementation ‐ 45%, Testing and analysis ‐ 35% 1 or 2 studenten
MASTERPROEF MASTER WISKUNDIGE INGENIEURSTECHNIEKEN TITEL De bouw van een state‐of‐the‐art spraakherkenner BEGELEIDING Voor meer informatie contacteer:
[email protected] Promotor Johan Suykens Co‐promotor
[email protected] Begeleider(s): CONTEXT (Info over het probleem, ev. info over het onderzoeksproject waarin deze masterproef kadert) In de laatste jaren is er een grote verschuiving opgetreden in de machine learning. Vroeger steunde machine learning vooral op gespecialiseerde algoritmen die door experts werden ontwikkeld. Deze algoritmen waren verschillend voor elk domein. Een specialist in beeldverwerking had moeite om spraakherkenning te begrijpen en vice‐versa. In het menselijk brein gaat het er evenwel anders aan toe. Hoewel ons brein verschillende gebieden heeft die elk instaan voor een bepaalde taak, werken al deze gebieden op een gelijkaardige manier en zijn zij gebaseerd op hetzelfde eenvoudige basisblok, het neuron. Momenteel worden bijna alle machine learning domeinen gedomineerd door (artificiële) neurale netwerken. Het idee erachter is zeer eenvoudig: bouw een enorm netwerk van zeer eenvoudige basisblokken die met elkaar verbonden zijn en laat dit netwerk zelf de taak leren die men wenst uit te voeren. Ook in de spraakherkenning is deze verschuiving naar neurale netwerken opgetreden en bedrijven zoals Google, Apple, Microsoft, Facebook en Amazon gebruiken neurale netwerken in vele van hun producten. De evolutie van het gebruik van neurale netwerken is in spraakherkenning van binnen uit ontstaan. Dit betekent dat in het begin de voorverwerking van de spraak (signaalverwerking) en de nabewerking (taalmodellering) nog altijd steunden op algoritmen gemaakt door experten. Tegenwoordig worden neurale netwerken echter gebruikt voor steeds grotere delen van de spraakherkenning en de machine learning algoritmen schuiven op naar end‐to‐end systemen waarbij alles wordt uitgevoerd door één enkel groot neural netwerk. DOEL Het doel van het eindwerk is net dat: de bouw van een end‐to‐end spraakherkenningssysteem waarbij zo weinig mogelijk handwerk komt kijken en zoveel mogelijk geleerd wordt door het neuraal netwerk. Een dergelijk netwerk mapt een input sequentie op een output sequentie. Dat wil zeggen dat een spectrogram wordt aangeboden aan de ingang van het systeem (of eventueel een andere mogelijke kenmerkenvoorstelling), en dat er aan de uitgang tekstkarakters verschijnen. UITWERKING Einde 2015 heeft Google zijn tools voor de bouw van hun neurale netwerken vrijgegeven (Tensorflow). We hebben reeds ervaring opgebouwd met Tensorflow bij de bouw van een meer traditionele spraakherkenner. Deze tools zullen gebruikt worden om het end‐to‐end spraakherkenningssysteem te bouwen. We starten met de eenvoudige taak van foneemherkenning waarbij een spectrogram aan de ingang wordt gemapt op een sequentie van symbolische voorstellingen van de klanken die het bevat. Daarna kan een model voor woordvolgorde worden toegevoegd om woordsequenties te mappen. PROFIEL (bv. eerder theoretisch/eerder implementatie, vereiste voorkennis (bv. gevolgde of te volgen vakken, methodieken,programmeertalen, ... ))
Literatuur en studie – 20%, Implementatie – 45%, Testen en analyse – 35% 1 of 2 studenten