A model's response for a given input pattern depends on the seen patterns in the training data. The larger the amount of training data, the more likely edge cases are covered during training. However, the more complex input patterns are, the larger the model has to be. For very simple use cases, a relatively small model can achieve very high test accuracy in a matter of minutes. On the other hand, a large model has to be trained for multiple days. The actual time to develop a model of that size can be considered to be even greater since often many different architecture types and hyper-parameter configurations have to be tried. An extreme case for a large model is the recently released GPT-3 model. This model consists of 175 billion parameters and was trained using 45 terabytes of text data. The model was trained to generate text and is able to write news articles and source code based only on a rough description. However, a model like this is only creatable for researchers with access to special hardware or immense amounts of data. Thus, it is desirable to find less resource-intensive training approaches to enable other researchers to create well performing models. This thesis investigates the use of pre-trained models. If a model has been trained on one dataset and is then trained on another similar data, it faster learns to adjust to similar patterns than a model that has not yet seen any of the task's pattern. Thus, the learned lessons from one training are transferred to another task. During pre-training, the model is trained to solve a specific task like predicting the next word in a sequence or first encoding an input image before decoding it. Such models contain an encoder and a decoder part. When transferring that model to another task, parts of the model's layers will be removed. As a result, having to discard fewer weights results in faster training since less time has to be spent on training parts of a model that are only needed to solve an auxiliary task. Throughout this thesis, the concept of siamese architectures will be discussed since when using that architecture, no parameters have to be discarded when transferring a model trained with that approach onto another task. Thus, the siamese pre-training approach positively impacts the need for resources like time and energy use and drives the development of new models in the direction of Green AI. The models trained with this approach will be evaluated by comparing them to models trained with other pre-training approaches as well as large existing models. It will be shown that the models trained for the tasks in this thesis perform as good as externally pre-trained models, given the right choice of data and training targets: It will be shown that the number and type of training targets during pre-training impacts a model's performance on transfer learning tasks. The use cases presented in this thesis cover different data from different domains to show that the siamese training approach is widely applicable. Consequently, researchers are motivated to create their own pre-trained models for data domains, for which there are no existing pre-trained models. / Die Vorhersage eines Models hängt davon ab, welche Muster in den während des Trainings benutzen Daten vorhanden sind. Je größer die Menge an Trainingsdaten ist, desto wahrscheinlicher ist es, dass Grenzfälle in den Daten vorkommen. Je größer jedoch die Anzahl der zu lernenden Mustern ist, desto größer muss jedoch das Modell sein. Für einfache Anwendungsfälle ist es möglich ein kleines Modell in wenigen Minuten zu trainieren um bereits gute Ergebnisse auf Testdaten zu erhalten. Für komplexe Anwendungsfälle kann ein dementsprechend großes Modell jedoch bis zu mehrere Tage benötigen um ausreichend gut zu sein. Ein Extremfall für ein großes Modell ist das kürzlich veröffentlichte Modell mit dem Namen GPT-3, welches aus 175 Milliarden Parametern besteht und mit Trainingsdaten in der Größenordnung von 45 Terabyte trainiert wurde. Das Modell wurde trainiert Text zu generieren und ist in der Lage Nachrichtenartikel zu generieren, basierend auf einer groben Ausgangsbeschreibung. Solch ein Modell können nur solche Forscher entwickeln, die Zugang zu entsprechender Hardware und Datenmengen haben. Es demnach von Interesse Trainingsvorgehen dahingehend zu verbessern, dass auch mit wenig vorhandenen Ressourcen Modelle für komplexe Anwendungsfälle trainiert werden können. Diese Arbeit beschäfigt sich mit dem Vortrainieren von neuronalen Netzen. Wenn ein neuronales Netz auf einem Datensatz trainiert wurde und dann auf einem zweiten Datensatz weiter trainiert wird, lernt es die Merkmale des zweiten Datensatzes schneller, da es nicht von Grund auf Muster lernen muss sondern auf bereits gelerntes zurückgreifen kann. Man spricht dann davon, dass das Wissen transferiert wird. Während des Vortrainierens bekommt ein Modell häufig eine Aufgabe wie zum Beispiel, im Fall von Bilddaten, die Trainingsdaten erst zu komprimieren und dann wieder herzustellen. Bei Textdaten könnte ein Modell vortrainiert werden, indem es einen Satz als Eingabe erhält und dann den nächsten Satz aus dem Quelldokument vorhersagen muss. Solche Modelle bestehen dementsprechend aus einem Encoder und einem Decoder. Der Nachteil bei diesem Vorgehen ist, dass der Decoder lediglich für das Vortrainieren benötigt wird und für den späteren Anwendungsfall nur der Encoder benötigt wird. Zentraler Bestandteil in dieser Arbeit ist deswegen das Untersuchen der Vorteile und Nachteile der siamesische Modellarchitektur. Diese Architektur besteht lediglich aus einem Encoder, was dazu führt, dass das Vortrainieren kostengünstiger ist, da weniger Gewichte trainiert werden müssen. Der wesentliche wissenschaftliche Beitrag liegt darin, dass die siamische Architektur ausführlich verglichen wird mit vergleichbaren Ansätzen. Dabei werden bestimmte Nachteile gefunden, wie zum Beispiel dass die Auswahl einer Ähnlichkeitsfunktion oder das Zusammenstellen der Trainingsdaten große Auswirkung auf das Modelltraining haben. Es wird erarbeitet, welche Ähnlichkeitsfunktion in welchen Kontexten empfohlen wird sowie wie andere Nachteile der siamischen Architektur durch die Anpassung der Trainingsziele ausgeglichen werden können. Die entsprechenden Experimente werden dabei auf Daten aus unterschiedlichen Domänen ausgeführt um zu zeigen, dass der entsprechende Ansatz universell anwendbar ist. Die Ergebnisse aus konkreten Anwendungsfällen zeigen außerdem, dass die innerhalb dieser Arbeit entwickelten Modelle ähnlich gut abschneiden wie extern verfügbare Modelle, welche mit großem Ressourcenaufwand trainiert worden sind. Dies zeigt, dass mit Bedacht erarbeitete Architekturen die benötigten Ressourcen verringern können.

No description available.

The study of phylogeny, i.e. the evolutionary history of species, is a central problem in biology and a key for understanding characteristics of contemporary species. Many problems in this area can be formulated as combinatorial optimisation problems which makes it particularly interesting for computer scientists. The reconstruction of the phylogeny of species can be based on various kinds of data, e.g. morphological properties or characteristics of the genetic information of the species. Maximum parsimony is a popular and widely used method for phylogenetic reconstruction aiming for an explanation of the observed data requiring the least evolutionary changes. A certain property of the genetic information gained much interest for the reconstruction of phylogeny in recent time: the organisation of the genomes of species, i.e. the arrangement of the genes on the chromosomes. But the idea to reconstruct phylogenetic information from gene arrangements has a long history. In Dobzhansky and Sturtevant (1938) it was already pointed out that “a comparison of the different gene arrangements in the same chromosome may, in certain cases, throw light on the historical relationships of these structures, and consequently on the history of the species as a whole”. This kind of data is promising for the study of deep evolutionary relationships because gene arrangements are believed to evolve slowly (Rokas and Holland, 2000). This seems to be the case especially for mitochondrial genomes which are available for a wide range of species (Boore, 1999). The development of methods for the reconstruction of phylogeny from gene arrangement data has made considerable progress during the last years. Prominent examples are the computation of parsimonious evolutionary scenarios, i.e. a shortest sequence of rearrangements transforming one arrangement of genes into another or the length of such a minimal scenario (Hannenhalli and Pevzner, 1995b; Sankoff, 1992; Watterson et al., 1982); the reconstruction of parsimonious phylogenetic trees from gene arrangement data (Bader et al., 2008; Bernt et al., 2007b; Bourque and Pevzner, 2002; Moret et al., 2002a); or the computation of the similarities of gene arrangements (Bergeron et al., 2008a; Heber et al., 2009). 1 1 Introduction The central theme of this work is to provide efficient algorithms for modified versions of fundamental genome rearrangement problems using more plausible rearrangement models. Two types of modified rearrangement models are explored. The first type is to restrict the set of allowed rearrangements as follows. It can be observed that certain groups of genes are preserved during evolution. This may be caused by functional constraints which prevented the destruction (Lathe et al., 2000; Sémon and Duret, 2006; Xie et al., 2003), certain properties of the rearrangements which shaped the gene orders (Eisen et al., 2000; Sankoff, 2002; Tillier and Collins, 2000), or just because no destructive rearrangement happened since the speciation of the gene orders. It can be assumed that gene groups, found in all studied gene orders, are not acquired independently. Accordingly, these gene groups should be preserved in plausible reconstructions of the course of evolution, in particular the gene groups should be present in the reconstructed putative ancestral gene orders. This can be achieved by restricting the set of rearrangements, which are allowed for the reconstruction, to those which preserve the gene groups of the given gene orders. Since it is difficult to determine functionally what a gene group is, it has been proposed to consider common combinatorial structures of the gene orders as gene groups (Marcotte et al., 1999; Overbeek et al., 1999). The second considered modification of the rearrangement model is extending the set of allowed rearrangement types. Different types of rearrangement operations have shuffled the gene orders during evolution. It should be attempted to use the same set of rearrangement operations for the reconstruction otherwise distorted or even wrong phylogenetic conclusions may be obtained in the worst case. Both possibilities have been considered for certain rearrangement problems before. Restricted sets of allowed rearrangements have been used successfully for the computation of parsimonious rearrangement scenarios consisting of inversions only where the gene groups are identified as common intervals (Bérard et al., 2007; Figeac and Varré, 2004). Extending the set of allowed rearrangement operations is a delicate task. On the one hand it is unknown which rearrangements have to be regarded because this is part of the phylogeny to be discovered. On the other hand, efficient exact rearrangement methods including several operations are still rare, in particular when transpositions should be included. For example, the problem to compute shortest rearrangement scenarios including transpositions is still of unknown computational complexity. Currently, only efficient approximation algorithms are known (e.g. Bader and Ohlebusch, 2007; Elias and Hartman, 2006). Two problems have been studied with respect to one or even both of these possibilities in the scope of this work. The first one is the inversion median problem. Given the gene orders of some taxa, this problem asks for potential ancestral gene orders such that the corresponding inversion scenario is parsimonious, i.e. has a minimum length. Solving this problem is an essential component 2 of algorithms for computing phylogenetic trees from gene arrangements (Bourque and Pevzner, 2002; Moret et al., 2002a, 2001). The unconstrained inversion median problem is NP-hard (Caprara, 2003). In Chapter 3 the inversion median problem is studied under the additional constraint to preserve gene groups of the input gene orders. Common intervals, i.e. sets of genes that appear consecutively in the gene orders, are used for modelling gene groups. The problem of finding such ancestral gene orders is called the preserving inversion median problem. Already the problem of finding a shortest inversion scenario for two gene orders is NP-hard (Figeac and Varré, 2004). Mitochondrial gene orders are a rich source for phylogenetic investigations because they are known for more than 1 000 species. Four rearrangement operations are reported at least in the literature to be relevant for the study of mitochondrial gene order evolution (Boore, 1999): That is inversions, transpositions, inverse transpositions, and tandem duplication random loss (TDRL). Efficient methods for a plausible reconstruction of genome rearrangements for mitochondrial gene orders using all four operations are presented in Chapter 4. An important rearrangement operation, in particular for the study of mitochondrial gene orders, is the tandem duplication random loss operation (e.g. Boore, 2000; Mauro et al., 2006). This rearrangement duplicates a part of a gene order followed by the random loss of one of the redundant copies of each gene. The gene order is rearranged depending on which copy is lost. This rearrangement should be regarded for reconstructing phylogeny from gene order data. But the properties of this rearrangement operation have rarely been studied (Bouvel and Rossin, 2009; Chaudhuri et al., 2006). The combinatorial properties of the TDRL operation are studied in Chapter 5. The enumeration and counting of sorting TDRLs, that is TDRL operations reducing the distance, is studied in particular. Closed formulas for computing the number of sorting TDRLs and methods for the enumeration are presented. Furthermore, TDRLs are one of the operations considered in Chapter 4. An interesting property of this rearrangement, distinguishing it from other rearrangements, is its asymmetry. That is the effects of a single TDRL can (in the most cases) not be reversed with a single TDRL. The use of this property for phylogeny reconstruction is studied in Section 4.3. This thesis is structured as follows. The existing approaches obeying similar types of modified rearrangement models as well as important concepts and computational methods to related problems are reviewed in Chapter 2. The combinatorial structures of gene orders that have been proposed for identifying gene groups, in particular common intervals, as well as the computational approaches for their computation are reviewed in Section 2.2. Approaches for computing parsimonious pairwise rearrangement scenarios are outlined in Section 2.3. Methods for the computation genome rearrangement scenarios obeying biologically motivated constraints, as introduced above, are detailed in Section 2.4. The approaches for the inversion median problem are covered in Section 2.5. Methods for the reconstruction of phylogenetic trees from gene arrangement data are briefly outlined in Section 2.6.3 1 Introduction Chapter 3 introduces the new algorithms CIP, ECIP, and TCIP for solving the preserving inversion median problem. The efficiency of the algorithm is empirically studied for simulated as well as mitochondrial data. The description of algorithms CIP and ECIP is based on Bernt et al. (2006b). TCIP has been described in Bernt et al. (2007a, 2008b). But the theoretical foundation of TCIP is extended significantly within this work in order to allow for more than three input permutations. Gene order rearrangement methods that have been developed for the reconstruction of the phylogeny of mitochondrial gene orders are presented in the fourth chapter. The presented algorithm CREx computes rearrangement scenarios for pairs of gene orders. CREx regards the four types of rearrangement operations which are important for mitochondrial gene orders. Based on CREx the algorithm TreeREx for assigning rearrangement events to a given tree is developed. The quality of the CREx reconstructions is analysed in a large empirical study for simulated gene orders. The results of TreeREx are analysed for several mitochondrial data sets. Algorithms CREx and TreeREx have been published in Bernt et al. (2008a, 2007c). The analysis of the mitochondrial gene orders of Echinodermata was included in Perseke et al. (2008). Additionally, a new and simple method is presented to explore the potential of the CREx method. The new method is applied to the complete mitochondrial data set. The problem of enumerating and counting sorting TDRLs is studied in Chapter 5. The theoretical results are covered to a large extent by Bernt et al. (2009b). The missing combinatorial explanation for some of the presented formulas is given here for the first time. Therefor, a new method for the enumeration and counting of sorting TDRLs has been developed (Bernt et al., 2009a).

Although not all hazardous effects on human health have been clearly defined so far, the health risks of particulate matter can be considered evident. Pulmonary and cardiovascular diseases, in particular, are caused or aggravated by inhaled particulate matter. The aim of this article is to describe the incorporation and the effects on organ function of inhaled particles. Furthermore, the potential risks of de novo synthesised nanoparticles are discussed in the context of the public controversy regarding environmental particulate matter pollution. / Obwohl die schädlichen Auswirkungen inhalierbarer Partikel auf unseren Organismus bisher noch nicht vollständig geklärt sind, kann eine Gesundheitsgefährdung durch Feinstäube als erwiesen angesehen werden. Insbesondere pulmonale und kardiovaskuläre Erkrankungen werden durch Feinstaubexposition ausgelöst oder verschlimmert. Dieser Artikel stellt Aufnahme und Auswirkungen inhalierter Partikel im menschlichen Organismus dar und erörtert potenzielle Gefahren de novo synthetisierter Nanopartikel im Kontext der auch in der breiten Öffentlichkeit kontrovers geführten Feinstaubdiskussion.

During the last decades, mitochondria and their DNA have become a hot topic of research due to their essential roles which are necessary for cells survival and pathology. In this study, multiple methods have been developed to help with the understanding of mitochondrial DNA and its evolution. These methods tackle two essential problems in this area: the accurate annotation of protein-coding genes and mitochondrial genome rearrangements. Mitochondrial genome sequences are published nowadays with increasing pace, which creates the need for accurate and fast annotation tools that do not require manual intervention. In this work, an automated pipeline for fast de-novo annotation of mitochondrial protein-coding genes is implemented. The pipeline includes methods for enhancing multiple sequence alignment, detecting frameshifts and building protein profiles guided by phylogeny. The methods are tested on animal mitogenomes available in RefSeq, the comparison with reference annotations highlights the high quality of the produced annotations. Furthermore, the frameshift method predicted a large number of frameshifts, many of which were unknown. Additionally, an efficient partially-local alignment method to investigate genomic rearrangements in mitochondrial genomes is presented in this study. The method is novel and introduces a partially-local dynamic programming algorithm on three sequences around the breakpoint region. Unlike the existing methods which study the rearrangement at the genes order level, this method allows to investigate the rearrangement on the molecular level with nucleotides precision. The algorithm is tested on both artificial data and real mitochondrial genomic sequences. Surprisingly, a large fraction of rearrangements involve the duplication of local sequences. Since the implemented approach only requires relatively short parts of genomic sequence around a breakpoint, it should be applicable to non-mitochondrial studies as well.

Am 10. November 2022, 16 Uhr, begann die diesjährige digitale StuFoExpo 2022 – die fünfte Ausstellung studentischer Forschungsprojekte an der TU Dresden. Paul Druschke, ein ehemaliger StuFo-Teilnehmer, übernahm die Rolle des Moderators und führte galant und charmant durch den Nachmittag bzw. Abend. Rektorin Frau Prof. Ursula Staudinger eröffnete die Veranstaltung mit einem Grußwort und zeigte sich von den engagierten Studierenden an der Technischen Universität Dresden begeistert. Außerdem betonte sie die wachsenden Herausforderungen unseres Planeten und die damit einhergehende Bedeutung, dass junge Leute in ihrer Forschung Verantwortung dafür übernehmen und vielfältige lösungsorientierte Perspektiven eröffnen. An diese Perspektiven anknüpfendend, stellte Dr. Franziska Schulze-Stocker das Programm FOSTER – Funds for Student Research vor, ein Programm zur Förderung studentischer Forschung an der TU Dresden. Anschließend folgte einer der Höhepunkte des Abends: die sechzehn Videopitches. Die 90-sekündigen Kurzvideos wurden von den Teilnehmenden als Ergänzung zu den Postern vorbereitet und sollten Neugierde und Interesse am eigenen Forschungsprojekt wecken. An Kreativität und Einfallsvermögen mangelte es den Teilnehmenden nicht und die Resultate sind spannend und interessant anzuschauen. Wer neugierig auf die Projekte und deren Darstellung geworden ist, kann sich hier Pitches, Poster und Inhalte der Projekte ansehen.

Humans can quickly adapt their behavior to changes in the environment. Classical reversal learning tasks mainly measure how well participants can disengage from a previously successful behavior but not how alternative responses are explored. Here, we propose a novel 5-choice reversal learning task with alternating position-reward contingencies to study exploration behavior after a reversal. We compare human exploratory saccade behavior with a prediction obtained from a neuro-computational model of the basal ganglia. A new synaptic plasticity rule for learning the connectivity between the subthalamic nucleus (STN) and external globus pallidus (GPe) results in exploration biases to previously rewarded positions. The model simulations and human data both show that during experimental experience exploration becomes limited to only those positions that have been rewarded in the past. Our study demonstrates how quite complex behavior may result from a simple sub-circuit within the basal ganglia pathways.

It is widely recognised that knowledge and highly skilled individuals as "carriers" of knowledge (i.e. knowledge spillover agents) play a key role in impelling the development and growth of cities and regions. In this paper we discuss the relation between the mobility of talent and knowledge flows. In this context, several issues are examined, including the role of highly skilled labour for regional development, the features that characterise knowledge spillovers through labour mobility, the key factors for attracting and retaining talent as well as the rise of "brain gain" policies. Although the paper deals with highly skilled mobility and migration in general, a particular attention will be paid to flows of (star) scientists. / Series: SRE - Discussion Papers

Large volumes of data are collected today in many domains. Often, there is so much data available, that it is difficult to identify the relevant pieces of information. Knowledge discovery seeks to obtain novel, interesting and useful information from large datasets. One key technique for that purpose is subgroup discovery. It aims at identifying descriptions for subsets of the data, which have an interesting distribution with respect to a predefined target concept. This work improves the efficiency and effectiveness of subgroup discovery in different directions. For efficient exhaustive subgroup discovery, algorithmic improvements are proposed for three important variations of the standard setting: First, novel optimistic estimate bounds are derived for subgroup discovery with numeric target concepts. These allow for skipping the evaluation of large parts of the search space without influencing the results. Additionally, necessary adaptations to data structures for this setting are discussed. Second, for exceptional model mining, that is, subgroup discovery with a model over multiple attributes as target concept, a generic extension of the well-known FP-tree data structure is introduced. The modified data structure stores intermediate condensed data representations, which depend on the chosen model class, in the nodes of the trees. This allows the application for many popular model classes. Third, subgroup discovery with generalization-aware measures is investigated. These interestingness measures compare the target share or mean value in the subgroup with the respective maximum value in all its generalizations. For this setting, a novel method for deriving optimistic estimates is proposed. In contrast to previous approaches, the novel measures are not exclusively based on the anti-monotonicity of instance coverage, but also takes the difference of coverage between the subgroup and its generalizations into account. In all three areas, the advances lead to runtime improvements of more than an order of magnitude. The second part of the contributions focuses on the \emph{effectiveness} of subgroup discovery. These improvements aim to identify more interesting subgroups in practical applications. For that purpose, the concept of expectation-driven subgroup discovery is introduced as a new family of interestingness measures. It computes the score of a subgroup based on the difference between the actual target share and the target share that could be expected given the statistics for the separate influence factors that are combined to describe the subgroup. In doing so, previously undetected interesting subgroups are discovered, while other, partially redundant findings are suppressed. Furthermore, this work also approaches practical issues of subgroup discovery: In that direction, the VIKAMINE II tool is presented, which extends its predecessor with a rebuild user interface, novel algorithms for automatic discovery, new interactive mining techniques, as well novel options for result presentation and introspection. Finally, some real-world applications are described that utilized the presented techniques. These include the identification of influence factors on the success and satisfaction of university students and the description of locations using tagging data of geo-referenced images. / Neue Techniken für effiziente und effektive Subgruppenentdeckung

Nowadays, data centers are becoming increasingly dynamic due to the common adoption of virtualization technologies. Systems can scale their capacity on demand by growing and shrinking their resources dynamically based on the current load. However, the complexity and performance of modern data centers is influenced not only by the software architecture, middleware, and computing resources, but also by network virtualization, network protocols, network services, and configuration. The field of network virtualization is not as mature as server virtualization and there are multiple competing approaches and technologies. Performance modeling and prediction techniques provide a powerful tool to analyze the performance of modern data centers. However, given the wide variety of network virtualization approaches, no common approach exists for modeling and evaluating the performance of virtualized networks. The performance community has proposed multiple formalisms and models for evaluating the performance of infrastructures based on different network virtualization technologies. The existing performance models can be divided into two main categories: coarse-grained analytical models and highly-detailed simulation models. Analytical performance models are normally defined at a high level of abstraction and thus they abstract many details of the real network and therefore have limited predictive power. On the other hand, simulation models are normally focused on a selected networking technology and take into account many specific performance influencing factors, resulting in detailed models that are tightly bound to a given technology, infrastructure setup, or to a given protocol stack. Existing models are inflexible, that means, they provide a single solution method without providing means for the user to influence the solution accuracy and solution overhead. To allow for flexibility in the performance prediction, the user is required to build multiple different performance models obtaining multiple performance predictions. Each performance prediction may then have different focus, different performance metrics, prediction accuracy, and solving time. The goal of this thesis is to develop a modeling approach that does not require the user to have experience in any of the applied performance modeling formalisms. The approach offers the flexibility in the modeling and analysis by balancing between: (a) generic character and low overhead of coarse-grained analytical models, and (b) the more detailed simulation models with higher prediction accuracy. The contributions of this thesis intersect with technologies and research areas, such as: software engineering, model-driven software development, domain-specific modeling, performance modeling and prediction, networking and data center networks, network virtualization, Software-Defined Networking (SDN), Network Function Virtualization (NFV). The main contributions of this thesis compose the Descartes Network Infrastructure (DNI) approach and include: • Novel modeling abstractions for virtualized network infrastructures. This includes two meta-models that define modeling languages for modeling data center network performance. The DNI and miniDNI meta-models provide means for representing network infrastructures at two different abstraction levels. Regardless of which variant of the DNI meta-model is used, the modeling language provides generic modeling elements allowing to describe the majority of existing and future network technologies, while at the same time abstracting factors that have low influence on the overall performance. I focus on SDN and NFV as examples of modern virtualization technologies. • Network deployment meta-model—an interface between DNI and other meta- models that allows to define mapping between DNI and other descriptive models. The integration with other domain-specific models allows capturing behaviors that are not reflected in the DNI model, for example, software bottlenecks, server virtualization, and middleware overheads. • Flexible model solving with model transformations. The transformations enable solving a DNI model by transforming it into a predictive model. The model transformations vary in size and complexity depending on the amount of data abstracted in the transformation process and provided to the solver. In this thesis, I contribute six transformations that transform DNI models into various predictive models based on the following modeling formalisms: (a) OMNeT++ simulation, (b) Queueing Petri Nets (QPNs), (c) Layered Queueing Networks (LQNs). For each of these formalisms, multiple predictive models are generated (e.g., models with different level of detail): (a) two for OMNeT++, (b) two for QPNs, (c) two for LQNs. Some predictive models can be solved using multiple alternative solvers resulting in up to ten different automated solving methods for a single DNI model. • A model extraction method that supports the modeler in the modeling process by automatically prefilling the DNI model with the network traffic data. The contributed traffic profile abstraction and optimization method provides a trade-off by balancing between the size and the level of detail of the extracted profiles. • A method for selecting feasible solving methods for a DNI model. The method proposes a set of solvers based on trade-off analysis characterizing each transformation with respect to various parameters such as its specific limitations, expected prediction accuracy, expected run-time, required resources in terms of CPU and memory consumption, and scalability. • An evaluation of the approach in the context of two realistic systems. I evaluate the approach with focus on such factors like: prediction of network capacity and interface throughput, applicability, flexibility in trading-off between prediction accuracy and solving time. Despite not focusing on the maximization of the prediction accuracy, I demonstrate that in the majority of cases, the prediction error is low—up to 20% for uncalibrated models and up to 10% for calibrated models depending on the solving technique. In summary, this thesis presents the first approach to flexible run-time performance prediction in data center networks, including network based on SDN. It provides ability to flexibly balance between performance prediction accuracy and solving overhead. The approach provides the following key benefits: • It is possible to predict the impact of changes in the data center network on the performance. The changes include: changes in network topology, hardware configuration, traffic load, and applications deployment. • DNI can successfully model and predict the performance of multiple different of network infrastructures including proactive SDN scenarios. • The prediction process is flexible, that is, it provides balance between the granularity of the predictive models and the solving time. The decreased prediction accuracy is usually rewarded with savings of the solving time and consumption of resources required for solving. • The users are enabled to conduct performance analysis using multiple different prediction methods without requiring the expertise and experience in each of the modeling formalisms. The components of the DNI approach can be also applied to scenarios that are not considered in this thesis. The approach is generalizable and applicable for the following examples: (a) networks outside of data centers may be analyzed with DNI as long as the background traffic profile is known; (b) uncalibrated DNI models may serve as a basis for design-time performance analysis; (c) the method for extracting and compacting of traffic profiles may be used for other, non-network workloads as well. / Durch Virtualisierung werden moderne Rechenzentren immer dynamischer. Systeme sind in der Lage ihre Kapazität hoch und runter zu skalieren , um die ankommende Last zu bedienen. Die Komplexität der modernen Systeme in Rechenzentren wird nicht nur von der Softwarearchitektur, Middleware und Rechenressourcen sondern auch von der Netzwerkvirtualisierung beeinflusst. Netzwerkvirtualisierung ist noch nicht so ausgereift wie die Virtualisierung von Rechenressourcen und es existieren derzeit unterschiedliche Netzwerkvirtualisierungstechnologien. Man kann aber keine der Technologien als Standardvirtualisierung für Netzwerke bezeichnen. Die Auswahl von Ansätzen durch Performanzanalyse von Netzwerken stellt eine Herausforderung dar, weil existierende Ansätze sich mehrheitlich auf einzelne Virtualisierungstechniken fokussieren und es keinen universellen Ansatz für Performanzanalyse gibt, der alle Techniken in Betracht nimmt. Die Forschungsgemeinschaft bietet verschiedene Performanzmodelle und Formalismen für Evaluierung der Performanz von virtualisierten Netzwerken an. Die bekannten Ansätze können in zwei Gruppen aufgegliedert werden: Grobetaillierte analytische Modelle und feindetaillierte Simulationsmodelle. Die analytischen Performanzmodelle abstrahieren viele Details und liefern daher nur beschränkt nutzbare Performanzvorhersagen. Auf der anderen Seite fokussiert sich die Gruppe der simulationsbasierenden Modelle auf bestimmte Teile des Systems (z.B. Protokoll, Typ von Switches) und ignoriert dadurch das große Bild der Systemlandschaft. ...