In der vorliegenden Masterarbeit wurde die Problematik der Reichweitenvorhersage von Artikeln der Internetseite "der WELT" analysiert und eine Lösung entwickelt, die Redakteuren eine reichweitenoptimierte Kuratierung ermöglicht. Hierzu wurde die Hypothese eines linearen Zusammenhangs zwischen gleichzeitigen Nutzern in den ersten 30 Minuten und der Reichweite nach 24 Stunden aufgestellt. Für die Überprüfung und Optimierung wurden vier Algorithmen zur linearen Regression ausgewählt, umgesetzt und die Modelle verglichen. Aus Artikeldaten und Webtracking-informationen wurden Merkmale erarbeitet, welche Einflüsse auf die Reichweite abbilden. Dabei konnte durch Webtrekking-Merkmale eine lineare Abhängigkeit hergestellt und als Basis für die Vorhersage verwendet werden. Da hohe Reichweiten besonders relevant sind, wurden die besten 10% getrennt betrachtet. Dabei wurden soziale Trends als größter Verbesserungsfaktor festgestellt. Unter anderem wurde außerdem die Ähnlichkeit von Artikeln mit Doc2Vec einbezogen, wobei nur ein geringer Zusammenhang zwischen der Reichweite eines Einzelnen und der Reichweite ähnlicher Artikel identifiziert werden konnte.

In der Umsetzung ist ein Dienst mit austauschbaren Algorithmen und Modellen entstanden. Für den Redakteur wurde die Vorhersage als Reichweitenpotential in das Kuratierungstool integriert.

Applying machine learning techniques and especially artificial neural networks to problems related with quantum mechanical calculations has become more popular recently.

Meanwhile, the development of software frameworks for efficient parallelized computations, like Numpy  and SciPy for the CPU as well as Theano, TensorFlow  or Keras  for a GPU made enormous progress in the last few years.

The main part of this work is devoted to translating the techniques developed in the above mentioned works on applying artificial neural networks to approximate the Potential Energy Surface of an atomic cluster to modern computing frameworks, making use of the possibility to parallelize and hence speed up the computations, using a GPU. First, we test the software developed within this thesis and obtain results comparable to those in  for the approximation of the Potential Energy Surface. However, we will see that the software developed within this thesis is much faster, the factor of speed improvement is about 30 to 70, depending on the given problem and, of course, on the size of the data set.

The next step is to approximate the forces acting on the nuclei of the atoms in the atomic cluster, which was done before by differentiating the Potential Energy Surface with respect to the coordinate axes.

In this thesis, we will take a different approach, not differentiating the Potential Energy Surface but rather directly approximating these forces with an artificial neural network. We will see that this takes more time than approximating the Potential Energy, but because of the speed improvements gained by using the modern and optimized software frameworks like Keras, and because of the usage of a GPU, this approach is still feasible in terms of computational time needed.

The features for the training of an artificial neural network that approximates the Potential Energy Surface have been obtained using so called symmetry functions that have been applied to atoms lying within a certain range of a sphere, yielding features that are invariant under translation and rotation, just as the target output, the Potential Energy, is as well. However, when approximating the forces, the output is not invariant under rotation. But still, rotating the atomic cluster will rotate the forces in the same way. For this reason, we introduce a modified approach, replacing the aforementioned spheres by ellipsoids, more precisely spheres stretched along one of the coordinate axes, to ensure the same behaviour of features and target outputs when rotating the input, namely the atomic cluster. Introducing this change in feature calculation did not improve the result in terms of error measurement by a meaningful amount, but the idea is very new and there is definitely a lot of room for improvement.

It must be said that this is only a first try of directly approximating the forces through the use of an artificial neural network, but it yields promising results and asks for optimization in this direction.

Artificial Intelligence suffers a dichotomy between symbolic and numeric methods. With the advent of Deep Learning and the good performance on several different fields, its applications in context of the automated theorem prover Leo-III are explored. By applying methods from functional programming, this thesis introduces a reconstruction routine for proofs which allows to execute them and subject them for further analysis. Furthermore, the role of clause selection in internal proof-guidance is explored by modeling the proof process as a Markov Decision Process without clause selection and the application of a Markov Decision Process for optimizing the clause  selection is described. In the end, further applications of machine learning techniques in Leo-III are outlined.

Nowadays, real-time capable visual odometry and visual simultaneous localization and mapping have become popular research topics. Since robots depend on the precise determination of their own motion, visual methods can be used for trajectory generation, localization or path planning. Different kinds of sensors can be used to tackle this - in general hard to solve - task, but it is always a trade-off between configuration effort and monetary cost of the system as well as other quality factors. Hence, it becomes increasingly popular to use cameras as sensors for the ego-motionn determination of a robot.

This thesis deals with the extension of a monocular direct sparse visual odometry to a stereo direct sparse visual-inertial odometry and the evaluation of the outcome.

The depth information from a stereo camera is used to eliminate the initialization step and to pre-initialize the depth of selected pixels of keyframes. Furthermore, depth information and inertial measurements significantly robustify pose pre-initialization for new frames. Due to the known depth, the unknown scale issue is solved and the scale drift is eliminated. The experiments carried out in this work show that the extensions significantly increase both robustness and tracking accuracy.

Fahrzeuge mit immer größeren Automatisierungsgrad befinden sich auf dem Vormarsch, jedoch ist die derzeitige Verkehrsplanung noch vollständig auf menschliche Fahrer ausgerichtet.

In naher Zukunft werden moderne Fahrzeuge immer mehr Fahraufgaben in dem von menschlichen Fahrern dominierten Straßenverkehr übernehmen. Das kooperierende Fahren mit anderen Verkehrsteilnehmern stellt hierbei (teil-)autonome Fahrzeuge vor große Herausforderungen, denn viele Autofahrer folgen den Verkehrsregeln nicht mit der strikten Genauigkeit eines Computers.

Dies offenbart sich deutlich bei der Betrachtung von Kreuzungen. Hier finden die oftmals spärlichen Spurmarkierungen häufig nur wenig Beachtung. Eine Steigerung dazu sind große Kreuzungen und Kreisverkehre, bei denen gänzlich auf Spurmarkierungen verzichtet wurde.

In diesen Szenarien müssen sich autonome Fahrzeuge nach dem Fahrverhalten der anderen Verkehrsteilnehmer richten, um sicher und für andere Fahrer vorhersehbar zu agieren. Schwarmbasierte Navigationstechniken können die Entwicklung eines solchen Fahrverhaltens unterstützen.

Im Rahmen dieser Masterarbeit wurde ein Verfahren entwickelt, das eine Schwarm-Kartographierung von Straßenabschnitten, insbesondere Kreuzungen und Kreisverkehren, erlaubt. Bei diesem schwarmbasierten Mapping werden die Verkehrsbewegungen anderer Fahrzeuge erfasst und in einer Karte gespeichert, die eine effiziente Auswertung der Daten erlaubt.

Es wird gezeigt, dass sich mit realen Sensoraufzeichnungen des autonomen Forschungsfahrzeugs "MadeInGermany" präzise Karten aus den Fahrtbewegungen der anderen Fahrzeuge erstellen lassen.

Diese Karten werden einer umfassenden Genauigkeitsanalyse unterzogen. Zudem wird demonstriert, dass sie sich zur Pfadplanung, zur Erkennung von Abweichungen im Verkehrsfluss und zur Vorhersage der Intention anderer Verkehrsteilnehmer eignen.

Autonomous robots attempt to understand their environment in real-time. The data structure PLANT, which this thesis introduces, segments large amounts of data efficiently. To detect homogeneous image regions fast, a PLANT combines an integral image with a binary search. Beyond that, a PLANT-based vision for soccer-playing robots is presented. It enables the robots to perceive their environment with a camera resolution of 1920×1080 pixels at a frame rate of 30 Hz. For superpixel creation, the algorithms _PLANT and _PLANTm are introduced. The time complexities of _PLANT and _PLANTm are O(n + k·log(n)) and O(n + k·log(n) + k^{^}^2), where n is the number of pixels and k the number of superpixels. PLANT-based algorithms benefit from the spatial locality of reference, which results in a high speed-up.

In this  thesis, an adaptive leading behavior model for a biometric robot is proposed to lead a guppy: A series of analyses of fish shoals has been performed to provide the theoretical support for this model. This model uses quantified reactions of fish to the robot as feedback. It is implemented using a state machine. The behavior of this model has been divided into sub-behaviors, while each sub-behavior corresponds to a state of the state machine. The adjustment of the robot's motion is made within most of the state and the transitions of the state dependent on the environment in real time. The results indicate that this model is effective to lead a live guppy.

We present an embedding of the second-order fragment of the Theory of Abstract Objects as described in Edward Zalta's upcoming work Principia Logico-Metaphysica (PLM[12]) in the automated reasoning framework Isabelle/HOL. The Theory of Abstract Objects is a metaphysical theory that reifies property patterns, as they for example occur in the abstract reasoning of mathematics, as abstract objects  and provides an axiomatic framework that allows to reason about these objects. It thereby serves as a fundamental metaphysical theory that can be used to axiomatize and describe a wide range of philosophical objects, such as Platonic forms or Leibniz' concepts, and has the ambition to function as a foundational theory of mathematics. The target theory of our embedding as described in chapters 7-9 of PLM [12] employs a modal relational type theory as logical foundation for which a representation in functional type theory is known to be challenging [8].

Nevertheless we arrive at a functioning representation of the theory in the functional logic of Isabelle/HOL based on a semantical representation of an Aczel-model of the theory. Based on this representation we construct an implementation of the deductive system of PLM ([12, Chap. 9]) which allows to automatically and interactively find and verify theorems of PLM.

Our work thereby supports the concept of shallow semantical embeddings of logical systems in HOL as a universal tool for logical reasoning as promoted by Christoph Benzmüller [1].

The most notable result of the presented work is the discovery of a previously unknown paradox in the formulation of the Theory of Abstract Objects.The embedding of the theory in Isabelle/HOL played a vital part in this discovery. Furthermore it was possible to immediately offer several options to modify the theory to guarantee its consistency. Thereby our work could provide a significant contribution to the development  of a proper grounding for object theory.

Computergestützte Navigation wird in bronchoskopischen Operationen genutzt, um Ärzten die Orientierung im weitverzweigten Bronchialsystem zu erleichtern. Dadurch kann die Zeit eines Eigriffs verkürzt und die Fehlerrate verringert werden. Aktuelle Navigationssysteme berechnen die Pose des Bronchoskops häufig mittels eines elektromagnetischen Sensors, der in den Arbeitskanal des Instruments eingeführt wird. Um die Navigation ohne einen solchen Sensor zu ermöglichen, soll das Bronchoskop visuell unter Verwendung des Kamerabildes lokalisiert werden.

Die Leistungsfähigkeit der vielzähligen Methoden des visuellen Trackings sind in diesem Kontext noch wenig erforscht.  Daher ist das Ziel dieser Masterarbeit, eine Reihe von Algorithmen zum visuellen Tracking in der Lunge quantitativ zu analysieren.

Zu den analysierten Methoden gehören das Keypoint Matching und der optische Fluss. Darüber hinaus wird als wichtigste Anwendung des visuellen Trackings die Odometrie mit den verschiedenen Algorithmen analysiert. Es wurde festgestellt, dass das Keypoint Matching nur weniger als 2% der Features korrekt zuordnen kann. Der optische Fluss hingegen kann eine deutlich höhere Anzahl von Punkten korrekt zuordnen. Die Odometrie mittels des optischen Flusses wurde, mit dem Ergebnis, das Brox und Farnebäck die geeignetsten Algorithmen sind, untersucht.

Zuletzt werden eine Reihe von weiterführenden Verfahren, mit denen das Tracking verbessert werden kann, vorgeschlagen.

In this thesis, the imagery of the fisheye camera system mounted to the autonmous car MadeInGermany of the Freie Universität Berlin is processed and calibrated. Hereby, distortions caused by fisheye lenses are automatically corrected and a surround view of the vehicle is created.

Over the next decade, autonomous cars are expected to dadically change mobility as we know it. While intelligent software systems made astonishing improvements over the past years, the eventual quality of autonomous cars depends on their sensors capturing the environment.

One of the most important sensors are cameras for visual input. Hence, it does not surprise that current autonomous prototypes often have multiple cameras, for example to prevent  having blind spots. For this specific reason, fisheye lenses with a large field of views are often used.

To utilize recordings of these camera systems by computer vision algorithms, a camera calibration is required. It consists of the intrinsic calibration, rectifying possible distortions and extrinsic calibration, detrmining position and pose of the camera.

The waggle dance patterns of the western honey bee are well understood and researched. New methods in computer vision systems allow long-term tracking of individual bees to capture the spatiotemporal position of each bee in the hive. Previous research on the detection of the waggle dance primarily focuses  on videos with a high frame rate. The few currently known methods on spatiotemporal date require a high temporal resolution of at least 13 Hz to capture the waggle behavior part. Contrary to the dancer, research on the detection of waggle dance followers is not existent.

This thesis introduces a new model to detect waggle-dancers and their followers in spatiotemporal data with a low temporal resolution by using domain knowledge  to engineer specific features that match these behaviors. We describe the model of the waggle-dance and their followers, the patterns behind it and the process to utilize this knowledge into specific features.

The proposed model allows the discovery of not only the waggle-dancers  but its followers; the combination of long-term tracking and the detection enables further research into the relation between each bee in the colony.

Im zweiten und dritten Kapitel seines Werkes Proslogion formulierte Sankt Anselm von Canterbury sein berühmtes ontologisches Argument. Ob dieses als Beweis anzusehen ist, bleibt aufgrund der informalen Verschriftlichung fraglich. Eder und Ramharter haben 2015 mehrere formale Rekonstruktionen und einen allgemeinen Anforderungskatalog für solche präsentiert. Die Probleme ihrer Rekonstruktion erörtere ich, erarbeite und verifiziere eine Verbesserung. Den Anforderungskatalog überarbeite ich zweckmäßig. Abschließend diskutiere ich die Ergebnisse und ihre Implikationen.

Ziel der vorliegenden Arbeit ist es, ein leicht anzuwendendes Programm zu entwickeln mit dem die Daten des Versuchs automatisiert aufbereitet und ausgewertet werden können. Ferner soll es eine Modularität aufweisen, so dass es ohne größere Arbeit für ähnliche Versuche wiederverwendet werden kann.

The BeesBook system provides high-resolution data about bee movements within a single colony by automatically tracking individual honey bees inside a hive over their entire life. This thesis focuses on the process of designing and implementing a network pipeline to extract interaction networks from this data. Spatial proximity is used as an indicator for interactions between bees. Social network analysis methods were applied to investigate the static and dynamic properties of the resulting social networks of honey bees on a global, intermediate and local level. The resulting networks were characterized  by a low hierarchical structure and a high density.  The global  structure of the colony seems to be stable over time. The local structure is highly dynamic, as bees change communities as they age. Communities in the honey bee network are formed by age groups that show a high spatial fidelity. The findings are in line with the established state of research that colonies  are organized around age-based task division. The results of the analysis validate the implemented pipeline and the inferred networks. Consequently, this work provides an excellent foundation for future research focusing on temporal network analysis.

In recent years, neural networks have been widely used as the method of choice for classification tasks in computer vision and image processing. The networks that currently provide the best results are very deep residual Convolutional Neural Networks. New network structures, based on the previously used topologies are developed each year, using different approaches. In this thesis the application of a canonical genetic algorithm to optimize the topology of neural networks is investigated. Current state of the art networks are used as a baseline to start the optimization form. The generated architectures achieve an improved classification accuracy, compared to reference networks. The developed network structures differ from commonly used architectures in their higher level architecture. In contrast to the reference networks they are not homogeneously from a sequence of different blocks. The improved classification accuracy suggests that the use of heterogeneous structures may be advantageous.  Furthermore it is shown that the generated neural networks converge in fewer iterations, producing better results.

Das BeesBook Projekt ist ein Forschungsprojekt des Biorobotics Lab an der Freien Universität Berlin und untersucht als Schwerpunkt das Sozialverhalten von Bienen. Seit 2014 versehen Mitarbeiter des Projektes jährlich die Individuen eines Bienenschwarms mit zirkulären Binärcode Markierungen (Tags). Die markierten Bienen werden anschließend in ihrem Stock über mehrere Wochen mit hochauflösenden Kameras im Infrarobereich beobachtet. Eine Pipeline genannte Anwendung bestimmt die Positionen der Tags in den Bilddaten und liest die Tag-Ids aus. Danach sind circa 80% der Zuordnungen von Beobachtungen zu den individuellen Bienen korrekt. Mit Strategien des Multi-Target Trackings ist es möglich, die Erkennungsrate zu erhöhen. In dieser Arbeit wird gezeigt, dass schon ein spartanischer Ansatz die Erkennungsrae um 10% verbessert.  Dafür wird eine Anwendung vorgestellt, die die Vorbereitung, Durchführung und Analyse des Trackings auf den Daten des Bees-Book Projektes ermöglicht. Mit dieser Anwendung wird in einem Proof-of-Concept die Erkennungsrate mit Strategien des Multi-Target Trackings in einem mehrstufigen Verfahren auf über 95% erhöht. Der Vergleich mit eines existierenden Verfahren zeigt die Vorteile des mehrstufigen Trackings bezüglich Qualität und Laufzeit.

Currently, big players in the automotive sector are racing each other for developing the first full autonomous vehicle. The most well-known among them are Tesla Inc., VW AG and even non-automotive companies, such as Alphabet Inc. and Apple Inc. are taking part. However, there are still many pitfalls and major challenges they are faced with. Specifically unexpected situations in everyday road traffic have to be considered. Nissan recently even plans to install a remote call center for providing human assistance for dealing with those siutations and even has lost the hope of eventually developing a full autonomous vehicle one day. One type of these critical siutations is given by road construction sites. A road construction site is characterized by a broad range of unforeseeable temporal changes to the infrastructure that can invalidate any preexisting map information for navigation purposes. Furthermore, an undefined behavior or slow reactions on infrastructure changes can directly affect the road safety or obstruct the traffic-flow. The present work specifically focuses on this problem domain, and examines how deep learning can be a part of a solution towards a fully autonomous handling of those situations. For this purpose, a dataset of hundreds of road construction site images has been created. Only images of construction sites that comply with german rules and regulations and images taken at daytime are considered. Further, a deep convolutional neural network (CNN) has been designed and trained to detect road construction sites as well as the corresponding changes to the actual driving lanes on a pixel level. The final outcome of this work basically demonstrates the powerfulness of deep learning with respect to the domain of construction sites and thus provides a proof of concept for future developments in autonomous driving.