Courses
Winter Term 2016/2017
- Vorlesung (Lecture): Reinforcement Learning / Optimierendes Lernen
- Praktikum (Lab course): Deep Learning Course, Control Section
- Oberseminar (Advanced Seminar, Reading Club): Advanced Topics in Machine Learning
- Masterprojekt: Decoding speech-related brain signals with recurrent neural networks
- Masterprojekt: Kernel approximation for large-scale Gaussian Processes
- Masterprojekt: Model Learning for locally optimal control of a humanoid robot
Previous Lectures
For a complete list of previous lectures and student projects please click here.
Finished Thesis Projects
Download links for PDFs are only available to users who are logged in. They can also be sent to visitors via email on request; please contact one of the supervisors if you are interested.
Spotlights of Recent Theses
| Daniel Stratz: Khepera III Navigation mit internen Sensoren In dieser Arbeit soll eine Implementierung vorgestellt werden, die den Khepera möglichst genau entlang der Wand einer Testumgebung fahren lässt und hierbei die Ecken der Testumgebung erkennt und die Position dieser bestimmt. Bachelor Thesis Supervisors: Martin Riedmiller, Manuel Blum |
| Jan Wülfing: Feature Learning for Music Classification In the area of music information retrieval, automatic categorization of music pieces into categories such as mood, artist or genre is a widely studied topic. Those categorization taks basically consist of two steps, feature selection/extraction and classification. In the past, a wide range of features have been proposed, most of them hand designed. More recently, there has been a growing interest in methods that learn those features from data. In his masters thesis, Jan works on a feature learning method that has been successfully used for object recognition in images, and investigates the applicability of this method for genre classification of music pieces. Master Thesis Supervisors: Martin Riedmiller |
 | Tobias Springenberg: Feature Learning
using Temporal Coherence In his thesis Tobias develops feature learning algorithms that capture the relation between input samples. Most current feature learning algorithms were developed for object recognition in still images and are thus unable to capture this temporal information. This problem becomes apparent in many applications for which feature learning might be desireable such as: human motion and action recognition in video data, tracking and stereo vision. Apart from these applications features learned using temporal information can also be used to improve object recognition performance under complex image transformations. Master Thesis Supervisors: Martin Riedmiller |
 | Hadi Kheyruri: Ball detection for RoboCup Hadi develops computer vision algorithms, that are able to detect an arbitrary soccer ball in a RoboCup setting. Previously RoboCup balls were limited to a specific color and texture which is known before a competition. In this work Hadi tries to expand the state-of-the-art techniques in two ways. First by the collection of a well suited data-set, to perform as a benchmark for other researchers interested in investigating ball detection. Second part of this work focuses on the development and implementation of detection algorithms. Master Thesis Supervisors: Martin Riedmiller, Thomas Brox |
| Jeremias Holub: Approximate Dynamic Programming with a Fuzzy Parameterization In this thesis, Jeremias implements and analyzes Fuzzy Q-iteration, a Q-iteration algorithm with a fuzzy parameterization of the state space and a discretization of the action space. A Fuzzy Q-iteration controller is implemented using the open source software CLSquare. The performance is tested within a two-link manipulator control problem and is compared with other state-of-the-art reinforcement learning algorithms. Bachelor Thesis Supervisors: Manuel Blum, Martin Riedmiller |
 | Jan Mattner: Application of Visual Deep Learning to a Real-World System Jan's thesis focuses on the problem of delays and inaccuracies which arises when controlling a dynamic system on basis of raw camera images. An augmented state space is used to overcome these difficulties. His experiments empirically demonstrate learning on automatically constructed feature spaces to be in the same order of complexity as learning on low-dimensional sensory data. Bachelor Thesis, finished 10/2011 Supervisors: Sascha Lange, Martin Riedmiller |
 | Manuel Watter: Object Tracking using Reinforcement Learning with a Pan-Tilt Camera This thesis explores the challenge to focus and follow a moving object in a actuated Pan-Tilt Camera. An algorithm was developed to center a general object in the camera and to learn an augmentation add-on controller to be able to follow once this object starts moving on a predefined path. Bachelor Thesis, finished 9/2011 Supervisors: Stefan Welker, Martin Riedmiller |
 | Nikolaus Mayer: Globale Explorationsstrategien (German) This thesis analyzes the effects of globally directed exploration on a learning agent's performance. We introduce the BELT growing-batch algorithm which, for the first time, employs a learning procedure to create an independent policy for exploration itself. Compared to undirected and locally directed exploration, we find global exploration to accelerate the learning of good policies. Bachelor Thesis, finished 09/2011 Supervisors: Sascha Lange, Martin Riedmiller |
 | Arne Voigtländer: Using Visual Deep Reinforcement Learning to Control a Real-World Dynamical System In his thesis, Arne applies our ground-breaking Deep Fitted Q Algorithm (DFQ) to a real-world dynamical system. DFQ combines batch reinforcement learning methods with deep neural networks in order to solve visual reinforcement learning tasks. With his thesis Arne has demonstrated that it is possible to learn near-optimal control policies from raw camera data just by trial and error. Master Thesis, finished 07/2010 Supervisors: Sascha Lange, Martin Riedmiller |
Full List of Theses
| Student | Title | Type | Completion |
| Christopher Schröder | Sentimentanalyse deutschsprachiger Twitter-Nachrichten | Bachelor | 2014 |
| Florian Geißer | General Game Playing under Uncertainty | Master | 2013 |
| Markus Grützner | Detektion von Multis in MMO durch Analyse von Handelsaktivitäten | Master | 2013 |
| Max Schumacher | A study of DFQ on playing pong | Bachelor | 2013 |
| Denis Stier | Semantic Attachments for Decision-Theoretic Planning | Master | 2013 |
| Jonas Delleske | Verbesserung des client-side Caching der Spielkarte im Browserspiel Wack-a-doo mit Techniken des Machine Learnings | Bachelor | 2013 |
| Daniel Stratz | Khepera III Navigation mit internen Sensoren | Bachelor | 2013 |
| Tatiana Jimenez Cardenas | Training Pose Specific Detectors | Master | 2013 |
| Mesbahuddin Anwari | Learning binding preferences of RNA-binding proteins using in vitro affinities and in vivo binding sites | Master | 2012 |
| Stefan Band | Ein KI-Ansatz für Strategiespiele | Bachelor | 2012 |
| Urban Himmelsbach | Entwicklung eines autonom fahrenden Fahrzeugs | Master | 2012 |
| Jost Tobias Springenberg | Feature Learning using Temporal Coherence | Master | 2012 |
| Jan Manuel Wülfing | Unsupervised Feature Learning for Music Classification | Master | 2012 |
| Martin Gloderer | Optimizing Sequences of Learned Movement Primitives Using Data Gathered from the Demonstrations | Master | 2012 |
| Hadi Kheyruri | Ball Detection Based on Histogram of Features on an own collected Dataset for Soccer Robots | Master | 2012 |
| Jeremias Holub | Implementierung eines approximativen Q-Iterationsalgorithmus mit einer unscharfen Parametrisierung | Bachelor | 2012 |
| Dragos Alexandru Sorescu | De Novo Molecular Design Using Graph Kernels | Master | 2012 |
| Johannes Andreas Stork | Learning Social Navigation Behaviors with Dynamic Features and Inverse Reinforcement Learning | Master | 2012 |
| Jens Silva Santisteban | Socially Acceptable Robot Navigation: A Learning Approach | Master | 2012 |
| Kiran Kumar Telukunta | Secondary structure motif determination in ncRNA via graph kernel based computational models | Master | 2012 |
| Julian Schmid | Kontrollieren von Einheiten in Starcraft mit Reinforcement Lernen | Bachelor | 2012 |
| Jörg Holschuh | Zustandskodierung für Reinforcement Learning mittels neuronaler Netze | Bachelor | 2012 |
| Johannes Aldinger | Algorithms for the Canadian Traveler´s Problem with Remote Sensing | Master | 2011 |
| Jan Peter Mattner | Application of Visual Deep Learning to a Real-World System | Bachelor | 2011 |
| Manuel Meyer | Überwachtes Lernen in pixelbasierter Objekterkennung | Bachelor | 2011 |
| Manuel Watter | Objektverfolgung mittels Reinforcement Learning am bewegten Kamerakopf | Bachelor | 2011 |
| Nikolaus Mayer | Globale Explorationsstrategien | Bachelor | 2011 |
| Li Zhang | RNA-Protein Interaction Prediction with Graph Kernels | Master | 2011 |
| Elmar Haussmann | Contextual Sentence Decomposition with Applications to Semantic Full-Text Search | Master | 2011 |
| Thomas Huber | Evaluation einer dateneffizienten neuronalen Lernmethode am Katana Roboterarm | Master | 2011 |
| Thorsten Engesser | Generalisierendes Neural Fitted Q Learning im TORCS-Competition-Framework | Bachelor | 2011 |
| Benedikt Mendorf | Analyse und Gegenüberstellung von selbstlernender neuronaler Reglern und PID-Reglern im pneumatischen Positionierungssystem | Bachelor | 2011 |
| Philipp Blohm | Maschinelle Lernverfahren in Information Retrieval | Master | 2011 |
| Oliver Mitevski | Semantic Indexing in Information Retrieval | Master | 2011 |
| Alexandru-Dorin Oprea | Recommender System for Research Papers | Master | 2010 |
| Roxana Dolores Bersan | Plan-based Control of Characters in Video Games | Master | 2010 |
| Norman Kohler | Bewegungsplanung für autonome Luftschiffe | Diploma | 2010 |
| Denis Stier | Einsatz von Reinforcment Learning zur Geschwindigkeitsregulierung in mikroskopischer Verkehrsimulation | Bachelor | 2010 |
| Christian Lutz | Autonom lernende Agenten in Computerspielen: Anwendung von Reinforcement-Lernverfahren in Star Ships Learning Framework | Bachelor | 2010 |
| Stefan Keller | Priorisierung von Lernzielen im Nebenfach Medizin für Informatiker | Diploma | 2010 |
| Manuel Blum | Merkmalsextraktion bei Klassifikationsproblemen | Master | 2010 |
| Thomas Lampe | Visual Servoing durch Neuronales Reinforcement Learning | Master | 2010 |
| Andreas Witsch (Uni Kassel) | Policy Gradient Methods | Diploma | 2010 |
| Jonas Gehring | Intelligente Objekterkennung für ein lernfähiges Carrerabahn-System | Bachelor | 2009 |
| Nathan Franklin | Vision Based Navigation | Master | 2009 |
| Hannes Schulz | Euclidean Embedding of Co-Proven Queries | Master | 2009 |
| Matthias Laquai | Intelligente Steuerung autonomer Carrera-Rennwagen mittels Lernverfahren | Diploma | 2009 |
| Jost Tobias Springenberg | Machine Learning on Massively Parallel Architectures - A Case Study | Bachelor | 2009 |
| Philipp Blohm | Beurteilung von Stärken und Schwächen von neuroevolutionären Verfahren zum Einsatz im Bereich des Reinforcement Learnings | Bachelor | 2008 |
| Manuel Blum | Klassifikation von Bilddaten mittels rekurrenter neuronaler Netze zur Steuerung eines Torwarts im Roboterfußball | Bachelor | 2008 |
| Christian Müller | Reinforcement-Lernen mit Pac-Man | Bachelor | 2007 |
| Sebastian Schnock | Implementation und Evaluation von Neuroevolution in CLS | Bachelor | 2007 |
| Hussain Iqbal | Evolutionary Algorithm using neuro-evolutionary techniques in dynamic environment | Master | 2007 |
| Michael Wiegand (FH Karlsruhe) | Vorhersage der Bewegung eines autonomen Roboters | Diploma | 2007 |
| Tim Kitzmann | View-Based Object Recognition: A Biologically Motivated Approach using iGRLVQ | Bachelor | 2006 |
| Andrea Albrecht | Erkennung und Klassifizierung von Objekten mit Neuronalen Netzen | Bachelor | 2007 |
| Srinivas Gudipally | Survey: Approaches of Artificial Intelligence with the focus on Machine Learning within the RoboCup domain | Master | 2006 |
| Andreas G. Nie | RoboPECS - Autonome Fußballfans im RoboCup Entwicklung eines Multiagentensystems zur Simulation von Fußballfans | Magister | 2005 |