Fakultätspreise der KIT-Fakultät für Wirtschaftswissenschaften
Preisträger 2022
Stefan Schwarze: Solution Methods for Discrete Nash Equilibrium Problems
Über die Masterarbeit: We study non-cooperative n-person games, where each player aims to maximize her payoff given the other players' strategies. Applied to competing firms in a market, the equilibrium points are interesting for the firms as well as for a central authority that may pose rules. Up to now, most algorithmic approaches for the computation of Nash equilibria require continuous strategy sets for all players. However, integer decisions appear in many real-world applications. We present a branch-and-prune procedure for discrete Nash equilibrium problems with a convex description of each player’s strategy set. The derived pruning criterion reduces the search space effectively by stating the activity of certain constraints. This results in a synchronous branching and pruning method.
1-Minütiges Erklärvideo über die Masterarbeit
Lukas Struppek: Embedding Convolutional Mixture of Experts into Deep Neural Networks for Computer Vision Tasks
Über die Masterarbeit: Das Mixture of Experts (MoE) Konzept basiert auf der Idee, für verschiedene Aufgaben verschiedene Experten-Modelle einzusetzen, welche dynamisch durch ein sogenanntes Gating-Network ausgewählt werden. Die vorliegende Arbeit bettet das MoE-Konzept in Form generischer MoE-Schichten in tiefe Convolutional Neural Networks zur Bildverarbeitung ein. Es wird gezeigt, dass selbst mit einem einzelnen Datensatz und Ende-zu-Ende Lernen – welches traditionelle MoE-Ansätze nicht unterstützen – eine implizite Spezialisierung der einzelnen Experten auf individuelle Domänen stattfindet. Dies bietet zusätzliche Einblicke in den Entscheidungsprozess innerhalb von Neuronalen Netzen. Zudem lässt sich die Anzahl der aktiven Experten dynamisch festlegen und entsprechend der verfügbaren Rechenleistung skalieren.
Kevin Wiegratz: Machine Learning Methods in Financial Economics: Recent Applications, Prospects, and the Valuation of Real Estate Assets
Über die Masterarbeit: Artificial intelligence is increasingly entering our day-to-day life with impressive applications such as face detection or voice recognition. The main technology behind artificial intelligence is machine learning (ML). Given the power of ML, this thesis investigates where and how to apply ML in financial economics. First, the thesis identifies the methodological differences of ML compared to traditional econometric approaches. Then, it develops a taxonomy of existing and future applications of ML in financial economics. Finally, it applies ML to a typical problem in financial economics: the pricing of real estate assets. The results show a significant improvement of ML over traditional methods, which demonstrates the strong potential of ML for future applications in financial economics.
Preisträgerinnen und Preisträger 2021
Patrick Deininger: Monte Carlo Tree Search for Production Scheduling in Matrix Production
Über die Masterarbeit: One challenge of production scheduling for matrix productions is the near real time generation of good machine schedules, a NP-hard problem. Patrick Deininger has developed a central, holistic decision instance, which is based on multi-objective Monte Carlo Tree Search (MCTS) and capable of finding pareto solutions with respect to individual preferences in near real-time. The thesis considers two modifications of MCTS. (1) The dynamic adaption of exploration and exploitation improves local optima. (2) The learning of promising combination rules for action generation facilitates adaptability to changing optimization contexts and improves the generalizability, search speed and solution quality significantly.
1-Minütiges Erklärvideo über die Masterarbeit
Patrick Hemmer: DEAL: Deep Evidential Active Learning for Image Classification
Über die Masterarbeit: Convolutional Neural Networks (CNNs) have proven to be state-of-the-art models for supervised computer vision tasks. However, large labeled data sets are generally needed for model training. Unlabeled data is available in many domains, but labeling is expensive, e.g., when specific expert knowledge is required. In this thesis, a novel Active Learning algorithm that learns from unlabeled data by capturing high prediction uncertainty is proposed to mitigate the problem of extensive data labeling. We demonstrate with publicly available data and in the context of a quality assurance industry use case that our approach leads to a significant reduction in the number of labeled images necessary to achieve a predefined performance contributing to a considerable decrease in labeling overhead.
1-Minütiges Erklärvideo über die Masterarbeit
Jingyi Lisa Zhang:Overcoming the Cold Start Problem in AI Technology - A Survey-based Approach on Example of Wearable Sleep Tracker AURA
Über die Masterarbeit: Wearable sleep technology such as AURA represents a novel technology that promises to deliver the most accurate sleep measurement as well as sleeping improvement suggestions and thus generates benefits for customers in need. However, there is a critical time period between users’ first interaction and the first valid piece of sleep-improving suggestion the algorithm can generate due to development of statistical significance which hence, leads to unclear acceptance of customers. After relevant case study, this work aims to explore various value perception and expectation of possible segments in terms of their preferences in sleep wearable area, in particular regarding the critical initial phase (“cold start”). First, the experimental study reveals users’ implicit valuations via conjoint analysis. Attributes of Design, measured sleep parameter and frequency of necessary input are the most important in a nutshell. The study also provides empirical evidence for the existence of three dissimilar clusters, that is, design preferred lifestylers, indifferent average followers and detail-oriented functional desirers. The findings imply different strategies of recommended course of action regarding the relevance of the cold start problem. Finally, implications are discussed and future research options are sketched out.