- Multiobjective Optimization im evolutionären Machine Learning (Bachelor- oder Masterarbeit, Ansprechpartner: Michael Heider)
- Benchmarking evolutionäres Machine Learning / Pittsburgh LCS (Bachelor- oder Masterarbeit / FM / PM, Ansprechpartner: Michael Heider)
- XAI Techniken zur gezielteren Steuerung von Optimierern im evolutionären Machine Learning (Ansprechpartner: Michael Heider)
- Compaction algorithms in XCS(F) (Bachelor- oder Masterarbeit, Ansprechpartner: Michael Heider)
- Benchmarking / extending the Heuristic Evolutionary Rule Optimization System (Bachelor- oder Masterarbeit, Ansprechpartner: Michael Heider)
- Die Erforschung von Reinforcement-Learning-Methoden im autonomen Fahren stützt sich stark auf Simulationsumgebungen wie
highway-envim Gymnasium-Framework, welche standardmäßig ein eher amerikanisch geprägtes Fahrverhalten abbilden. Ziel dieser Arbeit ist es, die Simulationslogik um spezifisch deutsche Verkehrsregeln und Verhaltensmuster – wie das Rechtsfahrgebot oder das Bilden einer Rettungsgasse – zu erweitern. Parallel dazu soll eine grafische Benutzeroberfläche (GUI) entwickelt werden, mit der sich Simulationsparameter (z. B. Fahrzeugdichte und Spurgeschwindigkeiten) dynamisch anpassen lassen. Zudem soll die GUI es ermöglichen, reale Verkehrsszenarien aus bestehenden Autobahn-Datensätzen exakt nachzustellen, um diese direkt in die modifizierte Umgebung zu laden. Das Gesamtsystem wird abschließend anhand dieser Szenarien evaluiert und auf seine Praxistauglichkeit für das Agententraining untersucht. (Projektmodul und/oder Masterarbeit, Ansprechpartner: Marco Hüller)
- Automated Bug Fixing and Feature Generation using Genetic Improvement (Masterarbeit, Ansprechpartner: Michael Heider)
- Automating the design of anytime optimization algorithms / metaheuristics with (Push) genetic programming (Forschungs- oder Projektmodul, Bachelor- oder Masterarbeit, Ansprechpartner: Jonathan Wurth)
- Large-scale anytime benchmarking of state-of-the-art algorithms for single-objective black-box optimization (Forschungs- oder Projektmodul, Bachelor- oder Masterarbeit, Ansprechpartner: Jonathan Wurth)
- Gradient-free versus gradient-based methods for maximizing acquisition functions in Bayesian optimization (Forschungs- oder Projektmodul, Bachelor- oder Masterarbeit, Ansprechpartner: Jonathan Wurth)
- Designing real-world optimization benchmarks for automated algorithm design, for example using engineering design, energy systems, or aerospace simulation frameworks (Projektmodul oder Masterarbeit, Ansprechpartner: Jonathan Wurth)
- An experimental comparison of population diversity measures in metaheuristic optimisation algorithms (Forschungsmodul/Bachelorarbeit; Kenntnisse in Python oder Rust notwendig, Ansprechpartnerin:
Helena Stegherr)
- The population diversity during the search process is an important behaviour measure for metaheuristic optimisation algorithms that provides indications for diversification or intensification of the search. Over the years, a variety of diversity measures has been proposed, all with different specific goals, strengths and weaknesses. While many of these have been theoretically explored, the impact of using different measures in experimental settings has only been quantified in small parts.
Starting points:
https://ieeexplore.ieee.org/document/10611897/
http://ieeexplore.ieee.org/document/6151099/
- The population diversity during the search process is an important behaviour measure for metaheuristic optimisation algorithms that provides indications for diversification or intensification of the search. Over the years, a variety of diversity measures has been proposed, all with different specific goals, strengths and weaknesses. While many of these have been theoretically explored, the impact of using different measures in experimental settings has only been quantified in small parts.
- Self-organising maps for clustering metaheuristic behaviour (time series) data (Forschungsmodul/Bachelorarbeit oder Masterarbeit; Kenntnisse in Python notwendig, Ansprechpartnerin:
Helena Stegherr)
- The behaviour of metaheuristic algorithms during the search process can be described as a time series of specific behaviour measure values. However, in large experimental setups, comparing those to identify similarities or differences can quickly become infeasible without the use of appropriate methods, e.g. unsupervised learning techniques such as clustering or dimensionality reduction. Self-organising maps are among those approaches and have been successfully employed to cluster time series data in other contexts.
Starting points:
https://linkinghub.elsevier.com/retrieve/pii/S0893608012002596
http://arxiv.org/abs/2108.11523
- The behaviour of metaheuristic algorithms during the search process can be described as a time series of specific behaviour measure values. However, in large experimental setups, comparing those to identify similarities or differences can quickly become infeasible without the use of appropriate methods, e.g. unsupervised learning techniques such as clustering or dimensionality reduction. Self-organising maps are among those approaches and have been successfully employed to cluster time series data in other contexts.
- Expanding upon Structural Bias analysis of metaheuristic optimisation algorithms (Forschungsmodul/Projektmodul; kann bei vielversprechenden Ergebnissen auf eine Bachelorarbeit/Masterarbeit erweitert werden; Kenntnisse in Python oder R notwendig; Ansprechpartnerin:
Helena Stegherr)
- Structural Bias describes the tendency of metaheuristic optimisation algorithms to focus on a specific area of the search space without any obvious (performance-related) reason. It is often only measures at the end of the search process, on a specific bias test function, and only for utilising the information of the best found solution. However, there are other possibilities to increase upon our understanding of structural bias that should be explored.
Starting points:
http://arxiv.org/abs/1408.5350
https://ieeexplore.ieee.org/document/9828803 - Next to the investigation of additional structural bias examination techniques, visualisation of bias could be enhanced. This could range from investigating specialised visualisation techniques to interactive approaches to visualisation.
- Structural Bias describes the tendency of metaheuristic optimisation algorithms to focus on a specific area of the search space without any obvious (performance-related) reason. It is often only measures at the end of the search process, on a specific bias test function, and only for utilising the information of the best found solution. However, there are other possibilities to increase upon our understanding of structural bias that should be explored.
Predictive Maintenance (PdM) nutzt Datenanalyse und Machine Learning, um den Zustand von Maschinen kontinuierlich zu überwachen und Ausfälle vorherzusagen. Ziel ist es, durch rechtzeitige Wartungsmaßnahmen unerwartete Stillstände zu vermeiden, die Lebensdauer von Maschinen zu verlängern und die Effizienz zu steigern. In einer Abschlussarbeit können aktuell offene Fragestellungen im Forschungsfeld bearbeitet werden, z.B. durch die praxisnahe Integration von PdM oder das Testen moderner Machine-Learning Modelle. (Projektmodul oder Abschlussarbeit Bachelor und Master; Ansprechpartner: Lukas Meitz)
)
Einige unserer diversen Praxispartner haben auch oft kurzfristig spannende Themen. Fragen Sie hierzu gerne nach. Aktueller konkret ausgeschrieben ist:
- Deep Learning Framework Evaluation for AI Combustion Modeling bei everllence (ehemals MAN ES) in Augsburg: https://jobs.everllence.com/job/Augsburg-Abschlussarbeit-Deep-Learning-Framework-Evaluation-for-AI-Combustion-Modeling-86153/1253469601/