Development and Analysis of Evolutionary Algorithms to Enhance Out-of-Distribution Detection in Autonomous Driving Systems
Abstract: In the development of autonomous vehicles, the ability to reliably recognize unexpected and unknown situations is crucial. This thesis focuses on Out-of-Distribution (OOD) detection, a critical element in ensuring the safety of AI-driven systems. OOD detection involves identifying input data that was not considered during the training of a Deep Learning model, thus posing a risk of incorrect decisions. The goal is to improve the detection rates of OOD situations through the mining of rules sets with genetic or memetic algorithms, thereby enhancing the reliability of autonomous vehicles.
Background and Motivation: The increasing integration of AI systems into safety-critical applications such as autonomous vehicles necessitates advanced methods to ensure their reliability and safety. In particular, OOD detection plays a pivotal role as it enables the system to recognize unknown data during operation and respond appropriately. Current Deep Learning models are limited in their ability to handle data not seen during training. This thesis addresses this gap by exploring how evolutionary algorithms can be used to generate rule sets that improve OOD detection.
Research Objective: The primary goal of this thesis is to develop and evaluate a system based on evolutionary algorithms to generate efficient and robust rule sets for OOD detection in autonomous driving systems. The research will examine how these algorithms can be adapted and optimized to enhance the performance and explainability of OOD detection.
Methodology:
- Introduction to existing Deep Learning techniques and their limitations in OOD detection.
- Development of an evolutionary algorithm for generating rule sets that enable OOD input detection in a Neuro-Symbolic framework
- Application of these rule sets in a simulated autonomous driving environment to assess their effectiveness.
- Comparison of results with traditional Deep Learning approaches to OOD detection.
Student Opportunities: Participating in this thesis provides a unique opportunity for students to engage deeply with cutting-edge technologies in artificial intelligence and autonomous systems. Students will gain hands-on experience with evolutionary algorithms and deep learning techniques, enhancing their skills in both developing and implementing advanced computational methods. This project also offers the chance to contribute to significant advancements in the safety and efficiency of autonomous vehicles, positioning students well for future careers in AI, machine learning, and automotive technology.
Practical Significance: The findings of this research could have far-reaching implications for the safety and efficiency of autonomous vehicles and provide significant insights into the development of future AI-based systems.
If you are interested in this project, feed free to contact Konstantin Kirchheim.
