Adversarial reinforcement learning for safety-critical scenarios

Context

The research area of machine learning is experiencing a renaissance due to hardware acceleration. One use case in the area of scenario-based testing is the identification and parameterization of relevant scenarios. Technical implementations of scenario-based testing approaches often rely on groups of these parameters, which are individually adapted in a time-consuming or labor-intensive way. Reinforcement learning methods offer the possibility to learn such parameterizations through interaction even in complex environments. Based on state-of-the-art reinforcement learning algorithms, this thesis will investigate an approach and develop a method to generate safety-critical scenarios with a hostile agent. Furthermore, another agent will learn strategies to solve these scenarios.

Tasks

• Familiarization with the theory of Adversarial Reinforcement Learning
• Literature review on the topic of Adversarial Generation
• Implementation of selected approaches in Python, using PyTorch for RL algorithms
• Examination, comparison and processing of the results as well as documentation

Prerequisites

• Enthusiasm for the field of machine learning
• Basic knowledge in Python or comparable programming languages
• Independent thinking and working
• Very good knowledge of German or English
• Motivation, willingness to perform and commitment