ETD

• control barrier functions
• machine learning
• neural networks
• reinforcement learning
• safe reinforcement learning
• soft actor critic
• uncertainty estimation
• unity ml toolkit

This thesis is situated in the field of machine learning, with a specific focus on reinforcement learning and its extension towards safe reinforcement learning. The work investigates how uncertainty estimation of the policy can be combined with control barrier functions to improve the reliability of policies learned through neural networks. The approach builds on the Soft Actor-Critic algorithm as the underlying reinforcement learning framework, enhancing it with an uncertainty-aware safety mechanism where control barrier functions are activated only when the uncertainty of the reinforcement learning policy is high, thus balancing efficiency and safety. The experimental validation is carried out using the Unity ML Toolkit, in combination with Python, which provides a flexible simulation environment for testing learning agents in complex and potentially unsafe scenarios.