Topics for a Master’s thesis

Master's Thesis: Learning AI Control Policies with Deep Reinforcement Learning

Background and Motivation

As AI agents become increasingly capable and autonomous, ensuring their safe and reliable behavior becomes an important challenge. AI Control investigates how potentially misaligned AI agents can be supervised and constrained to prevent harmful actions, even when their behavior cannot be fully trusted.

Thesis Description

The goal of this thesis is to investigate whether Deep Reinforcement Learning (DRL) can be used to learn effective AI control policies. Specifically, an DRL-based supervisor should learn when to trust, monitor, or intervene in the actions of a potentially misaligned AI agent while balancing safety, task performance, and supervision costs.

The student will develop and evaluate DRL-based control strategies in a simulated environment and compare them against established rule-based approaches. Potential research directions include constrained reinforcement learning, resource-efficient monitoring, and robustness against adversarial agent behavior.

Required Skills

• Experience with Deep Reinforcement Learning (e.g., through completing a relevant university lecture).

• Good Python programming skills.

Keywords: Deep Reinforcement Learning, AI Alignment, AI Control, Safe Reinforcement Learning, Adversarial Agents