Reinforcement learning on continuous and deterministic systems

Researcher

Hans Harder

Publications

• Hans Harder, Simon Jantsch, Christel Baier, Clemens Dubslaff (2023): A Unifying Formal Approach to Importance Values in Boolean Functions Full Publication
• Hans Harder, Sebastian Peitz (2024): On the continuity and smoothness of the value function in reinforcement learning and optimal control Full Publication

Collaboration

Tristan Kenneweg, Thorben Markmann, Michiel Straat

Research Theme

R2 Prosilience & Robustness

Tags

Reinforcement Learning

My research focuses on data-efficient methods in machine learning and reinforcement-learning, in particular for continuous and deterministic dynamical systems. Under investigation are different ways to achieve data sparsity: For example, by exploiting symmetries (such as translation invariance); learning surrogate models (in order to reduce the number of interactions with the real system); or by exploiting the deterministic nature of such systems.

I’m currently looking at variance bounds in the context of reinforcement learning for “near deterministic” dynamical systems. I’m also investigating optimal (i.e. variance minimizing) step-sizes in policy evaluation procedures