Marc Aurel Vischer: The Lottery Ticket Hypothesis in Reinforcement Learning

BCCN Berlin / Technische Universität Berlin

 

Abstract

Recent research in deep learning has demonstrated the existence of very sparse neural networks that train to performance levels comparable to those of their dense counterparts. These results challenge the role of overparameterization and generalization in supervised learning.

But do these insights transfer to agents which simultaneously perceive and act to maximize reward signals? This thesis establishes the existence of such ”winning tickets” in a visual navigation task for both on- and off-policy reinforcement learning.

We provide evidence that the resulting compressed observation space provides minimal task-relevant information which facilitates learning.

 

Additional Information

Master Thesis Defense

 

Organized by

Prof. Dr. Henning Sprekeler   & Prof. Dr. Klaus Obermayer   / Lisa Velenosi