Verifying Deep Reinforcement Learning Systems

Next Wednesday May 15, 2024:

Organized by: FAU DCN-AvH, Chair for Dynamics, Control, Machine Learning and Numerics – Alexander von Humboldt Professorship at FAU, Friedrich-Alexander-Universität Erlangen-Nürnberg (Germany)
Title: Verifying Deep Reinforcement Learning Systems
Speaker: Guy Amir
Affiliation: The Hebrew University of Jerusalem

Abstract. Deep neural networks (DNNs) have gained significant popularity in recent years, becoming the state of the art in a variety of domains. In particular, deep reinforcement learning (DRL) has recently been employed to train DNNs that realize control policies for various types of real-world systems. In this work, we present recent advances made for formally verifying complex properties of DRL systems, both from the theoretical perspective, as well as the applicability of our approach to real-world robotic navigation platforms.

The talk will be mostly based on two papers:
Towards Scalable Verification of Deep Reinforcement Learning (FMCAD 2021)
Verifying Learning-Based Robotic Navigation Systems (TACAS 2023)