Mirco Theile

TU Munich and UC Berkeley

Monday, October 23, 2023, 11:30 - 12:30

SR 01-012

Reinforcement Learning is an increasingly popular method being applied to optimization in various fields. Particularly, Deep Reinforcement Learning (DRL) promises to solve many challenging and hard-to-model problems by learning generalizing policies through interactions with the environment. Cyber-physical systems (CPS) pose many challenges for DRL agents, including hard safety constraints and stringent timing requirements. This talk focuses on the projects of my Ph.D. that formulate various problems in CPS for DRL. The problems include planning, control, and scheduling, and the presented methods enable safety guarantees through modeling, predictable timing through a distributed training architecture, and limited generalization to unseen problems.