Summary of the dissertation

Reinforcement learning (RL) is a promising method of machine learning in which an agent learns control strategies exclusively through interaction with its environment.

This doctoral thesis investigated the use of RL in large-scale systems with a high repetition rate. To demonstrate this concept, the KINGFISHER system, based on the novel AMD/Xilinx Versal platform, was designed and implemented to support experimental deployments.

It was deployed at KIT's KARA synchrotron, where it successfully learned to control beam instabilities without any prior knowledge - making it the first fully online hardware-trained system at a particle accelerator.

We congratulate Luca Scomparin on this great achievement!