Speaker
Weijia Wang
(Karlsruhe Institute of Technology (KIT))
Description
The KIT (Karlsruhe Institute of Technology) storage ring KARA (Karlsruhe Research Accelerator) aims at the stabilization of the emitted THz radiation to tackle the problem
of micro-bunching instability. A reinforcement learning based method are planed to be implemented through a RF feedback system located at storage ring. IPE (Institute of Data Processing and Electronics) developed a reinforcement learning hardware platfrom for the final realistic approach in this paper. To provide a proof of concept, a CartPole problem environment are built on a ZYNQ MPSoC platfrom, a fast Neural Network inferance are deployed on FPGA, and a Lite-Weighted training process are developed on ARM. The problem is solved by Policy Gradient method, showing the same functionality with simulation on standard PC and improve dramatically the inferance and training speed. The platform could also used by other reinforcement learning based applications senario.
Primary authors
Prof.
Marc Weber
(KIT)
Mr
Matthias Balzer
(KIT)
Dr
Michele Caselle
(KIT)
Mr
Nick Karcher
(KIT - IPE)
Dr
Oliver Sander
(KIT)
Weijia Wang
(Karlsruhe Institute of Technology (KIT))