Beam Control with Fast Reinforcement Learning Inference at the Edge

This thesis delves into real time beam control, at the KARA accelerator. A key focus is the use of reinforcement learning at the edge with the Versal VCK190 evaluation board, emphasizing heterogeneous computing and its architectural advantages. It explores the implementation of neural networks on the AI Engines and their application for machine control. Specifically, the implementation of the Gated Recurrent Unit (GRU) is detailed, addressing motivation, computation pipeline, and technical challenges such as memory constraints and activation functions.