Reinforcement learning control of atom cooling

We apply reinforcement learning to the preparation of a rubidium-87 ultracold quantum gas to realize a consistent and large number of atoms at microkelvin temperatures. This reinforcement learning agent determines an optimal set of thirty experimental control parameters in a dynamically changing environment that is characterized by thirty experimentally sensed parameters. By comparing this method to that of training supervised-learning regression models, as well as to human-driven control schemes, we find that both machine learning approaches accurately predict the number of cooled atoms and both result in occasional superhuman control schemes. However, only the reinforcement learning method achieves consistent outcomes in the laboratory, even in the presence of a dynamic environment.