Reinforcement learning in cold atom experiments

Cold atom traps are at the heart of many quantum applications in science and technology. The preparation and control of atomic clouds involves complex optimization processes, that could be supported and accelerated by machine learning. In this work, we introduce reinforcement learning to cold atom experiments and demonstrate a flexible and adaptive approach to control a magneto-optical trap. Instead of following a set of predetermined rules to accomplish a specific task, the objectives are defined by a reward function. This approach not only optimizes the cooling of atoms just as an experimentalist would do, but also enables new operational modes such as the preparation of pre-defined numbers of atoms in a cloud. The machine control is trained to be robust against external perturbations and able to react to situations not seen during the training. Finally, we show that the time consuming training can be performed in-silico using a generic simulation and demonstrate successful transfer to the real world experiment.