Computer-implemented techniques include, during a delayed processing window, receiving reward data for arm actions taken, where the arm actions were chosen based on a previous version of an arm choice policy, and the previous version of the arm choice policy was determined based on a previous set of reward data for a previous set of arm actions taken. When the delayed processing window has closed, a new arm choice policy is determined based at least in part on the action-reward data, and the previous set of reward data and/or the previous arm choice policy. After a request to choose an arm choice is received, a particular arm action to take is determined based on the new arm choice policy. This chosen arm is provided in response to the request.