According to an aspect, there is provided an apparatus for the performing the following. The apparatus implements, separately for at least one downlink beam, a reinforcement learning model, where a state defines which of the plurality of uplink beams belong to a priority beam set for uplink reception corresponding to a downlink beam, an action is defined as an addition of a new uplink beam to the priority beam set, a removal of an uplink beam from the priority beam set or doing nothing and a reward is calculated based on a change in uplink signal-to-noise ratio due to an action adjusted with a cost for taking the action. The apparatus calculates iteratively at least one optimal state using at least one reinforcement learning model based on uplink signal-to-noise ratio statistics and on the plurality of optimal downlink beams for transmission to said plurality of terminal devices.