A method includes obtaining at least one policy parameter of a neural network corresponding to a load balancing policy, receiving trajectories for each mobile device in a plurality of mobile devices of the wireless network, each trajectory corresponding to a sequence of states of a respective mobile device, wherein the sequence of states is generated based on a continuous interaction of an existing policy of the respective mobile device with the wireless network, estimating advantage functions for each mobile device in the plurality of mobile devices based on the trajectories for each respective mobile device, and updating the at least one policy parameter based on the estimated advantage functions such that the load balancing policy is determined based on states of each mobile device in the plurality of mobile devices.