According to an aspect, there is provided a computing device for controlling terminal device uplink transmission power. Each terminal device is configured to determine uplink transmission power based on two power control parameters—a target received power for full pathloss compensation and a pathloss compensation coefficient. The computing device initializes a deep Q-learning network in which a state is defined as cell-specific pairs of the power control parameters, an action is defined as a selection of valid values of power control parameters for a cell and a reward is calculated based on the information on data traffic. The computing device trains the deep Q-learning network to approximate a Q value function, determines optimal power control parameters based on thereon and causes transmitting them to access nodes.