Deep Reinforcement Learning for Downlink Power Control in Dense 5G Networks

Abstract: This thesis examines the problem of downlink power allocation in dense 5Gnetworks, and attempts to develop a data-driven solution by employing deepreinforcement learning. We train and test multiple reinforcement learningagents using the deep Q-networks (DQN) algorithm, and the so-called Rainbowextensions of DQN. The performance of each agent is tested on 5G UrbanMacro simulation scenarios, and is benchmarked against a fixed power allocationapproach. Our test results show that the DQN models are successful atimproving data rates at cell-edge, while generalizing well to previously unseensimulation scenarios. In addition, the agents induce throughput balancing effects,i.e., achieve fairness among users, in networks with full-downlink-buffertraffic by properly designing the reward signal.