Reinforcement Learning Based D2D Power Control Algrithm

Benefitting from the short distance transmission of Device-to-device(D2D) communications in cellular networks, the D2D transmitter can work at a lower power level to guarantee the signals received by the D2D receiver are of good quality. Therefore, D2D communication will not only enhance system throughput, but also improve spectrum and energy efficiency. However, when the cellular radio resources is reused by D2D users, the macro-cellular links will suffer the interference from the D2D links, which makes power control for the D2D communication of great importance.In this thesis, the power control problems in cellular network hybrid with D2D communication are studied, where the scenario of sharing uplink spectrum are considered. Firstly, the power control problem for maximizing system throughput is studied, in which the quality of service(QoS) of cellular user is guaranteed. Specifically,this paper takes a machine learning view to build mathematical model, and regards D2D communication in cellular networks as a multi-agents system,and then designs centralized and distributed Q-learning based power control methods. Simulation results show that Q-learning based power control methods can enhance the system throughput.Secondly, considering the increasing focus on green communication,this paper takes the view from energy efficiency to study the problems of minimizing the transmission power and maximizing the energy efficiency for D2D user, in which the QoS of D2D user is guaranteed. Respectively,two distributed Q-learning based power control methods are proposed.Simulation results show that the methods proposed can effectively decrease the energy consumption and improve energy efficiency for D2D user respectively.