Machine Learning for Power Control in D2D Communication Based on Cellular Channel Gains

Abstract

We consider a mobile network with users seeking to engage in a device-to-device (D2D) communication. Two D2D users (DUEs), a transmitter and a receiver, compose one D2D pair. We assume that the D2D pairs reuse a single communication channel to increase the spectral efficiency. Thus, a power control is needed to manage interference among the D2D pairs and to maximize capacity. We address the problem of D2D power control in the case when only standard cellular channel gains between the DUEs and base stations (BSs) are known while channel gains among DUEs are not available at all. We exploit supervised machine learning to determine transmission powers for individual D2D pairs. We show that the cellular channel gains can, in fact, be exploited to predict the transmission power setting for D2D pairs and, still, close-to-optimum sum capacity of the D2D pairs is reached. Moreover, even if our proposed power control requires no knowledge of the channel gains among DUEs and, thus, introduces no additional signalling, the sum capacity can be increased by 16% to 41.9% with respect to no power control, as demonstrated via simulations.