Deep Learning-Based Power Control Scheme With Partial Channel Information in Overlay Device-to-Device Communication Systems

Donghyeon Kim,Haejoon Jung,In-Ho Lee

doi:10.1109/access.2021.3109948

Abstract

In the overlay device-to-device (D2D) communication systems, transmit power control is critical to better manage interference, so that the sum rate is maximized. Such power control for sum-rate optimization is NP-hard, which is typically tackled by iterative algorithms such as weighted minimum mean square error (WMMSE) method. However, the iterative power control schemes inherently incur high complexity and excessive latency. To overcome the limitations, we propose a deep learning-based power control scheme with reduced complexity and latency, where partial and outdated channel state information (CSI) is considered. Using a deep neural network (DNN)-based approach, we formulate an optimization problem to maximize the spectral efficiency under the constraints of user fairness and energy efficiency, where the DNN-based method is based on unsupervised learning with no label data generation process. In addition, a CSI reporting method based on the channel-to-interference power ratio is proposed for partial CSI feedback, which considerably reduces the feedback overhead. Through simulations, we show the results of the spectral efficiency, energy efficiency, and fairness performance for various topographical sizes and channel correlation coefficients. Also, it is shown that the proposed scheme achieves better spectral efficiency and energy efficiency than the WMMSE scheme even when it uses a small amount of CSI feedback.

Highlights

An integration of device-to-device (D2D) communication in 5G networks has been widely investigated to provide higher spectral efficiency, lower latency, lower power consumption, and better spectrum utilization [1]
A deep neural network (DNN)-based power control scheme with partial and outdated channel state information (CSI) is proposed for the overlay D2D communication systems, where the power control optimization is to maximize the spectral efficiency under the constraints of the user fairness and the energy efficiency
We propose the partial CSI-based training scheme for deep learning, as shown in Fig. 3, where it is assumed that the base station (BS) has only partial CSI, which is obtained by channel-to-interference power ratio (CIR)-based CSI reporting, and the outdated and current versions of the partial CSI are available at the BS for training

Summary

INTRODUCTION

An integration of device-to-device (D2D) communication in 5G networks has been widely investigated to provide higher spectral efficiency, lower latency, lower power consumption, and better spectrum utilization [1]. In [28], DQN-based transmit power control scheme has been proposed considering outdated channel state information (CSI), which can occur in a real-time cellular network, and it resolved the problem of collecting full CSI and labelled data by using reinforcement learning. A DNN-based power control scheme with partial and outdated CSI is proposed for the overlay D2D communication systems, where the power control optimization is to maximize the spectral efficiency under the constraints of the user fairness and the energy efficiency. Simulation results of the DNN-based power control scheme are shown for the CSI reporting method with different CIR thresholds and various feedback delay in terms of spectral efficiency, user fairness, and energy efficiency.

SYSTEM MODEL

DNN STRUCTURE FOR POWER CONTROL

PARTIAL CSI-BASED TRAINING SCHEME

PARTIAL CSI-BASED TESTING PROCESS

SIMULATION RESULTS

CONCLUSION