QoS-Aware Power Management with Convolutional Neural Network

Abstract

Seeking an autonomous optimal power allocation strategy is one of the most important research issues in machine-to-machine communication networks. Recently, deep learning-based methods provide a promising way to address this research issue. In this paper, we focus on exploring the optimal performance of the convolutional neural network (CNN) by making use of the grid-like topology of the channel information. Extensive comparative experiments have been conducted to compare the advantages of different deep learning methods in allocating power resources. It is shown that, under a similar training setting, on average the optimal value of objective function of CNN is about a quarter better than feedforward neural network (FNN). For several different optimization models, CNN acquired this desirable performance with only about 1 millisecond ultra computational time.