Modified CNN to Maximize Energy Efficiency in D2D Underlying with Multi-Cell Cellular Network

Abstract

The usage of Device-to-Device (D2D) underlaying to reuse spectrum has a substantial influence on spectrum efficiency. On the other side, interference issues arise as a result of frequency reused by D2D users. Furthermore, wearable devices or communication devices have limited power sources, such as batteries. As a result, the fundamental problem formulation that must be solved is power allocation, with the goal function being to maximize the energy efficiency of the system. In order to provide optimum power allocation, conventional methods such as Convex Approximation (CA)-based algorithm need to run multiple iterations to solve the non-convex problem formulation. Therefore, Convolution Neural Network (CNN) as part of Deep Learning (DL) is utilized to approach (CA)-based algorithm for generating power allocation policies to maximize the systems energy efficiency. However, the conventional method of CNN has limitations in accepting arbitrary input size. Accordingly, to the limitation of CNN, this research proposed the combination of CNN with Spatial Pyramid Pooling (SPP) to overcome the limitation on the input size of conventional CNN. Specifically, the inputs of the model are the user's channel state information, and its outputs are power control policies. The simulation results show that both CNN-SPP and CNN can achieve similar performance to the traditional method up to 95 % accuracy. Furthermore, the combination of CNN and SPP can overcome the limitation on the input size of the conventional CNN method, reducing the number of models that must be trained to just one and applying it to all scenarios regardless of the number of CUEs D2D pairs.