Learning Decentralized and Scalable Resource Management for Wireless Powered Communication Networks

Abstract

This letter presents deep learning approaches for addressing resource allocation problems in wireless-powered communication networks. Conventional deep neural network (DNN) methods require the global channel state information (CSI), invoking impractical centralized operations. Also, their computations depend on the user population, which lacks the scalability of the network size. To this end, we propose decentralized and scalable DNN architectures. We interpret the ideal centralized DNN as a nomographic function that can be decomposed into multiple component DNNs. Each of these is dedicated to processing the local CSI of individual users, thereby leading to the decentralized architecture. To reduce coordination overheads among the H-AP and users, individual users leverage a DNN that compresses local CSI into low-dimensional messages shared with the H-AP. Since these DNN modules are designed to share identical trainable parameters, the proposed learning architecture can be universally applied to various configurations with arbitrary user populations. Numerical results show that the proposed decentralized method achieves almost identical performance to centralized schemes with reduced complexity.