Deep Learning-Based Radio Map for MIMO-OFDM Downlink Precoding

Abstract

Radio map is an advanced technology that mitigates the reliance of multiple-input multiple-output (MIMO) beamforming on channel state information (CSI). In this paper, we introduce the concept of deep learning-based radio map, which is designed to be generated directly from the raw CSI data. In accordance with the conventional CSI acquisition mechanism of MIMO, we first introduce two baseline schemes of radio map, i.e., CSI prediction-based radio map and throughput prediction-based radio map. To fully leverage the powerful inference capability of deep neural networks, we further propose the end-to-end structure that outputs the beamforming vector directly from the location information. The rationale behind the proposed end-to-end structure is to design the neural network using a task-oriented approach, which is achieved by customizing the loss function that quantifies the communication quality. Numerical results show the superiority of the task-oriented design and confirm the potential of deep learning-based radio map in replacing CSI with location information.