Prediction of Time-Varying Wireless Channels Using Recurrent Neural Networks

Abstract

High mobility environment raises multiple challenges in the field of wireless communications. One of the challenges is to compensate for channel variations needed in the use of adaptive communication techniques. Estimation of Channel State Information (CSI) at the receiver helps communication systems to adjust transmitter parameters for better performance. However, due to the feedback delay, outdated CSI cannot be effectively used. This issue has been extensively studied and reported in the literature. Ability to accurately predict the CSI values improves the performance of such systems. Multiple statistical and data driven algorithms for CSI predictions are available. The statistical modeling approach, such as the autoregressive parameter estimation, involves high algorithm complexity. Neural networks offer accurate solutions with reduced online computational cost. In this paper, we study applications of Recurrent Neural Networks (RNNs) for fading channel prediction. We focus on Long Short Term Memory (LSTM) neural networks, a subclass of RNNs, that can identify long-term data correlations. A joint classifier predictor (JCP) that uses a classifier and a set of LSTM networks trained at multiple Doppler frequencies is presented. Various design parameters of JCP are investigated. Numerical results demonstrate significant benefits of the JCP.