Impact of Channel and System Parameters on Performance Evaluation of Frequency Extrapolation Using Machine Learning

Channel extrapolation in the frequency domain is an important tool for reducing overhead and latency in frequency division duplex (FDD) wireless communications systems. Over the past years, various machine learning (ML) techniques have been proposed for this goal, but their effectiveness is usually evaluated only for a limited number of examples. This paper presents an extensive investigation of the impact of various parameters, both of the system and of the underlying channels, on the performance of three types of ML algorithms, namely K-nearest neighbor (KNN), convolutional multilayer perceptron (CNN/MLP), and autoencoder structures (AE). We analyze the impact of channel coherence bandwidth and coherence distance, and the number of multipath components, as well as system bandwidth, number of subcarriers, duplex distance, and signal-to-noise ratio (SNR) in uplink and downlink. We also consider both complex and magnitude normalized mean-square error (NMSE) as training and evaluation metrics. Physical interpretations of the obtained results are given. Most importantly, we find that the NMSE can vary by 10 dB or more over physically reasonable ranges of parameters but often shows saturation behavior over part of those ranges. We also find that, in particular, KNN results can be quantitatively and qualitatively different from CNN/MLP and AE. These investigations thus provide insights into meaningful parameter choices for the performance evaluation of new ML algorithms for frequency-domain channel extrapolation.