Wireless Channel Prediction Using Artificial Intelligence With Imperfect Datasets

This paper presents the study of machine learning (ML) algorithms for the prediction of vehicular channels under impairments that arise in realistic implementations. Emerging vehicular applications will operate in complex settings with potentially abrupt and quick propagation changes. These features can be difficult to capture by ideal (complete) datasets. Therefore, we consider sets of variable length (incomplete) to reflect the rapidly changing vehicular environment. Our assumption is that, in challenging settings, measurements collected by devices or base stations (BSs) might be the only information available to train models. Our approach covers multiple sub-cases including: i) short sets for rapidly changing settings, and ii) large sets for stationary conditions. Measurements are subject to two additional impairments: incorrect sampling and noise. We use a validated synthetic model for vehicular channels to analyze a spectrum of impairment settings that emulates the transition from non-ideal to ideal conditions. This stress test leads to new conclusions on channel prediction: i) how and why algorithms behave in different ways under diverse conditions (optimality region), ii) derivation of new bounds linked to channel features (coherence time, channel correlation, etc.), iii) optimum parameter settings for ML also linked to channel statistics, and iv) proposal of potential improvements. Linear regression (LR) is shown to have a better trade-off between performance and implementation issues when sets are short, oversampled, and with a high signal-to-noise ratio (SNR). A new method to improve the convergence of polynomial LR in sets close to the undersampling regime is proposed here. Results show that neural networks (NNs), particularly deep learning (DL), continuously reduce the mean square error (MSE) as the length of the set increases. They quickly outperform LR, even in sets near the undersampling condition with low SNR. The effectiveness of prediction is severely degraded when sets are undersampled or subject to low SNR. Convolutional NN (CNN) and particularly LSTM (Long Short-Term Memory) show more resilience to these impairments. One key objective of channel prediction is improving resource allocation to reduce latency and increase reliability, which are crucial metrics in applications such as autonomous vehicles. Our analysis contributes to the understanding (explainability) of how AI behaves under multiple impairments, which can also lead to the improvement of advanced vehicular applications.