Deep Learning for Probabilistic Interference Predictions in mmWave Networks

Abstract

In wireless communications, interference characterization, prediction, and feedback allow efficient radio resource management and enable advanced scheduling techniques. In this paper, we propose a new probabilistic deep learning algorithm for interference prediction on future resources. First, we study the interference characteristics in downlink 5G mmWave networks, and we show that it is highly dynamic due to the high scalability of the sub-carrier spacing and the beam-based transmissions. Then, we study the correlation properties of the 5G inter-cell interference and we show the effect of different scheduling parameters at the interfering cells on the interference correlation. Our proposed deep learning algorithm is designed to output a predicted interference power distribution on the future resources. The output distribution is then utilized to derive the level of confidence in the predicted interference power. We highlight the importance of the confidence level output of our model on the accuracy-precision performance trade-offs. System-level simulations show that our proposed probabilistic deep learning algorithm outperforms competitive machine learning (ML) and non-ML baselines.