Machine Learning Methods for Multi-UAV Network Throughput Prediction

A communication system that can proactively react to dynamic channel conditions is an essential characteristic of future mobile edge networks. In any anticipatory networking systems, the ability to accurately predict future channel quality evolution is a pre-requisite for eventual network optimization step. As such, many data-driven throughput predictive models have been proposed in the literature for cellular-based networks that utilize Machine Learning/Deep Learning (ML/DL) techniques. However, these models are not readily applicable in a multi-UAV network due to the highly dynamic UAV nodes and intermittent channel conditions which resulted in significant fluctuations in the achievable link throughput over time. In this paper, we investigate the use of three ML/DL models to perform univariate throughput prediction in a multi-UAV network. A custom set of datasets was procured using
$\text{OMNeT}++$
simulations, comprising one main training dataset and eleven other datasets with varying network or operating conditions for the purpose of cross-domain evaluations. The performances of the models were compared in 2-folds to determine the models' accuracy and robustness by using the test split from the main training dataset and the remaining collected datasets, respectively. The results showed that the implemented Seq2Seq model achieved the best performance in both the in-domain and cross-domain evaluations with a maximum improvement of 12.5% and 63%, respectively.