JISE

Anticipatory Multi-Unmanned Aerial Vehicles (UAVs) Network is the key to the re-alization of high-bandwidth and demanding multi-UAV applications in the future. An ac-curate and robust Channel Quality Prediction (CQP) model is essential in such anticipatory networks to facilitate the eventual optimization step. However, the models that are pro-posed in the literature are typically designed for static cellular networks and do not con-sider robustness or the cross-domain CQP accuracy as a key performance indicator. In this paper, we investigate the efficacy of three different Machine Learning (ML) models in CQP for multi-UAV networks by training them with univariate and multivariate network metrics data curated through OMNeT++ simulations. The models are then evaluated in two-folds via in-domain and cross-domain evaluations to test their accuracies and robust-ness, respectively. The results from the in-domain evaluations show that multivariate data is key to improving the in-domain performance of the ML models for multi-UAV network throughput prediction, whereas the cross-domain evaluations reveal that more complex models like the Seq2seq are necessary for achieving good robustness against multi-UAV network environments that have different operating conditions, with a maximum improve-ment in cross-domain CQP performance of 200% over the other implemented ML models.