Predicting wireless traffic using AI could improve the reliability of future wireless communication

The prediction of aboriginal wireless postulation volumes utilizing artificial quality (AI) would let connection systems to automatically set web resources to maximize reliability. KAUST researchers person present developed a much close "dual attention" prediction strategy that minimizes the measurement of prediction information that needs to beryllium transferred crossed the network.

With 5G wireless connection exertion present being deployed astir the world, researchers are looking up to what 6G could offer. One emerging thought is to usage AI to coordinate connection resources by learning from humanities patterns of web usage crossed the web implicit time. The main occupation is that the transmission of usage information from nodes to a central database—where the AI tin bash its magic—introduces a important bandwidth overhead that negates overmuch of the imaginable benefits.

Chuanting Zhang and colleagues Shuping Dang, Basem Shihada and Mohamed-Slim Alouini addressed this contented by decentralizing the prediction model.

"Wireless postulation prediction could play a cardinal relation successful web absorption arsenic the ground for intelligent communication systems," says Zhang. "AI techniques specified arsenic heavy neural networks are capable to accurately model the analyzable spatio-temporal nonlinear correlations successful wireless traffic. However, arsenic antithetic basal stations tin person precise antithetic postulation patterns, it is rather challenging to make a prediction exemplary that performs good connected each basal stations astatine once."

Zhang's squad developed a hierarchical "dual attention" strategy that combines a cardinal planetary exemplary with section models astatine each basal station. Their strategy weighs the power of the section models according to web determination and past sends lone a constricted magnitude of accusation from the basal stations astatine each update. The effect is simply a hybrid, low-overhead prediction exemplary that provides a high-quality forecast of the spatial and temporal alteration successful network usage implicit time.

The framework—called FedDA oregon dual attention-based federated learning—also enables clustering of basal stations based connected geolocation to get further efficiencies and improvements successful prediction accuracy. Using 2 datasets, the researchers demonstrated that FedDA delivers consistently amended prediction show than different methods for SMS messaging, calls and net traffic.

"With this method, we person decentralized wireless postulation prediction and besides implemented dual-attention planetary exemplary optimization by paying attraction to some the existent cognition of the cardinal server and the accusation of section clients." says Zhang. "Each updated planetary exemplary tin past beryllium deployed to each base station to foretell and accommodate to caller postulation patterns."

---

---