Predicting future wireless traffic volumes using AI would allow communication systems to automatically adjust network resources to maximise reliability, according to researchers from King Abdullah University of Science and Technology (KAUST), Saudi Arabia.
With 5G wireless communication technology being deployed around the world, researchers are looking ahead to what 6G could offer. One emerging idea is to use AI to coordinate communication resources by learning from historical patterns of network usage across the network over time.
The main problem is that the transmission of usage data from nodes to a central database introduces a substantial bandwidth overhead that negates much of the potential benefits.
KAUST researchers have developed a more accurate “dual attention” prediction scheme that minimises the volume of prediction data that needs to be transferred across the network.
Chuanting Zhang and colleagues Shuping Dang, Basem Shihada and Mohamed-Slim Alouini aimed to decentralise the prediction model.
“Wireless traffic prediction could play a central role in network management as the basis for intelligent communication systems,” said Zhang.
“AI techniques such as deep neural networks are able to accurately model the complicated spatio-temporal nonlinear correlations in wireless traffic. However, as different base stations can have very different traffic patterns, it is quite challenging to develop a prediction model that performs well on all base stations at once.”
The “dual attention” scheme combines a central global model with local models at each base station. It weighs the influence of the local models according to network location and then sends only a limited amount of information from the base stations at each update. The result is a hybrid, low-overhead prediction model that provides a high-quality forecast of the spatial and temporal change in network usage over time.
The framework—called FedDA or dual attention-based federated learning—also enables clustering of base stations based on geolocation to obtain further efficiencies and improvements in prediction accuracy. Using two datasets, the researchers demonstrated that FedDA delivers consistently better prediction performance than other methods for SMS messaging, calls and internet traffic.
“With this method, we have decentralised wireless traffic prediction and also implemented dual-attention global model optimisation by paying attention to both the current knowledge of the central server and the information of local clients,” added Zhang. “Each updated global model can then be deployed to each base station to predict and adapt to new traffic patterns.”
