Mobility Prediction at the Tactical Edge: A Handover for Centralized/Decentralized Networks

This paper presents a mobility prediction model as part of a handover mechanism combining centralized and decentralized network control to improve performance and resilience in tactical networks. Our approach utilizes mobility prediction to optimally switch between control modes, minimizing packet loss from live user data flows and maximizing the duration of centralized control connectivity. The prediction model ingests mobility trace files, extracts supplementary mobility features, and employs a Recurrent Neural Network (RNN) to forecast system states representing the network link quality. We evaluate the model using two mobility paradigms: Gauss-Markov and Manhattan-Grid. For the Gauss-Markov model, our method achieves 44% accuracy in predicting network states 60 seconds into the future and up to 72% accuracy for 5-second predictions. Our approach exhibits higher accuracy with the Manhattan-Grid model, attaining 67% and 82% for 60 and 5-second predictions, respectively. A comparative analysis with a previously implemented RSSI-based mechanism demonstrates that our prediction-based mechanism enhances centralized control connectivity duration by 56% and mitigates packet loss by 1.24%. This work contributes to advancing adaptive handover mechanisms for complex tactical networks.