Contextual Neural Bandit Approach for UAV Selection in Delay-Doppler-Based Vehicular Networks

This letter investigates the problem of optimal uncrewed aerial vehicle (UAV) selection in UAV-assisted vehicular communications, aiming to maximize ground vehicle (GV) data transmission by jointly optimizing its achievable data rate and coverage duration. However, this high-mobility environment introduces complex and nonlinear influences on the transmitted data coming from the highly dynamic factors, e.g., relative UAV-GV velocity, Doppler shifts, UAVs’ altitude, residual UAVs’ energy, etc. To address this, a contextual neural upper confidence bound (NeuralUCB) bandit algorithm is proposed to learn the intricate nonlinear relationship between these dynamic parameters and the achievable transmitted data. The framework incorporates delay-Doppler domain to estimate UAVs’ context features related to Doppler shifts and line-of-sight channel gains. This enables real-time selection of the optimal UAV based solely on the current contextual information and prior transmission history, facilitated by deep online learning. Simulation results confirm that the proposed approach significantly outperforms candidate benchmarks in terms of transmission efficiency.