ABSTRACT Millimeter‐wave (mmWave) Internet‐of‐Vehicles (IoV) communications rely on large antenna arrays to combat severe path loss and to provide high beamforming gains. However, practical vehicular deployments often adopt analog or hybrid beamforming with constant‐modulus constraints, leading to a highly non‐convex joint beamforming design problem whose iterative solvers incur prohibitive latency under fast channel variations. To enable low‐latency beamforming for dynamic IoV scenarios, we propose an end‐to‐end deep learning framework that directly maps instantaneous channel realizations to the transmit and receive beamformers. Three representative backbones are investigated, including a multilayer perceptron (MLP), a DCGAN‐style convolutional encoder, and a U‐Net encoder–decoder with multi‐scale feature extraction. Moreover, beyond conventional supervised imitation learning, we emphasize rate‐driven learning and develop training strategies that directly optimize the achievable‐rate objective. Extensive simulations demonstrate that the proposed rate‐driven learning substantially improves performance over the previous two‐stage scheme, with the U‐Net backbone consistently providing the best accuracy. Under the premise that the learned solutions achieve over 95% of the optimization‐based benchmark on average, the trained networks provide an approximately 800– inference‐time speedup compared with the iterative based method, highlighting their computational efficiency and practical potential for latency‐sensitive mmWave IoV systems.
Zheng et al. (2026) studied this question.