Lookahead VANET Routing: Mobility-aware Deep Neural Network-based Vehicular Communications

Abstract Vehicular ad hoc networks (VANETs) face persistent challenges in maintaining reliable message routing and communication. Existing methods often rely on exhaustive path comparisons or excessive control messages, making them unsuit- able for high-mobility environments. Motivated by these limitations, a Deep neural network (DNN)-based Lookahead routing scheme is proposed that utilizes local spatial-temporal information to enable real-time, globally effcient next- hop decisions. A novel grid-based representation of the vehicular environment is introduced to enable learning-based routing using local observations. First, an analytical model is introduced to minimize per-hop transmission time, effectively reducing end-to-end delay without requiring global topology awareness. Secondly, a U-Net-style DNN is proposed and trained to perform predictive, learning-based routing in dynamic vehicular environments, significantly reducing computational overhead. Empirical evaluations demonstrate that the proposed method outper- forms two state-of-the-art baselines by increasing throughput and packet delivery ratio by 5-10%. Additionally, the outage probability is reduced from 58% to 45% and from 91% to 82%, demonstrating robustness in high mobility.