A Delay-Aware Deep Reinforcement Learning-Based Adaptive Routing Scheme with Multi-Component Reward for Urban VANETs

Abstract

The rapid evolution of Intelligent Transportation Systems (ITS) has necessitated resilient Vehicular Ad Hoc Networks (VANETs) capable of maintaining high-reliability communication under extreme mobility and unpredictable urban topologies. While traditional protocols like AODV and GPSR struggle with control overhead and communication voids, existing Deep Reinforcement Learning (DRL) routing methods often suffer from catastrophic forgetting and ignore critical cross-layer physical constraints. This paper presents a novel delay-aware, Deep Q-Network (DA-DQN) routing scheme that integrates a multi-component reward system with advanced anti-overfitting mechanisms to ensure stable performance in non-stationary environments. Unlike previous models, the proposed framework utilizes a cross-layer delay evaluator that accounts for IEEE 802.11p-specific constraints, including MAC layer contention and localized queuing delays. By modeling routing as a latencyconstrained Markov Decision Process (MDP), the proposed agent balances a multi-objective reward involving delivery success, accumulated delay, and hop count. Simulation results conducted on a realistic urban network demonstrate that the proposed protocol achieves a Packet Delivery Ratio (PDR) of 84.83%, surpassing baselines by up to 25%, successfully balancing the trade-off between path length and link stability while maintaining end-to-end latency within the safety-critical requirements of urban vehicular applications.