Reliable and low-delay communication remains a critical challenge in the Internet of Vehicles due to high node mobility, rapidly changing topologies, and frequent channel congestion. Traditional routing and fixed-reward Deep Reinforcement Learning models often fail to adapt effectively, leading to unstable quality of service and reduced reliability. This work proposes an adaptive reward–based deep Q-Network (DQN) routing framework to enhance performance in dynamic vehicular environments. The model integrates a dynamic weighting mechanism that adaptively balances reliability, end-to-end delay, and network congestion based on their gradients during training. The simulation results show that the proposed scheme improves packet delivery ratio by up to 3.6%, and reduces delay by approximately 22%, compared to the DQN, Double DQN, and Dueling DQN approaches. The results also demonstrate that the proposed method achieves robust, real-time, and scalable routing, paving the way for more adaptive, scalable intelligent transportation systems.

Deep Q-Network (DQN),Internet of Vehicles (IoV),Machine learning techniques,Optimal Routing,Deep reinforcement learning