Energy-Efficient Resource Allocation in Vehicular Edge Networks Using Reinforcement Learning: A Framework for Sustainable Urban Mobility

Abstract

The proliferation of Connected and Autonomous Vehicles (CAVs) has led to an increased demand for efficient computation offloading strategies in Vehicular Edge Computing (VEC) networks with the goal of reducing latency and energy consumption. This need is further stressed by the essential function played by such strategies in the development of Smart and Sustainable City initiatives. Proposed reinforcement learning (RL)-based resource allocation model to dynamically adaptively allocate vehicular computation tasks to roadside edge servers by jointly optimising the transmission distance and the load of edge servers, without depending on the static and single-objective heuristic. We then train a Q-learning agent to learn an adaptive offloading policy. After 300 episodes, the average reward per decision converges stably from 0.676 to 0.822. Vehicle density is ranged from 20 to 120. Baseline Evaluation with Random, Round-Robin and Greedy (Least-Loaded) allocation. The simulation results show that the proposed RL model at 100 vehicles has an average energy consumption of 0.079 J/task, 98.28% task completion rate, Average total latency of 115.57 ms and 98.28 tasks per time step (33.61% Decrease in energy consumption compared to Random and 3.36% improvement compared to Round-Robin allocation). The proposed model achieves statistically comparable performance as Greedy heuristic (0.079 J/task vs. 0.078 J/task, 98.28% vs. 98.35% completion rate) and does not require hand-tuned load-balancing rule, thus it generalises across different vehicular densities. The results confirm the promise of RL-based resource allocation as a scalable, energy efficient and adaptive allocation mechanism on vehicular edge networks, towards sustainable and low carbon transport infrastructure in smart cities (SDG 11).