Latency-Energy Tradeoff in Connected Autonomous Vehicles: A Deep Reinforcement Learning Scheme

Vehicle Edge Computing (VEC)-assisted computational offloading brings cloud computing closer to user equipment (UEs) at the edge of the access network by delivering various services to the UEs with limited processing power and battery. However, in fifth-generation and beyond 5G (B5G) networks, where UEs’ service requests and locations change dynamically, the deployment of static edge server deployments may lead to an increase in latency and total energy consumption. This paper presents a latency-energy-aware, efficient task offloading scheme for connected autonomous vehicular networks. Firstly, vehicles are assembled into clusters, in which vehicle can transmit tasks to the other vehicle, while on the other hand, the VEC server is used for processing the data. We developed a joint resource allocation and offloading decision optimization problem to minimize network latency and total energy usage. Due to the non-convex character of the optimization issue, we employed the Markov decision process (MDP) to convert it to a reinforcement learning (RL) problem. Then, we used a soft-actor critic-based scheme to achieve the optimal policy for resource allocation and task offloading to reduce the total latency and energy consumption for connected autonomous vehicles. Simulation analysis reveals that the proposed scheme attains 46.6% and 17.2% lesser delay, and 28.8% and 20.0% consumes less energy than the Hybrid DRL with Genetic Algorithm (HDRL-GA) and DRL based collaborative Data Scheduling (DRL-CDSS) state-of-art schemes.