Actor-Critic Based DRL Algorithm for Task Offloading Performance Optimization in Vehicle Edge Computing

Abstract

Due to the rapid development of the Internet of Things (IoT), many latency-sensitive application businesses have recently emerged, such as the Telematics business. The traditional approach is to reduce latency by offloading tasks to MEC servers to meet the low latency requirements of the business. However, the limited number of MECs is not enough to cover all the roads, and adding MEC devices can greatly increase the equipment cost. In this study, a multi-vehicleassisted MEC system is proposed as a task offloading model for deep reinforcement learning (DRL)-based vehicle edge computing (VEC). The system includes both vehicles with limited computational power and VEC servers with more powerful processing power of roadside units (RSUs). We employ an Actor-Critic based DRL technique to determine when tasks are executed in the local vehicle or offloaded to the RSU unit for execution, expecting to improve the convergence speed of the algorithm while obtaining the lowest latency system performance improvement. Simulation results show that our proposed Actor-Critic based DRL approach can effectively accelerate the convergence speed of the system, improve the system performance, and reduce the overall cost of the system compared to the conventional DQN approach.