Deep reinforcement learning based migration and execution decisions for multi-hop task offloading in mobile vehicle edge computing

As the Internet of Things (IoT) drives the development of Vehicular Edge Computing (VEC), there is a surge in computational demand from emerging in-vehicle applications. Most existing studies do not fully consider the frequent changes in network topology under high mobility of vehicles and the underutilization of idle resources by single-hop offloading. To this end, we propose a task offloading scheme for vehicular edge computing based on multi-hop offloading. The scheme allows task vehicles to offload tasks to service vehicles with excess idle resources outside the communication range, and adapts to dynamic changes in network topology by introducing the concept of neighboring vehicle connection time. This study aims to minimize the delayed energy consumption utility value of the task under the conditions of satisfying the maximum task delay limit, vehicle computational and storage resource constraints. In response to this NP-hard problem, a two-stage reinforcement learning strategy MOCDD (combining Deep Q Network (DQN) and Deep Deterministic Policy Gradient (DDPG)) is proposed to divide the mixed action space into pure discrete and pure continuous action space to determine task migration, executive decision and vehicle transmission power. Simulation results verify the effectiveness of the proposed scheme.