Distributed Task Offloading for Large-Scale VEC Systems: A Multi-agent Deep Reinforcement Learning Method

Abstract

Vehicular Edge Computing (VEC) is a promising technology to meet the ultra-low delay requirements of many emerging Internet of Vehicles (IoV) resource-intensive tasks. Based on VEC, we propose a distributed intelligent task offloading and workload balance (DIOW) framework. In the framework, the base stations (BSs), mounted with mobile edge computing (MEC) servers, can execute the tasks from task vehicles (TV s). Moreover, the tasks can be transmitted from overloaded BSs to resource-idle BSs. Our optimum design is performed with respect to two types of decision variables: task offloading decisions of TV s and workload balancing decisions of BSs. The objective of DIOW is to minimize the system delay while satisfies the energy consumption constraint of each BS. To obtain the optimum design, the framework adopts a multi -agent deep deterministic policy gradient (MADDPG)-based algorithm. We analyze the effectiveness of the DIOW framework by giving numerical results. Comparisons with existing research schemes demonstrate the advantages of our framework.