A Dependency-Aware Offloading Algorithm based on Deep Reinforcement Learning for Vehicular Networks

Abstract

Recent years have witnessed the explosive growth of ubiquitous vehicles with extremely intelligent systems, which results in large amounts of data generated. Most of these vehicle applications have the characteristics of computation-intensive and latency-sensitive. Especially, with the rapid development of communication technology, the Internet of Vehicles (IoV) faces formidable challenges imposed by resource-constrained devices and the requirements for low time delay and energy consumption. To address this issue, regarded as a promising solution, Mobile Edge Computing (MEC) can enable vehicles to offload tasks to edge servers for processing. Computation offloading is a critical technology that decides which tasks should be offloaded for minimizing the total cost. However, conventional methods are inefficient to deal with multi-user vehicular networks. Since fine-grained task scheduling can reduce processing latency and power consumption significantly, in this paper, we employ Directed acyclic graphs (DAGs) to describe application and further design a distributed computing offloading algorithm. To be specific, the task offloading decision is illustrated as a Markov decision process (MDP), meanwhile an optimized deep reinforcement learning (DRL) method based on partition-based prioritized experience replay (PPER) is put forward to improve training efficiency of the network. Moreover, in order to extract deeper features of the state, this paper optimizes the actor and critic networks. Numerical results verify that, compared with the existing offloading methods, the proposed algorithm performs more efficiently on delay and energy consumption.