Moving direction‐based adaptive task migration in MEC. Issue 17 (10th July 2022)

• Record Type: Journal Article
• Authors: Li, Yang; Wei, Ziling; Su, Jinshu
• Is Part Of: IET communications. Volume 16:Issue 17(2022)
• Page Start: 2002
• Page End: 2018
• Publication Date: 2022-07-10
• Subjects: Telecommunication systems -- Periodicals; Speech processing systems -- Periodicals; 621.38205
• DOI: 10.1049/cmu2.12455
• Languages: English
• ISSNs: 1751-8628

Abstract

Abstract: Edge computing is expected to be a promising paradigm to provide low‐latency services. Tasks from resource‐limited users can be offloaded to edge servers for efficient execution within a limit time. This innovative technique has attracted widespread attention. Task migration is one of the important problems in mobile edge computing (MEC). Taking vehicle network as an example, during the moving process, a vehicle passes through multiple edge servers and decisions about where to migrate the task need to be made. The moving direction should be emphasized since it directly determines a vehicle's trajectory. Nevertheless, few existing works take the moving direction into consideration. In this paper, task migration issue during the vehicle's mobility process is investigated and the moving direction is specifically considered. The direction helps exclude meaningless selections. The moving process can be formulated as a Markov decision process (MDP) and effort is made to design an adaptive algorithm with direction consideration, aiming at minimizing the total communication time while satisfying the deadline of each task. Based on deep Q network (DQN), we devise a Soft update and parameter Noise applied algorithm DQN‐SN, trying to enlarge action exploration space and stabilize the target network's parameter placement in training process. Besides, with the goal of building credible MEC, a credit‐based scheme is also introduced to establish a trusted edge environment. Extensive experiments are conducted to evaluate the performance of our proposed algorithm. Compared to Greedy algorithm and DQN, the total consumed task communication time of DQN‐SN shows 10–20% reduction. Furthermore, the algorithms with the direction factor always outperform the algorithms without direction consideration.