Research Article
RADEAN: A Resource Allocation Model Based on Deep Reinforcement Learning and Generative Adversarial Networks in Edge Computing
@INPROCEEDINGS{10.1007/978-3-031-63989-0_13, author={Zhaoyang Yu and Sinong Zhao and Tongtong Su and Wenwen Liu and Xiaoguang Liu and Gang Wang and Zehua Wang and Victor C. M. Leung}, title={RADEAN: A Resource Allocation Model Based on Deep Reinforcement Learning and Generative Adversarial Networks in Edge Computing}, proceedings={Mobile and Ubiquitous Systems: Computing, Networking and Services. 20th EAI International Conference, MobiQuitous 2023, Melbourne, VIC, Australia, November 14--17, 2023, Proceedings, Part I}, proceedings_a={MOBIQUITOUS}, year={2024}, month={7}, keywords={edge computing resource allocation deep reinforcement learning generative adversarial networks}, doi={10.1007/978-3-031-63989-0_13} }- Zhaoyang Yu
Sinong Zhao
Tongtong Su
Wenwen Liu
Xiaoguang Liu
Gang Wang
Zehua Wang
Victor C. M. Leung
Year: 2024
RADEAN: A Resource Allocation Model Based on Deep Reinforcement Learning and Generative Adversarial Networks in Edge Computing
MOBIQUITOUS
Springer
DOI: 10.1007/978-3-031-63989-0_13
Abstract
Edge computing alleviates the network congestion and latency pressure on the remote cloud as well as the computation stress on end devices. However, facing numerous tasks, how to effectively allocate and manage the resources of edge servers is of great significance, which affects both the benefits of service providers and Quality of Service (QoS) of users. This paper proposes aresourceallocation model based onDeep Reinforcement Learning (DRL) and GenerativeAdversarialNetworks (G AN)(RADEAN). It considers the future resource occupancy of edge servers, and applies multi-replay memory with priority to eliminate the interaction between experiences and improve sampling efficiency. We maximize the average resource utilization of edge servers while ensuring the average transmission latency (ATL) and average execution time (AET) of tasks in a long-term view. Specifically, based on the state which consists the predicted resource occupation output by GAN, the current resource usage status of edge servers and the characteristics of the task queue, DRL agent makes resource allocation decision for each task. We conduct experiments using real-world data trace, and show that RADEAN outperforms traditional and state-of-the-art models with great generalization, reaching the maximum performance improvement of 48.21% compared with MMRA. The ATL and AET of tasks are also presented to reflect the QoS guarantee. Ablation experiments prove the effectiveness of multi-replay memory and priority sub-modules.
