Dependent Task Scheduling Using Parallel Deep Neural Networks in Mobile Edge Computing

doi:10.1007/s10723-024-09744-8

2024-02-12

PRE

Sheng Chai *

Jimmy Xiangji Huang

DOI：10.1007/s10723-024-09744-8

原文链接

原文求助

摘要

En 中文

Conventional detection techniques aimed at intelligent devices rely primarily on deep learning algorithms, which, despite their high precision, are hindered by significant computer power and energy requirements. This work proposes a novel solution to these constraints using mobile edge computing (MEC). We present the Dependent Task-Offloading technique (DTOS), a deep reinforcement learning-based technique for optimizing task offloading to numerous heterogeneous edge servers in intelligent prosthesis applications. By expressing the task offloading problem as a Markov decision process, DTOS addresses the dual challenge of lowering network service latency and power utilisation. DTOS employs a weighted sum optimisation method in this approach to find the best policy. The technique uses parallel deep neural networks (DNNs), which not only create offloading possibilities but also cache the most successful options for further iterations. Furthermore, the DTOS modifies DNN variables using a prioritized experience replay method, which improves learning by focusing on valuable experiences. The use of DTOS in a real-world MEC scenario, where a deep learning-based movement intent detection algorithm is deployed on intelligent prostheses, demonstrates its applicability and effectiveness. The experimental results show that DTOS consistently makes optimal decisions in work offloading and planning, demonstrating its potential to improve the operational efficiency of intelligent prostheses significantly. Thus, the study introduces a novel approach that combines the characteristics of deep reinforcement learning with MEC, demonstrating a substantial development in the field of intelligent prostheses through optimal task offloading and reduced resource usage.

Keyword:

Reinforcement learning

Task-offloading

Deep learning algorithms

Intent recognition

Multiple parallel deep neural networks