Loading...

System Optimization in Edge Computing Using Meta-Learning

Zouashkiani, Saeed Reza | 2024

99 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57248 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Hossein Khalaj, Babak
  7. Abstract:
  8. In the evolving landscape of edge computing, with an increasing number of devices and applications, the demand for real-time processing and resource optimization has rapidly grown. Computational offloading, which involves transferring processing tasks from these devices to more powerful edge servers, has been proposed to reduce execution latency. In edge computing environments, computational and communication resources are numerous and constantly changing. Additionally, real-world applications involve interdependent processing tasks. All these factors lead to the NP-hardness of the dependent task offloading decision. To tackle these challenges, traditional reinforcement learning (RL) methods have been proposed. However, these methods lack efficiency due to the need for extensive retraining. Meta-reinforcement learning, by adding an additional learning layer, enables rapid adaptation to new scenarios and requires less training time. Nonetheless, these methods alone cannot fully model the complex dependencies between tasks. Therefore, using graph neural networks (GNNs) is essential for effectively modeling these dependencies. This thesis presents a novel edge offloading algorithm that leverages the rapid adaptability of meta-reinforcement learning and the dependency modeling capabilities of GNN architectures. This hybrid approach effectively manages the complex dependencies in real-world applications and overcomes the limitations of traditional RL methods. To address training challenges, we employed a technique to freeze the initial parts of the model, reducing the number of training parameters by more than half while maintaining rapid adaptability. The experiments and evaluations we conducted demonstrate that our approach outperforms existing methods in execution latency when managing dynamic and interdependent task offloading scenarios
  9. Keywords:
  10. Meta Reinforcement Learning ; Graph Neural Network ; Computation Offloading ; Edge Offloading ; Task Dependency Modeling ; Edge Computing

 Digital Object List

 Bookmark

...see more