Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multi-task Domain Adaptation for Computation Offloading in Edge-intelligence Networks

Runxin Han, Bo Yang, Zhiwen Yu, Xuelin Cao, George C. Alexandropoulos, Chau Yuen

TL;DR

The paper tackles generalization of computation offloading in MEC under domain shift by introducing MTDA, a privacy-preserving teacher-student framework that enables online adaptation during inference while jointly optimizing binary offloading decisions and continuous resource allocation via a multi-task approach. It combines offline training on a source-domain MINLP-derived dataset with online domain adaptation using pseudo-labels from a weight-averaged teacher and a multi-task loss to align classification and regression tasks. The contributions include a novel MTDA architecture, offline training with cross-entropy and MSE losses, and an online adaptation mechanism featuring EMA teacher updates and stochastic recovery to mitigate forgetting. Empirical results show MTDA yields higher target-domain accuracy, lower MSE, and reduced offloading costs, particularly as the number of mobile users increases, illustrating practical potential for dynamic, privacy-aware edge-intelligence networks.

Abstract

In the field of multi-access edge computing (MEC), efficient computation offloading is crucial for improving resource utilization and reducing latency in dynamically changing environments. This paper introduces a new approach, termed as Multi-Task Domain Adaptation (MTDA), aiming to enhance the ability of computational offloading models to generalize in the presence of domain shifts, i.e., when new data in the target environment significantly differs from the data in the source domain. The proposed MTDA model incorporates a teacher-student architecture that allows continuous adaptation without necessitating access to the source domain data during inference, thereby maintaining privacy and reducing computational overhead. Utilizing a multi-task learning framework that simultaneously manages offloading decisions and resource allocation, the proposed MTDA approach outperforms benchmark methods regarding mean squared error and accuracy, particularly in environments with increasing numbers of users. It is observed by means of computer simulation that the proposed MTDA model maintains high performance across various scenarios, demonstrating its potential for practical deployment in emerging MEC applications.

Multi-task Domain Adaptation for Computation Offloading in Edge-intelligence Networks

TL;DR

The paper tackles generalization of computation offloading in MEC under domain shift by introducing MTDA, a privacy-preserving teacher-student framework that enables online adaptation during inference while jointly optimizing binary offloading decisions and continuous resource allocation via a multi-task approach. It combines offline training on a source-domain MINLP-derived dataset with online domain adaptation using pseudo-labels from a weight-averaged teacher and a multi-task loss to align classification and regression tasks. The contributions include a novel MTDA architecture, offline training with cross-entropy and MSE losses, and an online adaptation mechanism featuring EMA teacher updates and stochastic recovery to mitigate forgetting. Empirical results show MTDA yields higher target-domain accuracy, lower MSE, and reduced offloading costs, particularly as the number of mobile users increases, illustrating practical potential for dynamic, privacy-aware edge-intelligence networks.

Abstract

In the field of multi-access edge computing (MEC), efficient computation offloading is crucial for improving resource utilization and reducing latency in dynamically changing environments. This paper introduces a new approach, termed as Multi-Task Domain Adaptation (MTDA), aiming to enhance the ability of computational offloading models to generalize in the presence of domain shifts, i.e., when new data in the target environment significantly differs from the data in the source domain. The proposed MTDA model incorporates a teacher-student architecture that allows continuous adaptation without necessitating access to the source domain data during inference, thereby maintaining privacy and reducing computational overhead. Utilizing a multi-task learning framework that simultaneously manages offloading decisions and resource allocation, the proposed MTDA approach outperforms benchmark methods regarding mean squared error and accuracy, particularly in environments with increasing numbers of users. It is observed by means of computer simulation that the proposed MTDA model maintains high performance across various scenarios, demonstrating its potential for practical deployment in emerging MEC applications.
Paper Structure (11 sections, 15 equations, 3 figures, 2 algorithms)

This paper contains 11 sections, 15 equations, 3 figures, 2 algorithms.

Table of Contents

  1. Introduction
  2. System Model and Problem Formulation
  3. System Overview
  4. Jobs Processing Model
  5. Problem Formulation
  6. Multi-task Learning based Domain Adaptation
  7. MTDA Structure
  8. Offline Training
  9. Online Adaptation
  10. Numerical Results and Discussion
  11. Conclusion and Future Work

Figures (3)

  • Figure 1: The proposed MTDA neural network architecture.
  • Figure 2: The comparison of the model's classification accuracy in the new environment is shown in (a) and (b), and regression mean square error is shown in (c) and (d), where the number of mobile users ($N$) is 3 and 5.
  • Figure 3: System cost versus server computing resource and data amount is shown in (a) and (b), respectively, where $N$ = 3.