Heterogeneous Online Learning for 'Thing-Adaptive' Fog Computing in IoT

Tianyi Chen, Qing Ling, Yanning Shen, Georgios B. Giannakis

Research output: Contribution to journalArticle

6 Scopus citations


Internet of Things (IoT) is featured with its seamless connectivity of billions of smart devices, which offer different functionalities and serve various personalized tasks. To meet the task-specific requirements such as latency and privacy, the fog computing emerges to extend cloud computing services to the edge of the Internet backbone. This paper deals with online fog computing emerging in IoT, where the goal is to balance computation and communication at fog networks on-the-fly to minimize service latency. Due to heterogeneous devices and human participation in IoT, the online decisions here need to flexibly adapt to the temporally unpredictable user demands and availability of fog resources. By generalizing the classic online convex optimization (OCO) framework, the low-latency fog computing task is first formulated as an OCO problem involving both time-varying loss functions and time-varying constraints. These constraints are revealed after making decisions, and allow instantaneous violations yet they must be satisfied in the long term. Tailored for heterogeneous tasks in IoT, a 'thing-adaptive' online saddle-point (TAOSP) scheme is developed, which automatically adjusts the stepsize to offer desirable task-specific learning rates. It is established that without prior knowledge of the time-varying parameters, TAOSP simultaneously yields near-optimality and feasibility, provided that the best dynamic solutions vary slowly over time. Numerical tests corroborate that our novel approach outperforms the state-of-the-art in minimizing network latency.

Original languageEnglish (US)
Article number8421013
Pages (from-to)4328-4341
Number of pages14
JournalIEEE Internet of Things Journal
Issue number6
StatePublished - Dec 2018



  • Heterogeneous tasks
  • Internet of Things (IoT)
  • Mobile edge computing
  • Online learning
  • Saddle-point method

Cite this