BROWSE
Detail
Journal Article
Online Learning-Based Task Offloading for Energy-Efficient Edge Vision Analytics Under Latency Constraints in Indoor WLANs
Cited 0 time in 
Share 
Share
- Authors
- Ryangsoo Kim, Hark Yoo, Sung Chang Kim, Yonggang Kim
- Issue Date
- 2026-05
- Citation
- IEEE Transactions on Green Communications and Networking, v.10, pp.2804-2821
- ISSN
- 2473-2400
- Publisher
- IEEE
- Language
- English
- Type
- Journal Article
- Abstract
- Edge computing enables resource-constrained mobile devices to execute deep neural network (DNN)-based vision analytics by offloading tasks to nearby edge servers via wireless networks. However, in practical indoor IEEE 802.11 WLAN environments, the effectiveness of edge-assisted offloading is highly sensitive to time-varying link quality caused by device mobility, multipath fading, structural obstacles, and multi-user contention. In this paper, we develop an adaptive edge-assisted vision analytics framework that minimizes per-frame energy consumption while satisfying a strict latency constraint. Motivated by measurement-driven characterization of the link-quality–performance relationship on a real WLAN testbed, we explicitly formulate an RSSI-conditioned task offloading problem and design a lightweight online learning algorithm that operates without prior statistical knowledge. By integrating a penalty-based latency-aware decision rule with a sliding-window upper confidence bound (SW-UCB) mechanism, the proposed policy adapts to stationary and interval-wise non-stationary network conditions, with regret analysis guaranteeing sublinear performance loss. Extensive experiments on a real indoor WLAN testbed demonstrate that the proposed framework significantly reduces long-term energy consumption, strictly controls latency violations under mobility-induced NLoS and multi-user contention, and incurs negligible runtime overhead.
- Keyword
- Edge computing, energy efficiency, online offloading, WLAN