A Battery-Dependent Partial Offloading Scheme with sequential Multi-Task Learning for Mobile Edge Computing

doi:10.1016/j.future.2026.108679

A Battery-Dependent Partial Offloading Scheme with sequential Multi-Task Learning for Mobile Edge Computing

2026-06-25

0

OA

AI

DOI：10.1016/j.future.2026.108679

Original

Original request for help

Save

Abstract

En 中文

Mobile Edge Computing (MEC) enables computational task offloading from resource-constrained End Devices (EDs) to nearby edge servers, thereby reducing latency and energy consumption. This problem has been extensively studied in the literature, however, existing approaches often neglect realistic device constraints, particularly battery dynamics, and rely on limited and non-reproducible datasets. To address these gaps, this paper proposes a Battery-Dependent Partial Offloading Scheme (BDPOS) that integrates the current battery levels of EDs into the cost model, resulting in energy-aware and optimal offloading decisions. The proposed framework jointly optimizes three objectives: determining the optimal number of components, identifying task partitioning, and selecting offloading policies to identify the overall minimum-cost policy. To ensure reproducibility, BDPOS is further used to generate multiple synthetic datasets of varying sizes. A comprehensive comparative analysis of multiple AI models within a sequential Multi-Task Learning (MTL) framework is conducted on these datasets for intelligent offloading in MEC environments. The evaluated models include Deep Neural Networks (DNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), Gated Recurrent Unit (GRU), Bidirectional GRU (Bi-GRU), Minimal Gated Unit (MGU), and Temporal Convolutional Network (TCN). The datasets are pre-processed and optimized using Bayesian-based hyperparameter tuning before model training. The trained models are then evaluated using multiple performance metrics, the Wilcoxon signed-rank test, and computational cost analysis. Moreover, the energy consumption of the proposed algorithm is compared with existing schemes in the literature. Simulation results demonstrate that the proposed technique BDPOS, significantly reduces overall energy consumption compared to existing strategies, while the MTL-based GRU model achieves superior performance, attaining 87.96% accuracy in component optimization, a mean absolute error of 0.0780 for task partitioning, and 63.37% accuracy for offloading policy prediction. These findings highlight the importance of battery-aware modeling and provide actionable insights into the design of efficient, data-driven offloading strategies for next-generation MEC systems.

Keywords:

Mobile Edge Computing (MEC)

Multi-Task Learning (MTI)

Computational offloading

Optimal decision-making

Battery-aware offloading

Bayesian optimization

Partial offloading

Task partitioning

Energy efficiency

Neural networks

End device

AI models

A Battery-Dependent Partial Offloading Scheme with sequential Multi-Task Learning for Mobile Edge Computing

Abstract

Journal

Organization