.png){kind=icon}
.png){kind=icon}
Return
Reinforcement learning-driven interval multi-objective evolutionary algorithm for task offloading in uncertain cloud–edge
.png){kind=icon}
PRE
AI
DOI：10.1016/j.engappai.2026.115043.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Cloud–edge computing has emerged as a promising paradigm to overcome the high latency of centralized cloud computing and the limited computing capacity of edge computing. However, the uncertain factors, such as fluctuations of communication bandwidth and computing capacity available to tasks across the cloud and edge, are largely neglected in their impact on task offloading performance. To address this issue, we employ interval numbers to model uncertainty in cloud–edge environment and formulate the task offloading problem as an interval multi-objective optimization problem by considering three objectives: latency, energy consumption and cost simultaneously. Specifically, a reinforcement learning-driven interval multi-objective evolutionary algorithm (IMOEA-DQN) is developed for task offloading in uncertain cloud–edge. Firstly, to overcome the limitations of fixed or manually designed environment selection mechanisms in existing multi-objective evolutionary algorithms (IMOEAs) and to achieve adaptive coordination, an innovative deep Q-networks (DQN)-driven environment selection strategy is designed to enable the agent’s online learning by leveraging historical information and dynamically adjusting the prioritization of convergence, diversity, and uncertainty so as to realize a balanced trade-off among them. Furthermore, an uncertainty dominance relation is proposed to accurately quantify the uncertainty of individuals, providing a reliable basis for crossover operation and environment selection, and thereby reducing the uncertainty of the population. Finally, a dual-adaptive elite crossover strategy is designed to generate offspring with better overall quality in terms of convergence, diversity, and uncertainty and to further improve the evolutionary efficiency. Extensive experiments and ablation studies at various environmental scales show that IMOEA-DQN performs extremely competitively with several state-of-the-art algorithms.
Keywords:
task offloading
cloud–edge computing
interval multi-objective optimization
reinforcement learning
uncertainty modeling
Journal
IF:
8
Papers:
5.3K
Citations:
3.5W