.png){kind=icon}
.png){kind=icon}
Return
Hierarchical Deep Reinforcement Learning-Based Partial Task Offloading in Device-Edge-Cloud Computing Architecture
.png){kind=icon}
PRE
AI
DOI：10.1109/TCE.2025.3611973.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The rapid growth of the Intelligent Internet of Things (IIoT) and 5G applications demands high-performance computing, low latency, and energy efficiency, often beyond the capacity of local devices. Task offloading to edge or cloud servers offers a promising solution. However, achieving optimal partial task distribution across device, edge, and cloud remains challenging due to dynamic network conditions, heterogeneous resources, and conflicting performance objectives. This paper presents a Hierarchical Deep Reinforcement Learning (HDRL) framework for intelligent, resource-aware, latency-constrained partial task offloading in device–edge–cloud architectures. The problem is formulated as a two-level Markov Decision Process (MDP) for structured decision-making. At the high level, a Double Deep Q-Network (DDQN) agent observes the global network state and selects computation tiers. At the low level, a Proximal Policy Optimization (PPO) agent determines optimal task-splitting ratios, enabling fine-grained control of subtask distribution. The HDRL framework jointly optimizes energy consumption, task acceptance ratio, and resource utilization while ensuring end-to-end latency constraints. The system adapts dynamically to varying workloads and resource availability. Simulations showed that HDRL-PTO achieved 80–90% task acceptance, reduced energy consumption compared to DRL-PTO and DRL-TO, and consistently delivered higher episode rewards across Network I and II.
Keywords:
Edge-cloud computing
partial task offloading
hierarchical deep reinforcement learning
Markov decision process
resource allocation
deep reinforcement learning
Journal
IF:
10.9
Papers:
5.1K
Citations:
6.8K