.png){kind=icon}
.png){kind=icon}
Return
DNN task computation offloading and resource allocation optimization strategy based on probabilistic early exit
.png){kind=icon}
PRE
AI
DOI：10.1007/s11227-026-08777-y.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Edge Intelligence integrates Artificial Intelligence (AI) and edge computing to provide low latency and privacy preserving services for deep neural network (DNN) inference. To address the conflict between limited edge resources and massive computing demands of DNNs, early-exit DNNs and partition-based offloading have become active research topics. However, existing studies mostly optimize model structure or resource allocation in isolation, lacking joint optimization of task segmentation, scheduling, and server queue management in dynamic multi-user environments. Therefore, this paper focuses on probabilistic early-exit DNN tasks for multiple users, considering confidence, latency, energy consumption, and server queue status. We establish a Mixed Integer Nonlinear Programming (MINLP) model with the objective of a weighted sum of long-term average task completion rate, total latency and energy consumption. To solve this NP-hard problem, a probability-aware Multi-Agent Reinforcement Learning closed loop algorithm is proposed: Genetic Algorithm (GA) for global segmentation search, Multi-Agent Deep Deterministic Policy Gradient (MADDPG) for distributed resource optimization, and Probability Driven Scheduling (PDS) for server side prioritization. The simulation results show that compared with the baseline, this method improves the task completion rate by 4% in high load scenarios and achieves a better balance between latency and energy consumption.
Keywords:
Edge intelligence
DNN inference
Probabilistic early exit
Computation offloading
Resource allocation
Journal
IF:
2.7
Papers:
990
Citations:
1.0W