Dual-timescale caching and offloading optimization for semantic edge computing
Author Identifier (ORCID)
Abstract
Effective resource allocation is crucial to the efficiency of semantic edge computing (SEC), especially when faced with constrained energy, radio and computational resources. The unique characteristics of SEC, e.g., the demand of multitask execution and the finite cache capacity of edge servers (ESs), have brought new challenges to the optimization of resource allocation. This paper presents a new dual-timescale deep reinforcement learning (DRL)-based resource allocation algorithm aimed at minimizing the long-term average cost of SEC systems. Specifically, cache configurations are learned in the large timescale using Binary Particle Swarm Optimization (BPSO). In the small timescale, task offloading and computation strategies are distributively determined at local devices through Multi-Agent Proximal Policy Optimization (MAPPO), reducing signaling overhead and offering scalability and reliability. Simulations show that the proposed algorithm reduces the system cost by 33.8% compared to the benchmarks, and can adapt to both latency-sensitive and energy-sensitive tasks by fine-tuning the weighting coefficient.
Document Type
Journal Article
Date of Publication
1-1-2025
Publication Title
IEEE Transactions on Vehicular Technology
Publisher
IEEE
School
School of Engineering
Funders
National Key R&D Program of China (2022YFB2902002, 2022YFB2902303, 2023YFB2904300) / Shanghai Rising-Star Program
Copyright
subscription content
Comments
Xu, J., Ni, W., Qin, Z., Niyato, D., Sun, Y., & Zhang, S. (2025). Dual-Timescale caching and offloading optimization for semantic edge computing. IEEE Transactions on Vehicular Technology. Advance online publication. https://doi.org/10.1109/TVT.2025.3633960