Authorization

 

Journal

Volume 25, No.2, Special Issue on Artificial Intelligence and Computational Methods for the Medical Cyber-Physical Systems

 
Adaptive Task Offloading and Caching via Reinforcement Learning in Multi-Tier Medical Cyber-Physical Systems (pp.205-226)
S. Qiao1, Q. Xie2, X. He3,*, M. Wang4,*, Q. Du2, A. Almusharraf5, M. Shabaz6
https://doi.org/10.30546/1683-6154.25.2.2026.205


1School of Software, Tiangong University, Tianjin, China, e-mail: siboqiao@tiangong.edu.cn
2School of Control Sciences and Engineering, Tiangong University, Tianjin, China, e-mail: 2331091181@tiangong.edu.cn,
  duqifei@tiangong.edu.cn

3Qingdao Institute of Software, College of Computer Science and Technology, China University of Petroleum, Shandong, China
4School of Life Sciences, Tiangong University, Tianjin, China
5Department of Management, College of Business Administration, Princess Nourah bint Abdulrahman University, P.O. Box 84428,
  Riyadh 11671, Saudi Arabia, e-mail: aialmusharraf@pnu.edu.sa

6Marwadi University Research Center, Marwadi University, Rajkot, Gujarat, India, e-mail: bhatsab4@gmail.com
∗Corresponding author e-mail: b23070003@s.upc.edu.cn, wangmin@tiangong.edu.cn




Abstract.
In multi-tier medical cyber-physical systems (MCPS), heterogeneous devices such as wearable and implantable medical equipment continuously generate a large volume of computational tasks. Due to limited local computing resources, these tasks need to be offloaded to edge servers (ESs). However, the efficiency of medical task execution depends on whether the required AI services are cached at the ESs, while caching strategies are simultaneously influenced by the dynamic arrival of tasks. This interdependence makes the joint optimization of task offloading and service caching highly challenging under random task arrivals. To address this challenge, we propose a Deep Reinforcement Learning-based Task Offloading and Service Caching optimization algorithm (DRL-TOSC) that jointly optimizes medical task offloading and AI service caching in MCPS. By integrating the Temporal Difference (TD) method, the medical agent can gradually learn the optimal policy through online interaction with dynamic environments, ensuring fast convergence and robust decision-making. Furthermore, we design a Valid Action Reallocation Mechanism (VARM) that improves decision efficiency by eliminating invalid actions and renormalizing the probability distribution of valid ones. Experimental results demonstrate that DRL-TOSC reduces the average system cost by 42.3% and improves the service cache hit rate by 20.7% compared to existing benchmark algorithms.


Keywords:
Medical Cyber-Physical Systems, Deep Reinforcement Learning, AI Service Cache, Real-Time Data Processing, Intelligent Decision Support.


AMS Subject Classification
: 68M14, 68T05, 68W27.



COIA 2027

7-9 July 2027
Baku, Azerbaijan

Search