Energy-Efficient Task Offloading and DNN Inference in Dynamic STAR-RIS Assisted MEC With Decomposition-Based DRL
Yiting Huang, Baoshan Lu, Yuling Luo, Junli Fang, Qiang Fu, Sheng Qin, Junxiu Liu · IEEE Transactions on Wireless Communications · 2025
Deploying Deep Neural Network (DNN) models of varying capabilities to support collaborative inference for User Equipment (UE) tasks is becoming increasingly common in Mobile Edge Computing (MEC) systems. In this paper, we aim to minimize energy consumption in a dynamic Non-Orthogonal Multiple Access (NOMA)-based MEC system assisted by a Simultaneously Transmitting and Reflecting Reconfigurable Intelligent Surface (STAR-RIS) under complex Non-Line-of-Sight (NLoS) conditions, while meeting latency requirements and preserving inference accuracy for UEs. We formulate the problem as a non-convex optimization and address it using a Decomposition-Based Twin Delayed Deep Deterministic Policy Gradient (DB-TD3) approach. The problem is decomposed into two subproblems: 1) computation resource allocation and power optimization, and 2) optimization of the offloading ratio, time fractions allocated for reflection and transmission, phase shift, and transmission time. For the first subproblem, we derive optimal CPU frequencies and transmit power allocation through theoretical analysis. For the second subproblem, the offloading ratio, time fractions allocated for reflection and transmission, phase shift, and transmission time are optimized using the TD3 algorithm. Experimental results demonstrate that the DB-TD3 method significantly improves system efficiency and reduces average energy consumption by 59.3% compared to baseline algorithms. Furthermore, the proposed NOMA with STAR-RIS scheme outperforms other offloading methods, achieving an average energy reduction of 41.3%.