Optimizing Task Offloading in NOMA-Enhanced MEC Systems via Large Model-Assisted Reinforcement Learning

Abstract

Mobile edge computing effectively reduces delay and improves the system efficiency by offloading computing tasks. However, the standard communication channel, Orthogonal Frequency Division Multiple Access (OFDMA), is difficult to support large-scale connections in dense networks. Meanwhile, existing task offloading methods fail to fully exploit the structured semantics of the system state in task offloading domain. Therefore, this paper proposes an efficiency strategy: 1) designing an adaptive hybrid communication framework that integrates OFDMA and Non-Orthogonal Multiple Access modes, and dynamically select the optimal communication method based on the task sensitivity entropy; 2) modeling the offloading decision and communication mode selection as a hybrid decision-making process, and adopting reinforcement learning methods to minimize the weighted sum of energy consumption and delay; 3) introducing the large model-assisted proximal policy optimization (LMA-PPO) to encode semantic information of structured data. We conduct high-density experiments in different scenarios and compare them with widely adopted methods. The experimental results show that LMA-PPO performs better in performance, converges faster, and reduces the average cost by 22%. LMA-PPO significantly outperforms existing methods in terms of delay and energy consumption.