Diffusion Modeling Meets Deep Reinforcement Learning for Edge Task Scheduling

Mobile Edge Computing (MEC) reduces latency and provides high bandwidth by deploying resources at the network edge. Users can offload tasks via wireless networks to minimize both delay and energy consumption. Traditional optimization methods often struggle with real-time MEC systems, whereas Deep Reinforcement Learning (DRL) performs well due to its learning efficiency. This paper focuses on task scheduling in multi-user MEC environments to minimize system latency and energy consumption. To address the challenges of a large state space and discrete action space, we integrate a Generative Diffusion Model (GDM) with the Soft Actor-Critic (SAC) algorithm, thereby mitigating SAC entropy jitter and Q-function instability. The proposed SD-Diffusion algorithm improves multi-user task scheduling performance. Simulation results demonstrate that it outperforms SAC and other DRL methods in terms of both effectiveness and stability.