In this paper, we explore optimization-based and data-driven solutions in a reconfigurable intelligent surface (RIS)-aided multi-user mobile edge computing (MEC) system, where the user equipment (UEs) can partially offload their computation tasks to the access point (AP). We aim at maximizing the total completed task-input bits (TCTB) of all UEs with limited energy budgets during a given time slot, through jointly optimizing the RIS reflecting coefficients, the AP's receive beamforming vectors, and the UEs' energy partition strategies for local computing and offloading. A three-step block coordinate descending (BCD) algorithm is first proposed to effectively solve the non-convex TCTB maximization problem with guaranteed convergence. In order to reduce the computational complexity and facilitate lightweight online implementation of the optimization algorithm, we further construct two deep learning architectures. The first one takes channel state information (CSI) as input, while the second one exploits the UEs' locations only for online inference. The two data-driven approaches are trained using data samples generated by the BCD algorithm via supervised learning. Our simulation results reveal a close match between the performance of the optimization-based BCD algorithm and the low-complexity learning-based architectures, all with superior performance to existing schemes in both cases with perfect and imperfect input features. Importantly, the location-only deep learning method is shown to offer a particularly practical and robust solution alleviating the need for CSI estimation and feedback when line-of-sight (LoS) direct links exist between UEs and the AP.

中文翻译：

---

可重构智能表面辅助移动边缘计算：从基于优化的解决方案到仅基于位置学习的解决方案


在本文中，我们在可重构智能表面（RIS）辅助的多用户移动边缘计算（MEC）系统中探索基于优化和数据驱动的解决方案，其中用户设备（UE）可以将其部分计算任务卸载到接入点 (AP)。我们的目标是通过联合优化 RIS 反射系数、AP 的接收波束成形向量以及 UE 的本地能量分配策略，在给定时隙内最大化所有具有有限能量预算的 UE 的总完成任务输入比特 (TCTB)。计算和卸载。首次提出了三步块坐标下降（BCD）算法，以有效解决保证收敛的非凸TCTB最大化问题。为了降低计算复杂度并便于优化算法的轻量级在线实现，我们进一步构建了两种深度学习架构。第一个将信道状态信息 (CSI) 作为输入，而第二个仅利用 UE 的位置进行在线推理。这两种数据驱动方法是使用 BCD 算法通过监督学习生成的数据样本进行训练的。我们的仿真结果表明，基于优化的 BCD 算法和低复杂度的基于学习的架构之间的性能非常匹配，在完美和不完美输入特征的情况下，所有方案都具有优于现有方案的性能。重要的是，仅位置深度学习方法被证明提供了一种特别实用和强大的解决方案，当 UE 和 AP 之间存在视距 (LoS) 直接链路时，可以减轻 CSI 估计和反馈的需要。