Geometric Machine Learning Over Riemannian Manifolds for Wireless Link Scheduling
Geometric Machine Learning Over Riemannian Manifolds for Wireless Link Scheduling
复制标题
DOI：
10.1109/access.2022.3153324
复制
发表时间：
2022
期刊：
影响因子：
3.9
通讯作者：
R. Shelim;A. Ibrahim
中科院分区：
文献类型：
--
作者：
R. Shelim;A. Ibrahim
在本文中，我们提出了两种新颖的几何机器学习（G-ML）方法来解决设备到设备（D2D）网络中的无线链路调度问题。在动态D2D网络（例如车载网络）中，获取大量训练样本对于实时响应来说非常耗时，并且由于高移动性，获取准确的瞬时信道状态信息（CSI）具有挑战性。我们的目标是在黎曼流形上有效地表示 D2D 网络，并使用 G-ML 进行少量训练无线网络布局且无需 CSI 来实现动态网络中急需的最先进的总速率最大化性能。为此，我们首先通过黎曼流形上的一组正则化拉普拉斯矩阵将每个 D2D 对周围的局部图建模为一个点。我们计算 D2D 对之间的黎曼度量，例如对数欧几里得度量（LEM），这是这些对之间干扰的合适度量。我们使用几何支持向量机（G-SVM）方法中的LEM以监督学习的方式对链路调度决策进行分类。然后，针对某些动态网络中没有可用标记训练的情况，我们提出了无监督调度方法的几何 $k$ 聚类。仿真结果表明，与现有最先进的方法相比，所提出的方法在总速率最大化方面实现了有希望的性能，仅需要大约一百个训练无线网络布局进行训练并且不使用 CSI。
In this paper, we propose two novel geometric machine learning (G-ML) methods for the wireless link scheduling problem in device-to-device (D2D) networks. In dynamic D2D networks (e.g., vehicular networks), obtaining a large number of training samples is time-consuming for real-time response, and acquiring accurate instantaneous channel state information (CSI) is challenging due to high mobility. Our goal is to efficiently represent D2D networks on Riemannian manifold and use G-ML with few training wireless network layouts and no CSI to approach the sum rate maximization performance as state-of-the-arts which is much needed in dynamic networks. To this aim, we first model the local graph around each D2D pair as a point through a set of regularized Laplacian matrices on the Riemannian manifold. We compute the Riemannian metric, e.g., Log-Euclidean metric (LEM), among D2D pairs, which are suitable measures of interference among these pairs. We use LEM in the geometric support vector machine (G-SVM) method to classify the link scheduling decisions in a supervised learning manner. Then we propose geometric $k$ -means clustering for unsupervised scheduling method for the case when no labeled training is available in some dynamic networks. Simulation results demonstrate that the proposed methods achieve promising performance for sum rate maximization against the existing state-of-the-arts approaches with only around a hundred training wireless network layouts for training and without using CSI.