AI-based Robust Convex Relaxations for Supporting Diverse QoS in Next-Generation Wireless Systems
AI-based Robust Convex Relaxations for Supporting Diverse QoS in Next-Generation Wireless Systems
复制标题
DOI：
10.1109/icdcsw53096.2021.00014
复制
发表时间：
2021-07
期刊：
影响因子：
--
通讯作者：
Steve Chan;M. Krunz;Bob Griffin
中科院分区：
文献类型：
--
作者：
Steve Chan;M. Krunz;Bob Griffin
支持 5G 及其他无线系统中的不同服务质量 (QoS) 要求通常涉及解决一系列凸优化问题，并采用不同的方法来最佳地解决每个问题。即使当输入集被专门设计/架构为凸范式时，所得输出集仍然可能是非凸的，从而需要通过某些松弛技术转换为凸优化问题。这种转换本身可能会产生其他非凸优化问题，突出显示利用鲁棒凸松弛（RCR）框架的需求/机会。在本文中，我们探索了一类特殊的卷积神经网络（CNN），即深度卷积生成对抗网络（DCGAN），它不仅可以解决与 QoS 相关的凸优化问题，还可以利用相同的 RCR 机制来调整其自身的超参数。这种方法带来了各种技术挑战。例如，粒子群优化（PSO）通常用于超参数减少/调整。当在 DCGAN 上实现时，PSO 需要将连续/不连续超参数转换为离散值，这可能会导致粒子过早停滞在局部最优处。所涉及的实现机制，例如增加惯性权重，可能会产生其他凸优化问题。我们引入了一个 RCR 框架，该框架利用了基于“You Only Look Once”（YOLO）的 DCGAN 的前馈结构。具体来说，我们使用压缩深度卷积-YOLO-生成对抗网络（DC-YOLO-GAN），以下称为改进的压缩YOLO v3实现（MSY3I），与凸松弛对抗训练相结合，以改善每个连续神经网络层的边界紧缩，并通过MSY3I内的特定数值稳定性实现更好地促进全局优化。
Supporting diverse Quality of Service (QoS) requirements in 5G and beyond wireless systems often involves solving a succession of convex optimization problems, with varied approaches to optimally resolve each problem. Even when the input set is specifically designed/architected to segue to a convex paradigm, the resultant output set may still turn out to be nonconvex, thereby necessitating a transformation to a convex optimization problem via certain relaxation techniques. This transformation in itself may spawn yet other nonconvex optimization problems, highlighting the need/opportunity to utilize a Robust Convex Relaxation (RCR) framework. In this paper, we explore a particular class of Convolutional Neural Networks (CNNs), namely Deep Convolutional Generative Adversarial Network (DCGANs), to solve not only the QoS-related convex optimization problems but also to leverage the same RCR mechanism for tuning its own hyperparameters. This approach gives rise to various technical challenges. For example, Particle Swarm Optimization (PSO) is often used for hyperparameter reduction/tuning. When implemented on a DCGAN, PSO requires converting continuous/discontinuous hyperparameters to discrete values, which may result in premature stagnation of particles at local optima. The involved implementation mechanics, such as increasing the inertial weighting, may spawn yet other convex optimization problems. We introduce a RCR framework that capitalizes upon the feed-forward structure of the “You Only Look Once” (YOLO)- based DCGAN. Specifically, we use a squeezed Deep Convolutional-YOLO-Generative Adversarial Network (DC-YOLO-GAN), hereinafter referred to as a Modified Squeezed YOLO v3 Implementation (MSY3I), combined with convex relaxation adversarial training to improve the bound tightening for each successive neural network layer and to better facilitate the global optimization via a specific numerical stability implementation within MSY3I.