Collaborative Research: CPS Medium: Learning through the Air: Cross-Layer UAV Orchestration for Online Federated Optimization

在人工智能应用需求增长的推动下，过去十年见证了协作网络物理学习系统（CCPLS）的出现。CCPLS通过网络物理系统（CPS）设备之间的协调执行分布式、基于学习的处理任务，预计在未来几年内为商业和国防部门提供关键功能。然而，CCPLS生成的数据往往是大规模的、高维的、异构的和时变的，这对智能建模提出了严峻的挑战。同时，无人驾驶车辆，特别是无人驾驶飞行器（uav），已经显示出在CCPLS中扩大信息共享的希望，特别是在服务不足的地区，如农村地区。该项目的新颖之处在于为UAV-CCPLS集成建立了一个具体的基础，将相关的学习、网络和通信设计方面统一起来，围绕适当的智能度量标准。该项目的影响是为智能农业任务开发无人机辅助CCPLS，以及推进为CCPLS量身定制的无人机和其他无人驾驶车辆的制造。项目成果将通过发布开源软件、研究视频和教程来传播。研究人员将进一步参与课程开发、多样性和外联活动，包括指导本科生研究人员。研究调查围绕三个相互关联的推动力展开。Thrust 1为CCPLS开发了一种新的无人机辅助智能框架，并引入了精确的、面向任务的数据动态和异构概念。此外，这一推动力开发了一个“为学习而学习”的框架，旨在预测估计数据动态的速率。推力2研究了通过无人机轨迹的协同设计和无人机- cps网络的建立来共同优化资源利用和智能质量的方法。推力1号的数据动态框架通过一个在线的、网络感知的顺序决策框架集成到这个设计中。最后，Thrust 3开发了基于学习感知上行和下行无线波束形成和空中聚合方法的CCPLS通信协议。这些协议是针对无人机辅助学习系统的特定需求量身定制的，例如，在无人机对无人机和无人机对接入点通信链路上传输噪声信息。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Propelled by the growth in demand for artificial intelligence-enabled applications, the past decade has witnessed the emergence of Collaborative Cyber-Physical Learning Systems (CCPLS). CCPLS carry out distributed, learning-based processing tasks through coordination among Cyber-Physical System (CPS) devices, and are envisioned to provide critical functionality across the commercial and defense sectors in the next several years. However, the data generated by CCPLS is often large-scale, high-dimensional, heterogeneous, and time-varying, which poses critical challenges for intelligence modeling. Concurrently, unmanned vehicles, in particular Unmanned Aerial Vehicles (UAVs), have shown promise of scaling up information-sharing in CCPLS, especially in under-served regions such as rural areas. The project's novelties are in establishing a concrete foundation for UAV-CCPLS integration that unifies the associated learning, networking, and communication design aspects around appropriate intelligence metrics. The project's impacts are the development of UAV-assisted CCPLS for smart agriculture tasks, as well as advancing the manufacturing of UAVs and other unmanned vehicles tailored for CCPLS. Project outcomes will be disseminated by releasing open-source software and research videos and tutorials. The investigators will further engage in Curriculum development, diversity, and outreach activities including mentoring undergraduate researchers. Research investigations center around three interconnected thrusts. Thrust 1 develops a novel UAV-assisted intelligence framework for CCPLS and introduces a precise, task-oriented notion of data dynamics and heterogeneity. Additionally, this thrust develops a `learning for learning’ framework that aims to predictively estimate the rate of data dynamics. Thrust 2 investigates methodologies for jointly optimizing resource utilization and intelligence quality through co-design of UAV trajectories and UAV-to-CPS network establishment. The data dynamics framework from Thrust 1 is integrated into this design through an online, network-aware sequential decision-making framework. Finally, Thrust 3 develops CCPLS communication protocols based on learning-aware uplink and downlink wireless beamformers and over-the-air aggregation methods. These protocols are tailored to the specific needs of the UAV-assisted learning systems, e.g., the transmission of noisy information over UAV-to-UAV and UAV-to-access point communication links.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.