Collaborative Research: CCSS: Hierarchical Federated Learning over Highly-Dense and Overlapping NextG Wireless Deployments: Orchestrating Resources for Performance

联邦学习（FL）是被提出用于在下一代（NextG）无线通信系统中的移动的设备上训练机器学习（ML）模型的分布式框架。分层联邦学习（HFL）是一种在无线网络上实现FL的架构。然而，现有的HFL研究福尔斯未能有效地解决NextG通信环境带来的挑战，如高用户和边缘服务器密度，多样化的边缘服务器部署，以及重叠的无线覆盖。为了应对这些挑战，本项目旨在研究HFL中的资源分配问题，重点是选择移动的客户端参与HFL，将它们与边缘服务器相关联，并在这些苛刻的条件下分配足够的带宽。拟议框架的成功完成有可能改变NextG系统的部署和运营，并将为各种ML驱动的应用和服务提供支持。在无线网络上设计HFL性能的主要目标是优化收敛所需的总体训练时间。这可以通过将无线带宽有效分配给每个客户端（带宽分配问题）来最小化每个HFL轮的持续时间来实现。然而，由于高客户端密度、有限的无线频谱和移动性，并非所有客户端都能够参与每一轮。这导致需要确定哪些客户端应该参与每一轮（客户端选择问题），以及在给定重叠的无线覆盖和多个提供商的存在的情况下，哪些客户端应该与哪个边缘服务器相关联（客户端关联问题）。为了应对这些挑战，该项目侧重于三个关键研究领域：（i）在高密度和异构部署下设计用于HFL的短期带宽分配，（ii）开发长期优化框架以解决用于HFL的用户选择和关联，以及（iii）在移动的客户端可以与多个边缘服务器相关联的新兴场景下改进HFL。基金会的使命是履行其使命，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价，被认为值得支持。

Federated learning (FL) is a distributed framework proposed for training machine learning (ML) models on mobile devices in Next Generation (NextG) wireless communication systems. Hierarchical federated learning (HFL) is an architecture that shows promise in enabling FL over wireless networks. However, existing research on HFL falls short in effectively addressing the challenges posed by the NextG communication environment, such as high user and edge server density, diverse edge server deployments, and overlapping wireless coverage. To tackle these challenges, this project aims to investigate resource allocation problems in HFL, focusing on selecting mobile clients to participate in HFL, associating them with edge servers, and allocating sufficient bandwidth under these demanding conditions. The successful completion of the proposed framework has the potential in transforming the deployment and operation of NextG systems and will provide support for a wide range of ML-powered applications and services. The primary objective in designing the performance of HFL over wireless networks is to optimize the overall training time required for convergence. This can be achieved by minimizing the time duration of each HFL round through efficient allocation of wireless bandwidth to each client (the bandwidth allocation problem). However, due to high client density, limited wireless spectrum, and mobility, not all clients may be able to participate in every round. This leads to the need to determine which clients should participate in each round (the client selection problem) and which clients should be associated with which edge server given the overlapping wireless coverage and presence of multiple providers (the client association problem). To address these challenges, the project focuses on three key research areas: (i) designing short-term bandwidth allocation for HFL under highly dense and heterogeneous deployments, (ii) developing a long-term optimization framework to solve user selection and association for HFL, and (iii) improving HFL under the emerging scenario in which a mobile client can be associated with multiple edge servers.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.