CCSS: Collaborative Research: Towards a Resource Rationing Framework for Wireless Federated Learning

联邦学习（FL）是一种新兴的分布式机器学习范式，具有许多吸引人的特性。尽管早期的研究已经证明了联合优化通信和计算的潜力，但现有的设计并没有针对FL的独特特征进行定制。本项目旨在为无线FL开发一种新颖而严格的资源分配框架，我们称之为资源配给，以强调随着时间的推移平衡资源，以便明确捕获对最终学习结果的长期影响。资源配给建立在严格的理论基础之上，并指导算法的发展，解决物理层和MAC层的特定资源分配问题。联邦学习是一种新兴的无线通信应用程序，该项目有可能推动这种新用例的技术开发。同时，理论基础、算法和验证将广泛推动机器学习、通信理论和无线网络的发展。开发这种实用和有影响力的技术也将有助于保持美国在无线技术方面的领导地位，并保持提供高质量、训练有素和创新的工程师的管道。该项目为无线FL资源配给的成功设计和实施进行了协同活动。对每轮学习中不同资源的FL进行了新颖的收敛分析，建立了一般的后来顺受的原则。在理论基础的指导下，本项目进一步构建了从物理层比特加载、自适应编码调制到MAC层客户端选择、带宽分配、功率控制等具体资源配给设计的综合算法框架。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。

Federated learning (FL) is an emerging distributed machine learning paradigm that has many attractive properties. Despite the early studies that have demonstrated the potential of jointly optimizing communication and computation, existing designs are not tailored to the unique characteristics of FL. This project aims at developing a novel and rigorous resource allocation framework for wireless FL, which we term resource rationing to emphasize balancing resources over time so that the long-term impact to the final learning outcome is explicitly captured. Resource rationing is built on a rigorous theoretical foundation and guides the algorithmic development that solves specific resource allocation problems in both physical and Media Access Control (MAC) layers. Federated learning is an emerging new application for wireless communications, and this project has potential to advance the technology development of this new use case. Meanwhile, the theoretical foundation, algorithms, and validation will broadly advance the state of the art in machine learning, communication theory, and wireless networking. Developing such practical and impactful technology would also help maintain the leadership of the United States in wireless technologies as well as keep the pipeline to supply high-quality, well-trained, and innovative engineers.The project pursues synergistic activities for the successful design and implementation of resource rationing for wireless FL. Novel convergence analysis of FL with varying resource in each learning round is carried out, which establishes the general later-is-better principle. Guided by the theoretical foundation, the project further builds a comprehensive algorithmic framework for specific resource rationing designs, ranging from physical layer bit loading and adaptive coding and modulation to the MAC layer client selection, bandwidth allocation, and power control.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.