AI 中文总结
该研究针对边缘学习中多无线边缘服务器通信的多址瓶颈，采用分数功率控制，通过随机建模推导核心聚合延迟的解析表达式与上界，数值验证其可大幅降低延迟，为缓解该瓶颈提供了简单模型无关的机制。
AI 中文摘要
当多个无线边缘服务器与公共核心服务器通信时，它们的上行传输会产生多址瓶颈，影响分布式边缘学习系统的延迟。本文分析了该瓶颈，并研究在分层联邦学习（HFL）中采用分数功率控制（FPC）缓解该瓶颈的方法。利用随机无线模型对边缘服务器的空间部署和无线信道增益进行建模，在时分多址（TDMA）聚合策略下，我们解析地刻画了核心聚合延迟（CAL）的均值。我们将累积核心聚合延迟建模为更新奖励过程，其中每个学习轮次构成一个更新周期，任务完成定义为停止时间。该表示将平均累积聚合延迟精确分解为，在独立同分布（iid）服务器选择下，预期停止轮次与每轮预期聚合延迟的乘积。随后，我们推导了采用分数功率控制（FPC）时每轮预期延迟的解析上界。数值结果表明，即使是适度的FPC指数，也能在广泛的部署场景中大幅降低CAL。这些结果表明，分数功率控制是一种简单且与模型无关的机制，可用于缓解边缘学习系统中的无线聚合瓶颈。
英文摘要
When multiple wireless edge servers communicate with a common core server, their uplink transmissions create a multiple-access bottleneck that affects the la- tency of distributed edge learning systems. This paper analyzes this bottleneck and investigates the use of fractional power control (FPC) to mitigate it in hierarchi- cal federated learning (HFL). Modeling the spatial deployment of edge servers and wireless channel gains using stochastic wireless models, we analytically character- ize the mean core aggregation latency (CAL) under a TDMA aggregation policy. We formulate the cumulative core aggregation latency as a renewal reward pro- cess, where each learning round constitutes a renewal cycle and task completion defines the stopping time. This representation yields an exact decomposition of the mean cumulative aggregation latency into the product of the expected stopping round and the expected per-round aggregation latency under iid server selection. We then derive analytical upper bounds on the expected per-round latency under fractional power control (FPC). Numerical results demonstrate that even modest FPC exponents substantially reduce CAL across a wide range of deployment scenar- ios. These results highlight fractional power control as a simple and model-agnostic mechanism for mitigating wireless aggregation bottlenecks in edge learning systems.
CommentsThe authors have decided to withdraw this version because it no longer accurately reflects the intended scope and positioning of the work. A substantially revised manuscript will replace this version
