基于FL的无蜂窝网络用户调度与功率分配策略
User scheduling and power allocation strategy for cell-free networks based on federated learning
王华华 1黄烨霞 1李玲1
作者信息
- 1. 重庆邮电大学通信与信息工程学院,重庆 400065
- 折叠
摘要
为解决无蜂窝网络系统中用户链路质量差异和通信、计算资源占用不平衡导致的联邦学习(FL)训练性能受限问题,设计了一个联合用户调度和功率分配的优化方案.首先,提出了一种低复杂度的基于资源优先的二次抽样用户调度(RPSS-US)算法,根据用户计算资源的可用性和链路质量选择用户,优先调度对系统容量和全局模型更新贡献较大的用户参与FL任务,提高整体训练性能.随后,提出了一种基于二分法的功率分配(BM-PA)算法,通过优化功率分配改善用户链路质量差异,以提高数据传输速率,减少FL任务的总体时延.通过交替迭代优化这2个子问题,实现系统性能的联合优化.仿真结果表明,相较于其他对比算法,所提出的算法下行吞吐量提升了47.19%,上行吞吐量提升了22.60%,FL任务时间消耗减少了57.33%,并在达到相同模型精度时的时间开销最小.
Abstract
In order to address the issue of limited training performance in federated learning(FL)due to user link quality disparities and imbalanced communication,and computing resource utilization in cell-free network systems,a joint opti-mization problem for user scheduling and power allocation was designed.Firstly,a low-complexity resource priority based secondary sampling user scheduling(RPSS-US)algorithm was proposed.Users were selected based on the avail-ability of their computing resources and link quality,with priority given to those contributing more to system capacity and global model updates,thus improving overall training performance.Then,a power allocation algorithm based on the binary method(BM-PA)was proposed to optimize power allocation,improve user link quality differences,enhance data transmission rates,and reduce overall FL task delay.By iteratively optimizing these two sub-problems alternately,joint optimization of system performance was achieved.Simulation results demonstrate that compared to other comparison al-gorithms,the proposed algorithm achieves a 47.19%increase in downlink throughput,a 22.60%increase in uplink throughput,and a 57.33%reduction in FL task time consumption,while minimizing time overhead for achieving the same model accuracy
关键词
无蜂窝网络/联邦学习/用户调度/功率分配/联合优化Key words
cell-free network/federated learning/user scheduling/power allocation/joint optimization引用本文复制引用
基金项目
重庆市自然科学基金创新发展联合基金(中国星网)资助项目(CSTB2023NSCQ-LZX0114)
出版年
2024
国家科技期刊平台
NSTL
万方数据