Work

Spektrumeffektivt federerat lärande i storskaliga nätverk

Year: 2024
Type: dissertation
Abstract: In an era where the edge of computing is continuously expanding, the advent of federated learning (FL) has opened new horizons in the utilization of data generated at the network’s edge. The essence of federated learning lies in its ability to conduct machine learning (ML) across numerous edge devices, harnessing their data without centralization, which preserves privacy and reduces bandwidth usage. This thesis explores a design for federated learning in large-scale wireless networks, with a focus on mitigating the communication bottleneck that is exacerbated by the interference inherent in large-scale network deployments. The bottleneck hinders the efficient aggregation of model updates, thereby impairing both the speed and accuracy of the collaborative learning process. The research utilizes optimized channel access and power control strategies to reduce interference. In the simulation process, stochastic geometry is applied to model large-scale networks where Poisson cluster processes (PCPs) are leveraged as the stochastic geometry tools to analyze the location correlation between the devices and their servers in clusters. Federated averaging (FedAvg) and broadband analog aggregation (BAA) are employed as federated learning algorithms and over-the-air computation schemes, respectively. In the analysis process, we use the optimality gap as a metric of the learning performance in our system model. The thesis endeavors to minimize this metric and the adverse effects of interference to pave the way for enhanced learning accuracy and optimization in federated learning systems. In conclusion, this thesis not only mitigates a pressing problem in federated learning but also extends the horizons of what is possible in the realm of federated learning and wireless communication, including advanced smart city functionalities and Internet of Things (IoT) integration in industrial settings. Future work is directed towards refining methods of channel access and power control further, exploring optimal parameters to minimize optimality gaps and accelerate convergence rates, ultimately leading to more robust and efficient federated learning models.
Author: Feng, Yiqi
Language: English
Open Access status: green
OpenAlex ID: W7043454873
Cites: 39
Cited by:
Related to: 10