Federated Learning over MU-MIMO Vehicular Networks | JoVE Visualize

Area of Science:

Vehicular networks
Machine learning
Wireless communication

Background:

Federated learning (FL) offers improved accuracy for vehicular applications by leveraging data from multiple vehicles.
Challenges in vehicular FL include limited bandwidth, variable channel quality, and latency, impacting vehicle selection and resource allocation.
Optimizing FL requires characterizing vehicles by learning importance and wireless resource utilization.

Purpose of the Study:

To address the joint vehicle selection and resource allocation problem in multi-cell vehicular networks.
To develop an efficient algorithm for optimizing federated learning in vehicular environments.
To evaluate the impact of multi-user MIMO on FL performance for traffic sign classification.

Main Methods:

Characterized participating vehicles based on learning importance and wireless resource usage.
Formulated a joint vehicle selection and resource allocation optimization problem for multi-cell MU-MIMO networks.
Proposed a "vehicle-beam-iterative" algorithm to approximate the optimization solution.
Conducted extensive simulations using realistic road and mobility models for traffic sign object classification.

Main Results:

Multi-user MIMO (MU-MIMO) was shown to significantly improve the convergence time of the global federated learning model.
Application-specific accuracy targets were achieved faster in scenarios with uniform training data set sizes across vehicles compared to varied sizes.
The proposed "vehicle-beam-iterative" algorithm effectively approximated the solution to the complex optimization problem.

Conclusions:

The study demonstrates the effectiveness of joint vehicle selection and resource allocation in enhancing federated learning for vehicular applications.
MU-MIMO technology is crucial for improving the efficiency and performance of federated learning in vehicular networks.
Future research can explore adaptive strategies for varying data set sizes to further optimize federated learning performance in dynamic vehicular environments.