Federated Learning Control via Sidelink UE Model Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Federated learning in wireless communication systems is prone to performance degradation due to unreliable communication, leading to biased and reduced contributions from UEs with poor communication quality, which affects the convergence and accuracy of the global model.
Innovation Solution
Incorporating sidelink communication between UEs to form centralized UE groups, where a central UE aggregates local models from non-central UEs and uploads them to a control device, ensuring successful contribution of local models to the global model, while avoiding repetition and maintaining data efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If UEs with poor communication quality upload local models directly to the control device, then the communication overhead is reduced, but the reliability of model contribution decreases due to unreliable communication
Solution Approach 1:
The patent introduces central UEs as intermediary nodes between non-central UEs and the control device. Non-central UEs upload their local models to central UEs, which then forward aggregated models to the control device. This intermediary structure ensures that even if direct communication between non-central UEs and the control device is unreliable, the model contributions can still reach the control device through the central UEs, thereby improving the reliability of model contribution without significantly increasing communication overhead.
2Productivity
If all UEs upload their local models directly to the control device, then the convergence speed is improved, but the communication overhead increases significantly
Solution Approach 1:
The patent merges the model contributions of multiple non-central UEs at the central UE level before forwarding to the control device. Instead of each non-central UE independently uploading its local model to the control device (which would generate redundant communication overhead), the central UE aggregates these models and performs a single upload to the control device. This merging approach maintains the convergence speed by ensuring all model contributions are included, while significantly reducing the total data transmission volume through the network.
3Measurement precision
If UEs with poor communication quality are excluded from Federated learning, then the communication reliability is improved, but the accuracy of the global model decreases due to biased contributions
Solution Approach 1:
The patent segments the UE population into central UEs and non-central UEs based on their communication quality. This segmentation allows non-central UEs with poor communication quality to participate in Federated learning through the intermediary central UEs, rather than being excluded entirely. The segmentation ensures that all UE groups, including those with communication challenges, can contribute their local models to the global model training process, thereby preventing bias and maintaining global model accuracy while accommodating varying communication reliabilities.
Data Source
Figure 1~2
Figure 3~4A
Figure 4B~5A
AI summary
The present invention relates to an electronic device and method for a wireless communication system, and a storage medium. A control device for a wireless communication system comprises a processing circuit, and is configured to: receive a plurality of models associated with a plurality of terminal devices in the wireless communication system, wherein the plurality of models comprises at least one aggregation model, each aggregation model among the at least one aggregation model is generated by one corresponding terminal device aggregating a local model of said corresponding terminal device and respective local models from one or more other terminal devices, and each local model is obtained by one corresponding terminal device performing training on the basis of local data of the terminal device.