Training UE Selection Using NWDAF for Reliable Federated Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In mobile communication systems, federated learning is hindered by issues such as terminals not completing training normally or failing to upload trained models to the server, leading to deteriorated model performance.
Innovation Solution
A method for selecting user equipment (UE) using network data analytics functions (NWDAF) to provide QoS sustainability analytics, ensuring reliable model training by excluding UEs likely to fail, and enhancing the 5G system's control plane to manage model distribution and aggregation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If terminals participate in federated learning training, then model training diversity is improved, but training completion reliability deteriorates due to terminals not completing training normally or failing to upload models
Solution Approach 1:
The system performs preliminary actions by selecting training terminals in advance based on predicted training completion probabilities. The network device calculates completion probabilities for candidate terminals and pre-selects terminals with high probabilities before training begins, ensuring reliable participants are chosen beforehand. This prevents terminals likely to fail from being selected, thus maintaining training completion reliability while still achieving diversity through the selection process.
Solution Approach 2:
The system implements feedback mechanisms by continuously monitoring terminal training status and using historical completion data to update completion probability predictions. The network device receives feedback on whether terminals complete training and uses this information to refine future selections. This feedback loop ensures that terminals with demonstrated reliability are preferentially selected, resolving the contradiction between maintaining diversity and ensuring completion reliability.
2Adaptability or versatility
If all candidate terminals are selected for training, then training data diversity is improved, but system resource consumption increases due to multiple terminals potentially failing training
Solution Approach 1:
The system performs preliminary filtering of candidate terminals based on predicted completion probabilities before allocating training resources. By calculating completion probabilities and selecting only terminals above a threshold, the system avoids wasting computational resources on terminals likely to fail. This preliminary action maintains diversity among selected terminals while significantly reducing energy consumption by excluding poor performers in advance.
Solution Approach 2:
The system applies partial selection rather than selecting all candidate terminals. Instead of including every available terminal in training (excessive action), it selectively includes only those with high completion probabilities (partial action). This approach achieves sufficient diversity for effective federated learning while avoiding the resource waste of including terminals unlikely to complete training successfully.
3Device complexity
If terminal selection is done without analytics, then system complexity is reduced, but model performance deteriorates due to unreliable terminal selection
Solution Approach 1:
The system introduces an intermediary analytics function that acts as a bridge between simple terminal selection and complex performance optimization. The NWDAF analytics function provides completion probability predictions that guide terminal selection, achieving reliable model performance without requiring the entire system to become complex. This intermediary layer handles the analytical complexity while keeping the core selection mechanism relatively simple.
Solution Approach 2:
The system enables self-service by allowing terminals to self-evaluate their training completion likelihood based on their own characteristics and historical performance. The analytics function uses terminal-provided information and historical data to generate completion probabilities, reducing the need for complex centralized evaluation. This self-service approach maintains model performance reliability while minimizing system complexity.
Data Source
AI summary
A method and an apparatus for selecting UE through steps of requesting analytics for selecting the UE to participate in an AI/ML operation from an NWDAF in a core network and selecting the UE based on the analytics.