Aggregator Node Dynamic Client Selection for Federated Learning

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing distributed machine-learning and federated learning processes in 5G networks face challenges such as performance degradation due to dynamic changes in network resources and the inability of Client NWDAFs to provide sufficient computation or access to training datasets, leading to premature termination of learning processes and reduced accuracy of machine-learning models.

Innovation Solution

A method for dynamically adding new client NWDAFs to ongoing distributed machine-learning or federated learning processes, allowing for the selection of eligible nodes to continue training, thereby expediting the process and maintaining model accuracy by aggregating data from a broader pool of resources.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If distributed machine-learning processes are used to accelerate training speed, then productivity is improved, but reliability deteriorates due to dynamic changes in network resources and inability of client nodes to provide sufficient computation or data access

Engineering Contradiction:

Improvetraining speedVSAvoidprocess stability

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The system dynamically adapts the set of client NWDAFs participating in the federated learning process. The aggregator NWDAF can add or remove client nodes based on their current availability, computation resources, and data access capabilities. This dynamic adjustment ensures the learning process continues reliably even as network conditions change, while maintaining high productivity through optimal resource utilization.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The system changes operational parameters such as the number of participating client nodes, their selection criteria, and resource allocation based on real-time network conditions. By adjusting these parameters dynamically, the system maintains reliable training processes while optimizing training speed according to available resources.

Inventive Principle:

Principle #35Parameter changes

2Adaptability or versatility

If client NWDAFs are selected based on current resource availability, then adaptability is improved, but device complexity increases due to dynamic node selection and management

Engineering Contradiction:

Improvedynamic node selectionVSAvoidnode management complexity

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

Client NWDAFs autonomously monitor their own resource availability and data access capabilities, and self-report their eligibility to participate in federated learning rounds. This self-service approach reduces the management burden on the aggregator NWDAF, enabling adaptive node selection while keeping system complexity manageable.

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The system implements feedback mechanisms where client NWDAFs continuously report their resource status and the aggregator adjusts participant selection accordingly. This feedback loop enables adaptable node selection while automating complexity management through systematic information exchange rather than manual coordination.

Inventive Principle:

Principle #23Feedback

3Duration of action of moving object

If the learning process continues with available nodes instead of terminating prematurely, then duration of action is improved, but loss of information increases due to incomplete training rounds

Engineering Contradiction:

Improvelearning process durationVSAvoidmodel accuracy

Core Design Contradiction:

Duration of action of moving objectVSLoss of information

Solution Approach 1:

The system performs partial training rounds with available client nodes rather than waiting for complete participation. Each round contributes partially to the final model, and multiple partial rounds accumulate to achieve complete training. This approach extends the learning process duration while minimizing information loss through iterative progress.

Inventive Principle:

Principle #16Partial or excessive action

Solution Approach 2:

The federated learning process continues continuously with whatever client nodes are available, rather than terminating when expected nodes are unavailable. This continuous partial action ensures the learning process maintains progress over time, with each available node contributing to reducing information loss through incremental model improvements.

Inventive Principle:

Principle #20Continuity of useful action

Data Source

PatentUS20250103904A1First node, third node, fifth node and methods performed thereby for handling an ongoing distributed machine-learning or federated learning process

Publication Date: 2025.03.27 TELEFONAKTIEBOLAGET LM ERICSSON (PUBL)

AI summary

A computer-implemented method, performed by a first node. The method is for handling an ongoing distributed machine-learning or federated learning (DML/FL) process for which the first node acts an aggregator of data or analytics from a first group of second nodes. The first node operates in a communications system. The first node obtains one or more first indications about one or more third nodes. The one or more first indications include respective information about the third nodes. The respective information indicates that the third nodes are eligible to be selected to participate in the ongoing DML/FL process. The one or more first indications are obtained during the ongoing DML/FL process. The first node then provides, to a fourth node operating in the communications system, an output of the ongoing DML/FL process based on the obtained one or more first indications.