Distributed Training Host Scheduling for QoS-Aware Model Transfer

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Distributed learning in communication networks has not been extensively implemented, leading to suboptimal performance due to separate training of models at individual entities without data aggregation, and resource-intensive model transmission affects communication quality and priority tasks.

Innovation Solution

Implement a mechanism for distributed training of machine learning models through a Meta-Training Host (MTH) that aggregates sub-models from Distributed Training Hosts (DTHs), with UE-gNB communication optimizing model transmission scheduling to minimize resource usage and maintain communication quality.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If federated learning is implemented in communication networks, then model performance is improved through data aggregation, but resource consumption increases affecting communication quality

Engineering Contradiction:
Improvemodel performanceVSAvoidresource consumption
Core Design Contradiction:
Measurement precisionVSLoss of energy

Solution Approach 1:

The patent applies partial action by implementing federated learning selectively for specific models (e.g., deep learning models benefiting from large datasets) rather than all models, and by allowing optional participation of distributed training hosts. This enables model performance improvement where needed while avoiding unnecessary resource consumption for models that don't require aggregation.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The system dynamically adjusts federated learning parameters including aggregation frequency, participation thresholds, and model selection criteria based on network conditions, device capabilities, and data characteristics. This allows optimization of the balance between model performance gains and resource consumption under varying operational conditions.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If model transmission is performed frequently, then model performance is improved through updated aggregation, but communication resources are consumed affecting priority tasks

Engineering Contradiction:
Improvemodel performanceVSAvoidcommunication efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent implements periodic aggregation where the meta-training host requests sub-models from distributed training hosts at scheduled intervals rather than continuously. This periodic approach maintains model performance improvements while significantly reducing communication overhead and preserving resources for priority tasks.

Inventive Principle:
Principle #19Periodic action

Solution Approach 2:

The system incorporates feedback mechanisms where the meta-training host evaluates received sub-models and determines whether aggregation is necessary based on performance metrics and convergence criteria. This feedback-driven approach prevents unnecessary transmission and aggregation operations, optimizing the balance between model performance and communication resource usage.

Inventive Principle:
Principle #23Feedback

3Measurement precision

If all distributed hosts participate in federated learning, then model performance is improved through diverse data aggregation, but device resources are overwhelmed

Engineering Contradiction:
Improvemodel performanceVSAvoiddevice resource load
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent enables optional participation in federated learning where distributed training hosts can choose to participate based on their data quality, device capabilities, and current operational state. This partial participation approach maintains model performance through diverse data aggregation while preventing device resource overload by excluding unsuitable hosts.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The system evaluates and selects distributed training hosts based on local data quality metrics, device capability assessments, and participation readiness. This quality-based selection ensures that only hosts with suitable data and sufficient resources participate, improving model performance while managing device complexity and resource load appropriately.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP4189931B1Distributed training in communication networks
Publication Date: 2025.12.24 NOKIA SOLUTIONS & NETWORKS OY
  • EP4189931B1 patent drawingFigure 1
  • EP4189931B1 patent drawingFigure 2~3
  • EP4189931B1 patent drawingFigure 4

AI summary

It is provided a method comprising: monitoring if a distributed training host receives a request from a meta-training host to provide a machine learning model; checking whether a link from the distributed training host to the meta-training host is required for another data communication having higher priority than providing the machine learning model such that the other data communication will block the link for the providing the machine learning model to the meta-training host; informing the meta-training host, in response to the request, that the link is required for the other data communication if the link is required for the other data communication.