ML Context Handover Using Make-Before-Break Model Caching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for network automation are impractically large and inflexible, leading to performance issues and handover failures when transitioning between local contexts, particularly in mobility prediction and positioning scenarios.
Innovation Solution
Implement a make-before-break method for seamlessly transferring machine learning context between overlapping model instances, utilizing cache areas and standby models to minimize data transfer interruptions and ensure continuous inference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If a single monolithic ML model instance covers the complete network, then network coverage is improved, but model size becomes impractically large and flexibility is lost
Solution Approach 1:
The patent divides the network into multiple local validity areas, each with its own ML model instance. This segmentation allows the system to cover the complete network through multiple smaller models rather than one large monolithic model, resolving the contradiction between network coverage and model size.
Solution Approach 2:
The patent creates a framework where multiple ML model instances can be activated based on terminal location. The system universally handles different local contexts by selecting appropriate model instances, allowing the network to maintain comprehensive coverage while using only the necessary local models.
2Area of stationary object
If a single monolithic ML model instance is used, then network coverage is improved, but adaptability to different local contexts is reduced
Solution Approach 1:
The patent segments the network into distinct validity areas with location-specific ML model instances. Each model is trained on local data and adapts to local characteristics such as radio propagation environment, enabling the system to maintain network-wide coverage while adapting to diverse local contexts.
Solution Approach 2:
The patent implements local quality by training each ML model instance specifically for its local validity area with unique characteristics. This allows each model to be optimized for its specific context while the overall system provides comprehensive network coverage through the collection of specialized local models.
3Measurement precision
If ML model instances are transferred during handover, then local context accuracy is improved, but data transmission overhead increases
Solution Approach 1:
The patent implements make-before-break handover by activating the target ML model instance before the actual handover occurs. The terminal receives and prepares the second model instance in advance while still using the first model instance, ensuring continuous accurate predictions without interruption or excessive data transfer during the handover event.
Solution Approach 2:
The patent ensures continuous inference by maintaining overlap between old and new ML model instances during handover. The terminal can switch between models seamlessly without interruption, preserving prediction accuracy while minimizing data transmission overhead by avoiding complete model retransmission.
Data Source
AI summary
A method comprising: storing a received first and machine learning model instance and a received second machine learning model instance in a cache of a terminal, wherein the first machine learning model instance is associated to a first cell and configured to make, if activated, a first prediction for the terminal, and the second machine learning model instance is associated to a second cell different from the first cell and configured to make, if activated, a second prediction for the terminal; checking if a predefined first requirement is fulfilled; activating the first machine learning model instance to make the first prediction if the predefined first requirement is fulfilled; inferring a decision involving the terminal based on the first prediction if the predefined first requirement is fulfilled; inhibiting to infer the decision involving the terminal based on the second prediction if the predefined first requirement is fulfilled.