ML Estimation of Network Latency from Core Measurements
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for measuring network latency in 5G networks are limited, as they typically rely on end-to-end measurements that are outside the control of network operators, who often lack visibility into edge and server computing environments managed by third parties, and fail to distinguish uplink and downlink latencies or measure service response times accurately.
Innovation Solution
The use of machine learning techniques to estimate key performance indicators (KPIs) such as latency and response time by training models with partial path measurements from the core network, allowing for end-to-end performance estimation using supervised algorithms or deep reinforcement learning, and incorporating one-way delay estimation for more accurate user experience assessment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If end-to-end measurements are used to measure network latency, then complete path visibility is achieved, but network operators lose control and visibility into edge and server computing environments managed by third parties
Solution Approach 1:
The patent segments the network measurement path into multiple parts: controlled core network segments and uncontrolled edge/server segments. Machine learning models are trained on measurements from both segments, allowing the system to estimate end-to-end performance while maintaining operator control over the core network measurements and modeling approach.
Solution Approach 2:
The patent introduces machine learning models as intermediaries that bridge the gap between controlled core network measurements and uncontrolled edge/server environments. These models learn the relationship between partial path measurements and end-to-end performance, enabling operators to infer performance in third-party environments without direct access.
2Ease of manufacture
If traditional measurement methods are used, then implementation simplicity is maintained, but accuracy in distinguishing uplink and downlink latencies and measuring service response times deteriorates
Solution Approach 1:
The patent performs preliminary actions by collecting and labeling measurement data during the training phase, including uplink latency, downlink latency, and service response time measurements. This pre-collected labeled data enables the machine learning models to accurately distinguish between different latency components without requiring complex real-time measurement infrastructure.
Solution Approach 2:
The patent changes the measurement parameters by introducing multiple types of measurements (ICMP ping, TCP RTT, HTTP response time, service response time) and using machine learning to extract detailed components (uplink latency, downlink latency, service response time) from these measurements, thereby achieving high precision without proportionally increasing implementation complexity.
3Measurement precision
If machine learning models are trained with partial path measurements from core network, then end-to-end performance estimation is achieved, but model training complexity and data requirements increase
Solution Approach 1:
The patent applies partial action by using supervised learning algorithms that are trained on a subset of available data (labeled measurements from the core network) rather than requiring complete end-to-end measurement data. This approach achieves accurate end-to-end performance estimation while reducing training complexity and data requirements compared to methods requiring full path instrumentation.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Embodiments of the disclosed techniques include methods for estimating network performance of a path between a terminal device and an application server through a core network. In one embodiment, the method comprises measuring network performance of a portion of the path within the core network to obtain one or more partial network performance measurements in a first period, obtaining one or more end-to-end measurements of the path in the first period, where the end-to-end measurements are values of an end-to-end performance indicator; and estimating one or more values of the end-to-end performance indicator in a second period after the first period, using partial network performance measurements in the second period and the machine learning function.