1. Latency breakdown, a recap
In a network, Latency is a breakdown among
1. Signal propagation
2. Transmission / Transport
3. Processing / Queuing / Routing
Below the flow summarizes the breakdown of every segment of a telecom network.
2. Latencetech Dashboard Forecast
In the actual latest version Latencetech 1.12 integrates statistical model as displayed below:
Let’s have a look at the figures displayed:
1.41 ms is the amount of variation in data measured ie latency. It is calculated by the standard deviation formula. Higher interval could mean lower confidence.
19.53ms is the value is the forecast in 30s-time frame with a 95.86 % confidence level.
The confidence level is computed by MSE indicator, https://en.wikipedia.org/wiki/Mean_squared_error
The forecasted latency value is displayed on the right side (yellow dots), up to 30 seconds every 5 seconds.
Latencetech can forecast any latency or throughput of any telecom network segments, any protocol (TCP, UDP, HTTP, ICMP, TWAMP).
3. What next AI/ML in Telecom?
This is just the beginning of the story!
Latencetech data scientist engineers are working on various models to determine which will be the most effective in predicting latency indicators, incorporating not only raw values but also contextual values (topology, technology, routing, protocols, environment, etc.).
The objective is to show that ML can beat Stats-based approach.
AI/ML used in telecom required special attention to:
Near real-time aspects of telecom network
Dynamic nature of wireless networks
High volumetry of data
Need for high scalability and reliability
In H1-2024 Latencetech worked on a new PoC with a major supplier in the automobile industry. The objective was to test and compare new ML models versus stats and therefore improving forecasting using ML Model with increased accuracy for the next 30 secs only (in steps of 5 secs). At the end, show that AI/ML can beat Stats-based approach.
4. Case study: Connected Vehicle
The Context of the case is self-Driving system (ADAS, Advanced Driver Assistance System) may be improved with real-time connectivity
The Goal is to send immediate alarms if the forecasted connectivity latency is higher than set upper limit.
Therefore, ADAS system will use data to adjust itself (mainly slow down) the vehicle.
What as the Challenges at stake?
To provide real-time network latency measurements (tcp, udp IP stacks) combined with car position (GPS) + speed info+ other contextual data (e.g., weather) to forecast latency in next +90 seconds and send results within 40 ms to car ADAS
Figure: diagram of the PoC
5. Challenges for AI/ML Telecom Real Time
Below some key findings:
High volumes of data Gb> need data streaming and filtering (dataset)
Real-time predictions: in the range of 5-25 ms > need ML distributed framework,
Beyond lab to field > perform trials with early adopters in real environments
The selection of a prediction method depends on several factors:
Data type: Quantitative, qualitative, time series, images, text.
Data size: Machine learning methods often require large datasets.
Prediction objective: Accuracy, interpretability, computational efficiency.
Domain knowledge: Understanding the underlying domain can inform hypothesis formulation.
In practice, combining multiple methods and evaluating their performance through techniques like cross-validation is often necessary.
Any subjects you're interested in, let's engage.
Marc Soulacroup
Paris, France
Comments