Uncertainty-Aware Conformalized Quantile Regression for ML-based Latency Prediction: For Interface Selection in Unlicensed Spectrum
Authors
Kristensen, Anders Peter Bundgaard ; Lyholm, Nicolai Dalsgaard
Term
4. semester
Education
Publication year
2025
Submitted on
2025-06-04
Pages
101
Abstract
Mange enheder deler det ulicenserede 2,4 GHz ISM-bånd, som ofte er overfyldt. Det gør det svært at opnå pålidelig kommunikation med lav latens for WiFi (IEEE 802.11) og Bluetooth Low Energy (BLE, IEEE 802.15). Denne afhandling udvikler en metode til at forudsige 95 %-kvantilen af latens (Q95) – den forsinkelse som 95 % af pakkerne ikke overskrider – sammen med en statistisk solid usikkerhedsangivelse. Vi introducerer en usikkerhedsbevidst, konformaliseret kvantilregression, som kombinerer kvantilregression med konformal prediktion for at levere forudsigelsesintervaller med et valgt konfidensniveau. Der blev opbygget en omfattende simuleringsramme, der modellerer link-niveau-adfærd for WiFi og BLE, og som genererede data til at træne fem latenstidsmodeller: to parametriske (MV-Param, GMM-Param) og tre kvantilregressionsmodeller (MV-QR, GMM-QR MSE, GMM-QR Pinball). Alle modeller blev kalibreret med split-konformal prediktion for at sigte mod 90 % konfidens for det forudsagte Q95-interval. Resultaterne viser, at kvantilregressionsmodellerne – især GMM-QR Pinball – giver den bedste præcision. Denne model oversteg mål-dækningen med 90,37 % (andelen af tilfælde hvor den sande Q95 ligger inden for det forudsagte interval) for både WiFi og BLE og gav samtidig smallere median-usikkerhedsintervaller end de parametriske alternativer: 24,27 ms for WiFi og 11,59 ms for BLE. Arbejdet demonstrerer en valideret metode til usikkerhedsbevidst latenstidsforudsigelse, som kan understøtte mere intelligent og pålidelig valg af trådløse grænseflader i uforudsigelige radiomiljøer.
The unlicensed 2.4 GHz ISM band is crowded with devices, making reliable, low-latency communication challenging for WiFi (IEEE 802.11) and Bluetooth Low Energy (BLE, IEEE 802.15). This thesis develops a way to predict the 95th-percentile latency (Q95)—the delay that 95% of packets do not exceed—together with a statistically sound measure of uncertainty. We introduce an uncertainty-aware, conformalized quantile regression approach that combines quantile regression with conformal prediction to produce prediction intervals at a chosen confidence level. A comprehensive simulation framework was built to model WiFi and BLE link-level behavior and generate data to train five latency predictors: two parametric models (MV-Param, GMM-Param) and three quantile regression models (MV-QR, GMM-QR MSE, GMM-QR Pinball). All models were calibrated using split conformal prediction to target 90% confidence for the predicted Q95 interval. Findings show that quantile regression models—especially GMM-QR Pinball—achieve the best accuracy. This model exceeded the target coverage, attaining 90.37% (the share of cases where the true Q95 lies within the predicted interval) for both WiFi and BLE, while providing tighter median uncertainty intervals than parametric alternatives: 24.27 ms for WiFi and 11.59 ms for BLE. Overall, the work offers a validated method for uncertainty-aware latency prediction, enabling smarter and more reliable interface selection in unpredictable wireless environments.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Keywords
Latency Prediction ; Interface Selection ; Unlicensed Spectrum ; Uncertainty-Aware ; Conformalized Quantile Regression ; Quantile Regression ; Conformalized ; Machine Learning ; Wireless Communication ; WiFi (IEEE 802.11) ; Bluetooth Low Energy (BLE, IEEE 802.15) ; Medium Access Control (MAC) Protocols ; Carrier-Sense Multiple Access with Collision Avoidance (CSMA/CA) ; Frequency-Hopping Spread Spectrum (FHSS) ; Quantile Regression Models ; Conformal Prediction (CP) ; Uncertainty Intervals ; Confidence Interva ; Radio Access Network (RAN)
Documents