Uncertainty-Aware Conformalized Quantile Regression for ML-based Latency Prediction: For Interface Selection in Unlicensed Spectrum

Kristensen, Anders Peter Bundgaard ; Lyholm, Nicolai Dalsgaard

4. semester

Electronic Systems, Master

2025

2025-06-04

101 pages

Ensuring reliable, low-latency communication in the congested and unlicensed 2.4 GHz ISM band using interfaces based on the IEEE 802.11 and 802.15 standard is a significant problem. This thesis addresses the problem by developing a method to predict the 95th latency quantile (Q95) with a statistically sound measure of uncertainty. We propose an "Uncertainty-Aware Conformalized Quantile Regression" approach. A comprehensive simulation framework was created to model WiFi and Bluetooth Low Energy (BLE) link-level behaviors, generating data to train five distinct latency predictor models: two parametric (MV-Param, GMM-Param) and three quantile regression (MV-QR, GMM-QR MSE, GMM-QR Pinball). These models were then calibrated using split conformal prediction to achieve a 90\% confidence level for the predicted Q95 interval. Key findings show that quantile regression models, particularly the GMM-QR Pinball, offer superior predictive accuracy. This model exceeded the target coverage with a 90.37\% coverage for both WiFi and BLE, whilst yielding tighter median uncertainty intervals (24.27 ms for WiFi, 11.59 ms for BLE) than parametric alternatives. This work demonstrates a validated methodology for uncertainty-aware latency prediction, enabling more intelligent and reliable interface selection in unpredictable wireless environments.

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)

Download
View record in AAU Student Projects