Ultra-High Reliability by Predictive Interference Management Using Extreme Value Theory

TL;DR

This paper introduces a risk-sensitive interference prediction method using extreme value theory and kernel density estimation to enhance ultra-reliable low-latency communications, significantly reducing outage rates and resource usage.

Contribution

It develops a novel EVT-based interference prediction algorithm with quantile estimation and confidence control, outperforming traditional Markov chain models in URLLC resource management.

Findings

Reduces outage rates up to 100-fold compared to DTMC

Achieves target outage probability as low as 10^{-7}

Minimizes radio resource usage by approximately 15%

Abstract

Ultra-reliable low-latency communications (URLLC) require innovative approaches to modeling channel and interference dynamics, extending beyond traditional average estimates to encompass entire statistical distributions, including rare and extreme events that challenge achieving ultra-reliability performance regions. In this paper, we propose a risk-sensitive approach based on extreme value theory (EVT) to predict the signal-to-interference-plus-noise ratio (SINR) for efficient resource allocation in URLLC systems. We employ EVT to estimate the statistics of rare and extreme interference values, and kernel density estimation (KDE) to model the distribution of non-extreme events. Using a mixture model, we develop an interference prediction algorithm based on quantile prediction, introducing a confidence level parameter to balance reliability and resource usage. While accounting for the risk sensitivity of interference estimates, the prediction outcome is then used for appropriate resource allocation of a URLLC transmission under link outage constraints. Simulation results demonstrate that the proposed method outperforms the state-of-the-art first-order discrete-time Markov chain (DTMC) approach by reducing outage rates up to 100-fold, achieving target outage probabilities as low as \(10^{-7}\). Simultaneously, it minimizes radio resource usage \(\simnot15 \%\) compared to DTMC, while remaining only \(\simnot20 \%\) above the optimal case with perfect interference knowledge, resulting in significantly higher prediction accuracy. Additionally, the method is sample-efficient, able to predict interference effectively with minimal training data.