Characterizing and Taming the Tail in URLLC
2023
In this chapter, we go beyond the average-based system design for ultra-reliable low latency communication (URLLC). We first motivate the need of considering the tail distribution, higher-order statistics, extreme events with very low occurrence probabilities, worst-case metrics, and reliability/latency-affected metrics, e.g., age of information, for URLLC. To investigate these metrics, we introduce the entropic risk measure in financial mathematics as well as the generalized extreme value (GEV) distribution and generalized Pareto distribution (GPD) in extreme value theory. Subsequently, statistical learning methods with gradient descent are explored for characterizing the parameters/models of the GEV distribution and GPD. We further invoke the notion of federated learning in order to tackle the data shortage issues of model training in URLLC regimes. Finally, the performance of beyond-average metrics and various tradeoffs are evaluated and investigated accordingly.