Calibration analysis of probablistic machine learning models for 5G predictive latency

The goal of this thesis is to empirically validate the hypothesis proposed by Skocaj and his team in [1], which states that the hypoexponential distribution effectively models latency in 5G networks, which is essential for making efficient decisions in network management. Through experimental validation using real-world network data, the Gamma BPNN exhibits better performance w.r.t. Gaussian BPNN, particularly in terms of calibration. Calibration plots revealed that the Gamma BPNN predictions closely aligned with actual latency outcomes, highlighting its strong ability to accurately reflect the probabilities of predicted latency values. This achievement not only underscores the Gamma BPNN's suitability for latency prediction in 5G networks but also marks a significant advancement in the field of network performance management.