Can we exploit machine learning to predict congestion over mmWave 5G channels?

It is well known that transport protocol performance is severely hindered by wireless channel impairments. We study the applicability of Machine Learning (ML) techniques to predict congestion status of 5G access networks, in particular mmWave links. We use realistic traces, using the 3GPP channel models, without being affected using legacy congestion-control solutions. We start by identifying the metrics that might be exploited from the transport layer to learn the congestion state: delay and inter-arrival time. We formally study their correlation with the perceived congestion, which we ascertain based on buffer length variation. Then, we conduct an extensive analysis of various unsupervised and supervised solutions, which are used as a benchmark. The results yield that unsupervised ML solutions can detect a large percentage of congestion situations and they could thus bring interesting possibilities when designing congestion-control solutions for next-generation transport protocols.