Learning congestion over millimeter-wave channels

This paper studies how learning techniques can be used by the congestion control algorithms employed by transport protocols over 5G wireless channels, in particular millimeter waves. We show how metrics measured at the transport layer might be valuable to ascertain the congestion level. In situations characterized by a high correlation between such parameters and the actual congestion, it is observed that the performance of unsupervised learning methods is comparable to supervised learning approaches. Exploiting the ns-3 platform to perform an in-depth, realistic assessment, allows us to study the impact of various layers of the protocol stack. We also consider different scheduling policies to discriminate whether the allocation of radio resources impacts the performance of the proposed scheme.