Deep Learning TCP for Mitigating NLoS Impairments in 5G mmWave

Abstract

5G and beyond 5G are revolutionizing cellular and ubiquitous networks with new features and capabilities. The new millimeter-wave frequency band can provide high data rates for the new generations of mobile networks but suffers from NLoS caused by obstacles, which causes packet drops that mislead TCP because the protocol interprets all drops as an indication of network congestion. The principal flaw of TCP in such networks is that the root for packet drops is not distinguishable for TCP, and the protocol takes it for granted that all losses are due to congestion. This paper presents a new TCP based on deep learning that can outperform other common TCPs in terms of throughput, RTT, and congestion window fluctuation. The primary contribution of deep learning is providing the ability to distinguish various conditions in the network. The simulation results revealed that the proposed protocol could outperform conventional TCPs such as Cubic, NewReno, Highspeed, and BBR.