Machine Learning-Driven Packet Loss Classification via TCP Jersey and Multi-Layer Perceptron

With the huge increase in the amount of data that is transferred every second, congestion in the network becomes unavoidable. There is always a need for algorithms to control the congestion in the network when the data flow increases drastically. The traditional Transmission Control Protocol (TCP) uses a slow start and congestion avoidance mechanism that is not capable of differentiating the cause of packet loss. In a heterogeneous network, the packet loss could result either from congestion in the wired network or a transmission error in the wireless network. In this paper, we propose a machine learning based improvement over TCP-Jersey, a variant of TCP that can distinguish the cause of packet loss and handle it appropriately. TCP-Jersey contains two key components, the Available Bandwidth Estimation (ABE) algorithm and Congestion Warning (CW) router configuration. Our machine learning model learns to distinguish wireless packet loss from congestion packet loss using Multi-Layer Perceptron (MLP). It predicts the cause of the packet loss and handles the congestion using bandwidth estimation algorithm. The simulation results show that our algorithm outperforms TCP-Jersey, TCP-Reno, TCP-Tahoe and TCP-Vegas in terms of throughput, goodput, fairness index and friendliness index.