Stabilizer of SMPC Throughput using Neural Networks

Abstract

In view of the successful application of deep learning mainly in the field of image recognition, applications of deep learning in the fields of communication and computer networks are being explored. In these fields, systems that follow proper theoretical calculations and procedures are set up. However, due to the large amount of data to be processed, proper processing takes time, and the aspect that deviates from the theory due to the inclusion of uncertain disturbances sometimes appears. Therefore, deep learning or nonlinear approximation by neural networks may be useful in some cases. The communication method SMPC we have studied measures the throughput by a set of packets called train at the destination node and feeds it back to the source node. By comparing it with the transmission rate, the source node detects congestion on the transmission route and adjusts the packet transmission interval. However, there is a problem that the throughput fluctuates during packets pass through the route, and if it is fed back directly, the transmission rate fluctuates significantly, which causes the fluctuation of the throughput larger and the average throughput even lower. It is not desirable to change the transmission rate unnecessarily. In this study, we tried to stabilize the transmission rate by incorporating prediction and learning by neural network. Optimal transmission rates were predicted by a neural network from the throughput data measured by the destination node, and the results are learned by another neural network to generate a stabilizer. A simple moving average method and a stabilizer using three types of neural networks, MLP, RNN, and LSTM, were built into the transmission controller of the SMPC. It was shown that not only can the fluctuation be reduced, but also the average throughput improves. It is shown that deep learning can be used as a device to predict and output stable values from data with complicated time fluctuations that are difficult to analyze.