Expanding the Use of Deep Learning for Classification and Control of Congestion in High Band-High BER TCP/IP Networks

Abstract Current research proves the gains that Deep Learning Neural Networks, especially the Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN) architectures, applied to Congestion Control(CC) can generate in favor of better use of the high band network infrastructure, with high Bit Error Rates (BER), like Free Space Optical communication links. Following this line, the present research seeks to test the congestion prediction capacity of the mentioned Deep Learning architectures, diversifying the complexity of the scenarios, operating at different transmission rates (10, 100, 500, and 1000Mbps), carrying varied quantities of connections, reaching 80. This paper extends previous research, attesting that models extracted from Deep Learning neural networks maintain high performance, both in terms of accuracy in distinguishing between underutilization and overload movements (above 90% for unseen data) and in boosting throughput (reaching up to 6 $$\times $$ more), even in complex-upgraded scenarios, with doubled number of flows at transmission rates ten times higher compared to recent works.