A BiLSTM Combining WRELM-Based Method For Online TCP State Prediction

Abstract

Round-trip time (RTT) and throughput are two of the most important parameters that in networks with transmission control protocol (TCP). Deep learning-based time-series forecasting methods such as long-short term memory (LSTM) have recently been widely applied in TCP state prediction due to their strong pattern recognition and accurate prediction ability. However, the practical network environment can be dynamic and may deviate from the situations in which the deep model has been trained, resulting in deteriorated predictions. Furthermore, online retraining of the deep model to adapt to current working environment is usually unfeasible due to the nature of heavy computational complexity. In this paper, we propose a method which can online rectify the TCP predictions with a very small computational overhead (time consumption) by combining Bidirectional LSTM (BiLSTM) with the weighted regularized extreme learning machine (WRELM). Experiments show that the proposed method can greatly increase the online prediction accuracy of TCP states especially when the knowledge of the trained deep model diverges from the conditions of its original working environment.