Utilizing Machine Learning as a Prediction Scheme for Network Performance Metrics of Self-Clocked Congestion Control Algorithm

Abstract

Listen

Translate article

Congestion Control (CC) is a fundamental mechanism to achieve effective and equitable sharing of network facilities. As future networks evolve towards more complex paradigms, traditional CC methods are required to become more powerful and reliable. On the other hand, Machine Learning (ML) has become increasingly popular for solving challenging and sophisticated problems, and scientists have started to turn their interest from rule-based approaches to ML-based methods. This paper employs machine learning models to construct a performance evaluation scheme to predict network metrics for the Self-Clocked Rate Adaptation for Multimedia (SCReAM) algorithm. It uses a rigorous data preprocessing pipeline and a systematic application of ML methods to enhance the performance of the regression model for SCReAM’s performance metrics. Also, we constructed a dataset that provides SCReAM’s input parameters and output metrics, such as network queue delay, smoothed Round Trip Time (sRTT), and network throughput. Each prediction process has several phases: choosing the best initial regressor model, hyperparameter tuning, ensemble learning, stacking regressors, and utilizing the holdout data. Each model’s performance was evaluated through various regression metrics; this study will mainly focus on the coefficient of determination (R2) score. The improvement between the initial best-selected model and the final improved model determined that we were able to increase R2 up to 96.64% for network throughput, 99.4% for network queue delay, and 100% for sRTT.