Date of Award
5-2022
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
School of Computing
Committee Chair/Advisor
Jim Martin
Committee Member
Long Cheng
Committee Member
Rong Ge
Abstract
Applications ranging from video meetings, live streaming, video games, autonomous vehicle operations, and algorithmic trading heavily rely on low latency communication to operate optimally. A solution to fully support this growing demand for low latency is called dual-queue active queue management (AQM). Dual-queue AQM's functionality is reduced without network traffic throughput prediction.
Perhaps due to the current popularity of machine learning, there is a trend to adopt machine learning models over traditional algorithmic throughput prediction approaches without empirical support. This study tested the effectiveness of machine learning as compared to time series forecasting algorithms in predicting per-flow network traffic throughput on two separate datasets. It was hypothesized that a machine learning model would surpass the accuracy of an autoregressive integrated moving average algorithm when predicting future network per-flow throughput as measured by the mean absolute difference between the actual and predicted values of two independent datasets created by sampling network traffic.
Autoregressive integrated moving average (ARIMA), a deep neural network (DNN) architecture, and a long short-term memory (LSTM) neural network architecture were used to predict future network throughput in two different datasets. Dataset one was used in establishing the initial performance benchmarks. Findings were replicated with a second dataset. The results showed that all three models performed well. ANOVA failed to demonstrate a statistically significant advantage of machine learning over the algorithmic model. From dataset one, ANOVA F = 0.138 and p = 0.983. From dataset two, F = 0.087 and p = 0.994. The coefficient of determination tested the fit of models in the two datasets. The r squared value ranged from 0.971 to 0.983 in the machine models to 0.759 to 0.963 in the algorithmic model.
These findings show no evidence that there is a significant advantage of applying machine learning to per-flow throughput prediction in the two datasets that were tested. While machine learning has been a popular approach to throughput prediction, the effort and complexity of building such systems may instead warrant the use of algorithmic forecasting models in rapid prototyping environments. Whether these findings can be generalized to more extensive and variable datasets is a question for future research.
Recommended Citation
Wagner, Cayden, "A Quantitative Comparison of Algorithmic and Machine Learning Network Flow Throughput Prediction" (2022). All Theses. 3808.
https://tigerprints.clemson.edu/all_theses/3808