Machine Learning for Network Optimization: An Empirical Performance Analysis Across Diverse Traffic Scenarios
IEEE Access · 2026 · 14 · 85478-85494
DOI: 10.1109/access.2026.3696723 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Ansa Niaz und 3 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
The objective of this scientific study is to determine under which traffic conditions machine-learning-based routing control provides sufficient benefit over conventional equal-cost multi-path routing to justify its additional operational complexity. To address this objective, routing is formulated as a sequential decision-making problem and evaluated through a controlled empirical comparison in Network Simulator 3 using a fat-tree data-center topology. The study compares three controllers: a conventional equal-cost multi-path baseline, a supervised learning controller based on a temporal convolutional network, and a deep reinforcement learning controller based on a deep Q-network. Traffic workloads are derived from public Internet backbone, long-haul, and campus-network trace repositories, with additional distributed-denial-of-service traffic injected from a public attack dataset to construct four operating regimes: Core Internet, Long-Haul Diurnal, Enterprise Campus, and Stress/Attack. Performance is assessed using average and 99th-percentile flow completion time, receiver goodput, packet loss rate, and inference latency under heterogeneous trace-driven traffic conditions. The results show that both learning-based controllers improve latency and reliability relative to the conventional baseline, with the largest gains under volatile and adversarial traffic. In the aggregate medium-load evaluation, the deep reinforcement learning controller reduces average flow completion time by 33.4% and packet loss by 80.7% relative to equal-cost multi-path routing. The supervised controller is more competitive in stable regimes and delivers about half the inference latency of the reinforcement controller (0.8 ms versus 1.6 ms). These findings indicate that supervised learning is attractive when traffic is predictable and operational simplicity is important, whereas reinforcement learning is better justified when traffic non-stationarity, burstiness, and attack-driven stress make long-horizon adaptation more valuable.
Zitieren
APA
Niaz, A., Niaz, M. S., Talpur, K. R., & Talpur, B. A. (2026). Machine Learning for Network Optimization: An Empirical Performance Analysis Across Diverse Traffic Scenarios. IEEE Access, 14, 85478-85494. https://doi.org/10.1109/access.2026.3696723
Vancouver
Niaz A, Niaz MS, Talpur KR, Talpur BA. Machine Learning for Network Optimization: An Empirical Performance Analysis Across Diverse Traffic Scenarios. IEEE Access. 2026;14:85478-85494. doi:10.1109/access.2026.3696723
BibTeX
@article{Niaz2026,
author = {Niaz, Ansa and Niaz, Muhammad Saqib and Talpur, Kazim Raza and Talpur, Bandeh Ali},
title = {Machine Learning for Network Optimization: An Empirical Performance Analysis Across Diverse Traffic Scenarios},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {85478--85494},
doi = {10.1109/access.2026.3696723}
}
EUVIMED