[Paper Review] Problem-Adapted Artificial Intelligence for Online Network Optimization

Spyridon Vassilaras, Luigi Vigneri+5 more|arXiv (Cornell University)|May 30, 2018

Advanced Bandit Algorithms ResearchDecision Sciences11 references3 citations

TL;DR

This paper proposes problem-adapted artificial intelligence (AI) for online network optimization (ONO) in 5G networks, combining domain-specific modeling with data-driven AI to dynamically balance load and steer traffic. In a real-world evaluation, problem-adapted AI achieved the lowest average delay (0.87 ms) and minimal rejected traffic (0.02%), outperforming robust optimization and online mirror descent in both performance and robustness while scaling efficiently with network size.

ABSTRACT

Future 5G wireless networks will rely on agile and automated network management, where the usage of diverse resources must be jointly optimized with surgical accuracy. A number of key wireless network functionalities (e.g., traffic steering, power control) give rise to hard optimization problems. What is more, high spatio-temporal traffic variability coupled with the need to satisfy strict per slice/service SLAs in modern networks, suggest that these problems must be constantly (re-)solved, to maintain close-to-optimal performance. To this end, we propose the framework of Online Network Optimization (ONO), which seeks to maintain both agile and efficient control over time, using an arsenal of data-driven, online learning, and AI-based techniques. Since the mathematical tools and the studied regimes vary widely among these methodologies, a theoretical comparison is often out of reach. Therefore, the important question `what is the right ONO technique?' remains open to date. In this paper, we discuss the pros and cons of each technique and present a direct quantitative comparison for a specific use case, using real data. Our results suggest that carefully combining the insights of problem modeling with state-of-the-art AI techniques provides significant advantages at reasonable complexity.

Motivation & Objective

• To address the challenge of maintaining optimal network performance amid high spatio-temporal traffic variability in 5G networks.
• To compare the performance of online learning, robust optimization, and problem-adapted AI in real-world online network optimization (ONO) scenarios.
• To demonstrate that integrating domain-specific modeling with AI improves efficiency, robustness, and scalability in dynamic network optimization.
• To evaluate the trade-offs between performance, robustness, and computational complexity across different ONO techniques using real traffic data.

Proposed method

• Proposes a framework called Online Network Optimization (ONO) to maintain agile, efficient, and adaptive control in time-varying 5G networks.
• Employs a two-step approach: first predict traffic demand using historical data, then solve the optimization problem for the predicted conditions.
• Introduces problem-adapted AI that directly predicts optimal base station (BS) loads instead of predicting traffic, reducing error propagation and improving robustness.
• Uses online mirror descent (OMD) for online learning, which adapts without assumptions on traffic statistics but lacks built-in protection against sudden surges.
• Applies robust data-driven optimization to limit rejected traffic to under 0.1% by incorporating constraints on performance degradation.
• Employs real-world traffic traces from a dense urban network in Milan to evaluate all three methods under realistic conditions.

Experimental results

Research questions

• RQ1 Which ONO technique—robust optimization, online learning, or problem-adapted AI—achieves the best trade-off between average delay and robustness to traffic surges?
• RQ2 How does problem-adapted AI compare to traditional methods in terms of scalability and performance under real-world traffic variability?
• RQ3 To what extent can integrating domain-specific modeling with AI reduce prediction errors and improve system stability in online network optimization?
• RQ4 How do the three methods perform during periods of high traffic variability, such as weekday mornings with sudden demand spikes?

Key findings

• Problem-adapted AI achieved the lowest average delay (0.87 ms) among all methods, significantly outperforming robust optimization (7.96 ms) and online mirror descent (8.63 ms).
• Problem-adapted AI maintained the lowest rejected traffic rate (0.02%), effectively protecting against SLA violations even during traffic surges.
• Online mirror descent (OMD) showed fast adaptation but failed to prevent large-scale traffic rejection during rapid morning increases on Monday and Tuesday, leading to high delay spikes.
• Robust optimization successfully limited rejected traffic to below 0.1%, but at the cost of the highest average delay (7.96 ms), indicating a poor performance-robustness trade-off.
• Problem-adapted AI scaled effectively with network size, unlike robust optimization, which suffers from high computational complexity as the number of locations increases.
• The results confirm that combining AI with deep modeling insights—such as predicting optimal loads directly—significantly reduces error accumulation and improves decision-making under uncertainty.