[Paper Review] Survey on Machine Learning for Traffic-Driven Service Provisioning in Optical Networks

Tania Panayiotou, Maria Michalopoulou+1 more|arXiv (Cornell University)|Sep 12, 2022

Advanced Optical Network Technologies4 citations

TL;DR

This survey presents a comprehensive review of machine learning (ML)-based techniques for traffic-driven service provisioning in optical networks, focusing on predictive and prescriptive frameworks that enable proactive and adaptive resource allocation. By leveraging ML for traffic forecasting and intelligent decision-making, the approach reduces network over-provisioning, improves resource utilization, and maintains quality-of-service under dynamic traffic conditions.

ABSTRACT

The unprecedented growth of the global Internet traffic, coupled with the large spatio-temporal fluctuations that create, to some extent, predictable tidal traffic conditions, are motivating the evolution from reactive to proactive and eventually towards adaptive optical networks. In these networks, traffic-driven service provisioning can address the problem of network over-provisioning and better adapt to traffic variations, while keeping the quality-of-service at the required levels. Such an approach will reduce network resource over-provisioning and thus reduce the total network cost. This survey provides a comprehensive review of the state of the art on machine learning (ML)-based techniques at the optical layer for traffic-driven service provisioning. The evolution of service provisioning in optical networks is initially presented, followed by an overview of the ML techniques utilized for traffic-driven service provisioning. ML-aided service provisioning approaches are presented in detail, including predictive and prescriptive service provisioning frameworks in proactive and adaptive networks. For all techniques outlined, a discussion on their limitations, research challenges, and potential opportunities is also presented.

Motivation & Objective

• To address network over-provisioning caused by static virtual topologies in optical networks due to fluctuating traffic demands.
• To explore how machine learning can enable proactive and adaptive service provisioning by anticipating traffic patterns and optimizing resource allocation.
• To identify key challenges in ML application at the optical layer, including model uncertainty, fairness in decentralized training, and real-time decision-making.
• To highlight research gaps such as lack of large-scale field trials, adversarial ML threats, and the need for explainable and trustworthy ML systems.
• To provide a roadmap for future research by identifying opportunities in edge learning, customized neural networks, and human-in-the-loop ML integration.

Proposed method

• Systematic review of ML techniques applied to traffic prediction, resource allocation, and service provisioning in optical networks.
• Categorization of ML-based approaches into predictive (forecasting traffic) and prescriptive (optimizing decisions) frameworks for proactive and adaptive networks.
• Analysis of ML models including supervised learning, reinforcement learning, federated learning, and multi-agent RL for distributed decision-making.
• Evaluation of model uncertainty quantification, explainable AI, and human-in-the-loop (HITL) mechanisms to improve trust and reliability.
• Incorporation of network telemetry, SDN/NFV integration, and optical hardware capabilities as enablers for real-time ML inference.
• Discussion of adversarial ML risks and defense mechanisms such as adversarial training to ensure robustness in critical network applications.

Experimental results

Research questions

• RQ1 How can machine learning improve resource utilization and reduce over-provisioning in optical networks under dynamic traffic conditions?
• RQ2 What are the key differences and trade-offs between predictive and prescriptive ML-based service provisioning in optical networks?
• RQ3 What are the main challenges in deploying ML at scale in optical networks, particularly regarding model uncertainty, fairness, and real-time performance?
• RQ4 How can adversarial machine learning threats be mitigated to ensure trustworthy and secure service provisioning in optical networks?
• RQ5 What role do edge-based and tiny ML techniques play in reducing latency and improving scalability for real-time network control?

Key findings

• ML-based predictive and prescriptive provisioning significantly improves network resource utilization, energy efficiency, throughput, and latency compared to static or reactive approaches.
• Federated learning and multi-agent reinforcement learning enable decentralized decision-making with fairness and reduced communication overhead, though model convergence and accuracy disparities remain challenges.
• Model uncertainty quantification and explainable AI are critical for building trustworthy ML systems in mission-critical optical networks.
• Adversarial machine learning poses a serious threat to network reliability, especially in high-latency or safety-critical applications such as autonomous vehicles or remote surgery.
• Current field trials at scale for proactive or adaptive optical networks are lacking, highlighting a key gap between theoretical feasibility and real-world deployment.
• Tiny ML and edge-based inference are emerging as essential enablers for low-latency, energy-efficient, and scalable ML deployment in optical network control planes.