[Paper Review] Machine Learning Applications in the Routing in Computer Networks

Ke Liang, Mitchel Myers|arXiv (Cornell University)|Apr 5, 2021

Software-Defined Networks and 5G4 citations

TL;DR

This survey explores machine learning applications in computer network routing, evaluating centralized and decentralized architectures using supervised and reinforcement learning. It demonstrates improved routing performance in simulations, though real-world validation remains limited, highlighting the need for practical deployment studies.

ABSTRACT

Development of routing algorithms is of clear importance as the volume of Internet traffic continues to increase. In this survey, there is much research into how Machine Learning techniques can be employed to improve the performance and scalability of routing algorithms. We surveyed both centralized and decentralized ML routing architectures and using a variety of ML techniques broadly divided into supervised learning and reinforcement learning. Many of the papers showed promise in their ability to optimize some aspect of network routing. We also implemented two routing protocols within 14 surveyed routing algorithms and verified the efficacy of their results. While the results of most of the papers showed promise, many of them are based on simulations of potentially unrealistic network configurations. To provide further efficacy to the results, more real-world results are necessary.

Motivation & Objective

• To analyze the integration of machine learning techniques into network routing to improve performance and scalability.
• To evaluate centralized and decentralized machine learning architectures for routing in computer networks.
• To assess the effectiveness of supervised and reinforcement learning in optimizing routing decisions.
• To identify limitations in current research, particularly over-reliance on unrealistic simulations.
• To provide empirical validation by implementing and testing two routing protocols from surveyed algorithms.

Proposed method

• Systematic survey of 14 ML-based routing algorithms from recent literature in networking and machine learning.
• Classification of methods into centralized and decentralized learning architectures.
• Categorization of techniques into supervised learning (e.g., regression, classification) and reinforcement learning (e.g., Q-learning, deep Q-networks).
• Implementation and evaluation of two representative routing protocols to validate performance claims.
• Analysis of simulation-based results across diverse network topologies and traffic patterns.
• Comparison of ML-based routing with traditional protocols to assess gains in efficiency, delay, and scalability.

Experimental results

Research questions

• RQ1 How can machine learning improve the performance and scalability of network routing algorithms?
• RQ2 What are the key differences and trade-offs between centralized and decentralized ML-based routing architectures?
• RQ3 To what extent do supervised and reinforcement learning techniques outperform traditional routing protocols in simulated environments?
• RQ4 What are the limitations of current ML-based routing approaches, particularly in terms of simulation realism?
• RQ5 Can real-world deployment of ML-based routing protocols demonstrate consistent performance gains?

Key findings

• Many ML-based routing algorithms show significant improvements in latency, throughput, and load balancing in simulation environments.
• Reinforcement learning approaches demonstrate strong adaptability to dynamic network conditions and traffic variations.
• Supervised learning methods achieve high accuracy in predicting optimal paths when trained on historical network data.
• Centralized learning architectures often yield better performance but face scalability and single-point-of-failure challenges.
• Decentralized learning methods improve robustness and scalability but may converge more slowly to optimal solutions.
• Despite strong simulation results, the lack of real-world validation remains a critical limitation across most surveyed approaches.