AI-driven routing pipeline in software-defined networks using DQL: a mini review.

Area of Science:

• Computer Science
• Network Engineering
• Artificial Intelligence

Background:

• Data center networks face challenges with dynamic traffic, leading to inefficiencies like latency and congestion.
• Traditional routing algorithms (Dijkstra, ECMP) lack real-time adaptability in Software-Defined Networking (SDN).
• Reinforcement Learning (RL) offers adaptability, but scalability issues are present, leading to the development of Deep Q-Learning (DQL).

Purpose of the Study:

• To review and synthesize recent Deep Q-Learning (DQL) approaches for Software-Defined Networking (SDN).
• To examine DQL architectures, algorithmic variants, and emulation environments.
• To provide a structured taxonomy and practical synthesis of trade-offs and deployment issues for DQL in SDN.

Main Methods:

• Review of recent Deep Q-Learning (DQL) approaches applied to Software-Defined Networking (SDN).
• Examination of network architectures, algorithmic variations, and emulation environments like Mininet with Ryu.
• Analysis of empirical trade-offs, focusing on throughput, latency, and convergence.

Main Results:

• Reported studies indicate DQL improves throughput by 15-22% and reduces latency by 10-12% compared to ECMP.
• DQL enables SDN controllers to learn routing strategies from live network states using neural networks.
• Trade-offs include increased training time, inference delays, and persistent scalability challenges.

Conclusions:

• DQL presents a promising approach for adaptive routing in SDN, offering significant performance improvements.
• Current DQL implementations face hurdles in training efficiency and scalability, limiting immediate production readiness.
• Emerging directions like federated learning, graph neural networks, and explainable AI are key to advancing DQL for practical SDN solutions.