ML-Driven Context-Aware Adaptive Routing for Multi-Objective Optimization in Software-Defined Networks

Abstract

Software-Defined Networking (SDN) enables centralized and programmable network control, making it highly suitable for data-center and edge-cloud environments where traffic demands change rapidly. However, conventional routing methods such as shortest-path routing and Equal-Cost Multi-Path (ECMP) forwarding are often unable to respond effectively to fluctuating congestion, resulting in higher latency, uneven link utilization, and reduced fault tolerance. This paper proposes an ML-driven context-aware adaptive routing framework for SDN using multi-objective reinforcement learning to improve routing decisions in real time. The proposed system integrates a lightweight learning agent within the SDN controller and uses telemetry such as link utilization, packet loss, delay, and failure probability to select efficient routes dynamically. A multi-objective reward function is designed to jointly optimize latency, load balancing, and resilience, while maintaining low computational overhead through a compact state representation. The model is implemented in a Mininet-based SDN environment and evaluated against OSPF shortest-path routing, ECMP, and a single-objective ML-based baseline. The results indicate improvements in latency, link utilization balance, packet loss reduction, and recovery time after link failures. The framework provides a scalable and modular solution for intelligent routing in next-generation SDN infrastructures.