AI-Driven Traffic Engineering For Large-Scale Wide Area Networks Using Reinforcement Learning

Abstract

Wide area networks that connect large enterprise estates face traffic-optimization pressures that static routing protocols were never designed to absorb, including volatile demand, link degradation, and application-specific quality-of-service requirements. Conventional traffic engineering, built on fixed link metrics in OSPF and BGP or on preconfigured policy in early software-defined WANs, adapts slowly and leaves capacity unevenly used. This paper proposes a reinforcement learning framework that treats wide area traffic engineering as a sequential decision problem and drives path selection through a centralized software-defined controller. The network state is assembled from streaming telemetry, flow records, and quality-of-service counters. An agent maps this state to rerouting actions based on a reward that favors throughput while penalizing latency and loss. The controller enforces the selected actions on edge devices within a bounded action space. The framework was evaluated on an emulated dual-data-center topology with fifteen branch nodes and injected congestion. Measured against a static baseline, the learned policy reduced latency by approximately 25 to 30 percent, lowered packet loss by around 20 percent, and raised throughput by roughly 15 percent while balancing load across parallel paths. Results suggest that constrained reinforcement learning is a practical route to adaptive, operator-supervised traffic engineering at enterprise scale.