NeuroRisk: Physics-Informed Neural Optimization for Risk-Aware Traffic Engineering
Published in arXiv, 2026
This work presents NeuroRisk, a physics-informed deep unrolled optimizer for risk-aware traffic engineering in production WANs with correlated failure scenarios. NeuroRisk exploits the embedded Sort-and-Select structure of risk-aware formulations, enforces feasibility through gated edge-local reservations, and uses permutation-invariant, gradient-aligned cues to represent scenario sets.
Evaluations on production-style WANs show small optimality gaps relative to solver-based approaches with orders-of-magnitude speedups on risk objectives, while outperforming neural baselines on nominal throughput.
Authors: Yingming Mao; Ximeng Liu; Jingyi Cheng; Xiyuan Liu; Jiashuai Liu; Yike Liu; Zhen Yao; Yuzhou Zhou; Siyuan Feng; Qiaozhu Zhai; Shizhen Zhao.