Carnegie Mellon's TANGCO AI stops network cascades better than heuristics

Researchers from Carnegie Mellon University (Orkun Irsoy, Leman Akoglu, and Osman Yagan) introduced **TANGCO** (Topology-Aware Neural Graph-Guided Capacity Optimization), a machine learning framework designed to prevent cascading failures in networked systems such as power grids, cloud clusters, and traffic networks. Unlike traditional heuristic methods—which often rely on static, hand-crafted rules—TANGCO uses a graph neural network (GNN) policy trained via policy-gradient learning to dynamically allocate capacity based on network topology and load redistribution patterns.

In evaluations across 450 synthetic graph instances and 45 real-world networks (including power grids, road networks, air traffic, and internet topologies), TANGCO outperformed four state-of-the-art heuristic methods in 440 cases, achieving robustness gains ranging from 1.6% to 246%. Notably, the model generalizes effectively: pre-trained variants (TANGCO$^{pre}$) matched per-network training performance on unseen real networks, and training scaled near-linearly with graph size. The GNN-based approach also revealed critical insights into when local risk assessment suffices versus when topology-aware policies are essential, even enabling improvements to existing closed-form heuristics.