What is it about?
Cloud companies move huge amounts of data between distant datacenters and must split traffic to avoid overload. Instead of predicting future traffic, we train an AI model to directly decide how to route it from recent traffic observations. Tested on real networks, our theory-grounded approach moves more traffic and reacts much faster than traditional methods.
Featured Image
Photo by Green Prophet on Unsplash
Why is it important?
Backbone networks typically rely on forecasting future traffic and optimizing for predicted traffic. Our work removes traffic forecasting entirely, using modern AI to directly learn high-quality routing decisions from real data and optimize for network performance itself. This is timely as traffic grows more volatile and AI tools have matured for large-scale deployment. The result is a practical method that carries more data and reacts much faster than traditional approaches, improving the efficiency and reliability of cloud services.
Read the Original
This page is a summary of: Learning to Flow (Between Datacenters), Communications of the ACM, February 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3765706.
You can read the full text:
Contributors
The following have contributed to this page
