~/projects/ai-systems/stale-but-safe-te.py
Research PaperThe Cost of Stale Telemetry in Traffic Engineering
Does a learned policy beat a re-tuned heuristic when link counters arrive late, noisy and incomplete? A research paper with an audit tool.
- ❯A learned split-ratio policy trained through a differentiable fluid model, tested zero-shot on 24 Topology Zoo networks, replayed SNDlib traces and 27 Mininet TCP scenarios
- ❯Clean or moderate telemetry: the re-tuned heuristic matches or wins. Delay, noise and loss together: the learned policy leads by +0.033 dScore, but only 0.006 at packet level
- ❯An ECMP-fallback guard halves tail regret in simulation but not at packet level, and fluid features, not message passing, drive the result
Technologies & Frameworks
Core / Frontend
Python🧠GNNHTML5LaTeX
Backend / Engine
🌐Mininet + OVS🛰️Ryu SDNDocker
Core / Frontend
Backend / Engine
1 / 7
The Cost of Stale Telemetry in Traffic Engineering
Does a learned policy beat a re-tuned heuristic when link counters arrive late, noisy and incomplete? A research paper with an audit tool.
- ❯A learned split-ratio policy trained through a differentiable fluid model, tested zero-shot on 24 Topology Zoo networks, replayed SNDlib traces and 27 Mininet TCP scenarios
- ❯Clean or moderate telemetry: the re-tuned heuristic matches or wins. Delay, noise and loss together: the learned policy leads by +0.033 dScore, but only 0.006 at packet level
- ❯An ECMP-fallback guard halves tail regret in simulation but not at packet level, and fluid features, not message passing, drive the result
Technologies & Frameworks
Core / Frontend
Python🧠GNNHTML5LaTeX
Backend / Engine
🌐Mininet + OVS🛰️Ryu SDNDocker
Core / Frontend
Backend / Engine