Convergence of telemetry-aware neural routing under communication delays

Determine whether neural routing algorithms that consume real-time telemetry data can be trained to converge to high-quality routing policies when communication delays in propagating state information and disseminating or installing actions are explicitly modeled and enforced.

Background

The paper frames near-real-time, telemetry-aware routing as a closed-loop control problem where decisions must be made on millisecond timescales. Prior neural approaches either assume delay-free global state visibility or rely strictly on local telemetry, raising questions about their practicality in real deployments that inevitably involve communication and inference delays.

To move toward realistic conditions, the authors present a simulation framework that models both communication and inference delays and introduce a neural routing algorithm (LOGGIA). While their experiments provide empirical evidence under specific settings, the broader question of guaranteed convergence to high-quality policies in the presence of delays remains explicitly open.