When does better network information lead to better decisions?

The problem

A routing change moves load between paths. A controller must choose using information about the current state and a model of the response to that change. Prediction accuracy alone does not tell us how either factor affects the cost of the selected action.

Why this matters to me

Following the systems work in DT4N, this project examines the decisions made with a network model. The goal is to understand the separate and interacting effects of state information and response-model fidelity within a specified decision setting.

Research question

For load-changing routing actions, how do improvements in current-state information and action-response model fidelity differ and interact in their effect on expected decision regret?

My role

I work on queueing-based models, simulation, and the evaluation workflow for this ongoing study. The repository records experimental designs, exploratory analyses, diagnostic results, and their limitations.

Research Advisor

Dr. Huynh Van Dang
Faculty of Computer Networks and Communications, University of Information Technology, VNU-HCM

Academic homepage

Technical approach

1. Use a controlled two-path queueing model in which switching a managed load changes the load on both paths.
2. Compare four decision pipelines under the same plant, action set, objective, and probability law, varying state information and response modeling separately.
3. In the specified model comparison, hold parameters and posterior information fixed; compare mean-rollout costs with true expected future costs.
4. Evaluate expected decision regret against a reference with current-state and response-law information. The reference does not know realized future traffic. Preserve predictions, configurations, and result artifacts for inspection.

Results & current status

The repository records exploratory and diagnostic results. Current status is ongoing research; no publication, novelty claim, or confirmed hypothesis is asserted here.

Limitations & open questions

• The Phase 6 result is inconclusive and calls for reformulating the predictor. Phase 7 remains exploratory.
• Numerical and model checks do not by themselves establish research validity or real-network generalization.
• The current-state benchmark includes hidden-mode oracle information. It is not a deployed telemetry refresh mechanism, and it does not establish that fresher telemetry is harmful.
• The model comparison has a fixed scope: shared parameters and posterior, with different representations of future cost. It does not test arbitrary parameter or structural errors, or establish a general ranking of state information versus model fidelity.