infocom-experiments

INFOCOM Experiments

Use this before submission when the evaluation is not yet locked. INFOCOM reviewers span the full

analysis → simulation → testbed spectrum, and the evaluation is where a good idea is won or

lost — especially since there is (traditionally) no rebuttal to repair a weak comparison. The

organizing principle is evidence proportional to the claim, delivered by the method the claim

demands: a bound needs a proof, a design needs a fair experiment, a behavior claim needs

measurement.

Evaluation audit

• Match the method to the claim. An optimality/approximation claim needs a proof under

stated assumptions; a performance claim needs a fair experiment; a real-world behavior

claim needs measurement on real traffic. A simulation cannot substitute for a proof, and a

proof under unrealistic assumptions cannot substitute for an experiment.

• State and justify every assumption. For analytical work, list the assumptions (A1, A2, ...)

and defend each; a theorem true only under an assumption no network satisfies is a scored

weakness a reviewer will name — and you cannot rebut it.

• Name the simulator and pin the setup. State whether it is ns-3, OMNeT++, a custom simulator,

or a testbed; report the topology, traffic model, number of runs, seeds, and confidence intervals.

"A large-scale simulation" with no setup is not evidence.

• Choose honest baselines, including the strongest prior scheme and a simple-but-reasonable

alternative, each tuned with a documented, equal budget. An untuned baseline is the classic

INFOCOM reject.

• Report variance, not point estimates. Confidence intervals over multiple seeded runs;

appropriate statistics; say what the intervals represent.

• For testbeds/measurement, describe the hardware, the trace or workload, the collection

methodology, and the confounds — and bound them.

Claim-to-evidence design table

| Networking claim | Matching evidence | Reject pattern avoided |

|---|---|---|

| "Algorithm is near-optimal" | Proof of an approximation/competitive ratio under stated assumptions | "Only shown on examples the authors chose" |

| "Scheme improves throughput/delay" | Seeded simulation or testbed vs. a tuned baseline, with CIs | "Untuned baseline; single run; no variance" |

| "Scales to large networks" | Runtime/quality across realistic topology sizes | "Only a 10-node topology tested" |

| "Holds in real deployments" | Testbed or trace-driven measurement with real traffic | "Synthetic traffic claimed to generalize" |

| "The learner adds the value" | Ablation vs. a non-ML heuristic on the same inputs | "ML component's marginal value never isolated" |

Simulation and reproducibility floor

[Simulator]  name it (ns-3/OMNeT++/custom); pin the version; describe the model, not just the tool
[Topology]   real or realistic (e.g., from a topology dataset); state size and generation method
[Traffic]    the workload/trace and its source; synthetic models justified against real behavior
[Runs/seeds] multiple seeded runs; report the number and log the seeds; give confidence intervals
[Baselines]  the strongest competitor + a simple one, each tuned with a documented equal budget
[Compute]    the runtime/scale actually reached, not vague feasibility language

Analytical-work floor

• List assumptions explicitly and justify each against real network conditions.
• Give proof sketches in the body and full proofs in a tight in-budget appendix (it counts toward

the nine pages) or state where the full proof lives.

• Validate the theory in simulation: show the bound is tight/loose where the analysis predicts.

Vignette: evaluating a caching/offloading policy

Suppose the paper claims an online caching policy with a competitive ratio that also beats prior

heuristics. The matching plan: prove the ratio under stated assumptions; then simulate on a real

request trace with multiple seeds, comparing against the strongest prior policy tuned with an equal

budget and a simple LRU baseline; report hit-rate and delay with confidence intervals; show where

the empirical gap matches the analytical bound; and state the regime (skewed vs. uniform demand)

where the guarantee is loose — every number traceable to a logged run.

Statistical reporting floor

• Confidence intervals over seeded runs for every simulated/measured comparison; say what they

represent.

• The number of runs and the source of randomness for any stochastic component.
• The topology sizes and traffic models actually used, not vague "large-scale" language.

Output format

[Evaluation readiness] strong / adequate / weak
[Claim -> method map] <claim: proof | simulation | testbed/measurement>
[Assumptions] <listed and justified? yes/no>
[Simulator/setup] <named? seeds logged? CIs reported? yes/no>
[Baseline fairness] <baseline -> tuned? equal budget? documented?>
[Decision-critical next run] <one experiment or proof to add before the deadline>