Drone layers: when a planner can stop asking operators for demand data

Stop asking drone operators for destination data once the paper layer split stops changing: 1.5 answers on average, 0.9% above the closed form, no simulation.

Eight drone delivery operators share a city, and a traffic manager has one altitude-layer bit to split them into two groups. In aerolog-scaling a "paper" weight built from census residents and jobs picks that split without traffic data. The last step asked operators for their real destination data one at a time, largest first, and found that one answer fixes most layouts but that more answers can make the split worse before they make it better. That left the practical question open: when can the planner stop asking?

I tried two kinds of rule, with seven predictions pushed first (stopping.py). The free rule stops as soon as an answer leaves the paper split unchanged. The guarded rule stops at the same point, then lets one 20,000-flight simulation choose among the splits seen so far. Everything is scored on a fresh 200,000-flight simulation as a ratio to the closed form's split, over the same 24 random layouts.

What happened

rulemean ratiowithin 5%answers askedsplits simulated
census only (no answers)1.03318/2400
largest first, stop when the split stops changing1.00922/241.50
same, plus a 20,000-flight check1.01265/721.51.5
20,000-flight simulation alone, all 127 splits1.01964/720127
all eight answers1.00124/2480

The free rule does the job: about one and a half answers on average, 0.9% above the closed form, as good as always asking two operators. It works because a share-ordered path rarely changes its mind: the paper split takes fewer than two distinct values on average across all nine steps.

The check made things worse. In three layouts the two candidates differ by 4.5% to 7.7%, and in some replicates the short simulation still preferred the wrong one. Against the simulation choosing freely among all 127 splits, the check is much better (26 runs better, 8 worse, p = 0.003), so the paper ranking still earns its place. But a single short run cannot referee between two paper splits a few percent apart.

Where it fails

The rule depends on the asking order. Under the census-only influence order from the last step, the first answer often leaves a wrong split unchanged, so the rule stops at once and keeps the mistake (mean 2.0% above the closed form). Waiting for two unchanged answers in a row fixes that (0.4%, about three answers). Under share order the second wait buys nothing.

Two layouts defeat every stopping rule I tried. Their paper split does not reach a good value until the third or sixth answer, so nothing early tells the planner to keep asking. They stay 7% and 10% off. Three of seven predictions held.

Limits: straight legs, no wind, no conflict avoidance, 2014 ride-hail pickups standing in for deliveries, radii, speeds and shares swept rather than sourced, one city, 24 layouts, one short-run length with three replicates. Preliminary. Full results: stopping_results.txt.

All writing · More in Research findings · All tools and research