Census jobs data tells you which drone operators will conflict. It doesn't yet tell you how to split them

Weighting drone customers by census jobs makes paper conflict weights match the simulator (r 0.99, was 0.96); jobs plus residents match the split's median.

This morning's step (milestone 11i) ended on a guess. Spreading each delivery operator's customers over New York's land, rather than its whole service disc, barely helped a paper formula choose which of eight drone operators should share an altitude layer. The missing information seemed to be how densely customers sit on land. The cheapest public map of that is the census.

So we weighted each operator's destinations by two Census Bureau datasets:

  • where people live: 2020 population at each block group's population centroid (3.19 million residents inside our 10 km window);
  • where people work: LEHD LODES 2022 job counts at each census block (3.06 million jobs in the same window).

Everything else was the same as before: the same 24 random eight-operator layouts, simulator, kinetic-gas integral and scoring. We wrote nine predictions down and pushed them before computing a single weight.

What jobs fixed. The paper weights now rank which operator pairs conflict most almost exactly as the simulator does. The median correlation is 0.990, against 0.964 with the land mask. The worst layout is 0.948, against 0.692. The scale is right too: total predicted conflicts are 1.01 times the closed form's, against 0.74 with the land mask and 0.54 with uniform discs. Residents alone did not help (r 0.965). At block-group resolution, where people live carries little that the land mask didn't already.

What it did not fix. The two-layer split those weights choose is barely better. Its median is 1.2% more en-route conflicts than the closed form's pick, against 1.9% with the land mask. It is within 5% of the closed form in 17 of 24 layouts, the same as before. Jobs repaired three layouts where every earlier paper weight had been 15% to 24% off, but broke three others. Two layouts are missed by every demand map we tried, so their error comes from something other than demand. Five of the nine predictions failed.

The best paper weight so far: residents plus jobs. Treating each flight as going either to a home or to a workplace, with the counts simply added, gives a split whose median matches the closed form (excess 0.0%). It is within 5% in 18 of 24 layouts and picks the simulator's best split in 12 of 24. It was never worse than jobs alone and better in 6 layouts (sign test p = 0.03). We did not pre-register that comparison, so treat it as a lead.

What a planner can take from this. Public census data is enough to work out which operators' traffic will collide, without any traffic records and in half a second per layout. Choosing one split by max-cut on those weights is still fragile. Our untested guess is near-ties: several splits score almost the same, so a small error in one pair weight changes the choice and costs 5% to 20% out of sample. The next step tests that guess. A planner would pick the split that stays good when the weights are perturbed, rather than the single max-cut.

Straight legs, no wind, no avoidance; 2014 ride-hail pickups stand in for deliveries. Service radii, speeds and market shares are swept, not taken from real operators. One city, 24 layouts; preliminary. Code, data extract and full results: research/aerolog-scaling, experiments/hub_airspace/demand_weights.py.

All writing · More in Research findings · More on aerolog-scaling · All tools and research