A planner who has to split several drone delivery operators across altitude layers rarely knows where any operator's parcels go. What they can get is public census data: where people live (2020 Census) and where they work (LODES 2022). The last two steps of aerolog-scaling put those maps into a paper estimate of conflicts between operators. Jobs alone overshot one part of the estimate and residents undershot everything. So today's question was the obvious next one: what mix of residents and jobs is right, and does the answer hold for a layout the fit has not seen?
The setup is the programme's usual one: 24 random eight-operator layouts over New York, 2014 ride-hail pickups standing in for deliveries, a 50 m separation radius, 240 flights an hour, and an exact estimator as the reference. Every operator's demand map is a mixture: a fraction a of its jobs map plus 1 − a of its residents map, each normalised over its service disc. The conflict terms are linear in each operator's map, so four map combinations per layout give every mix exactly, in 1.3 s. Each layout's figure comes from a mix fitted on the other 23.
What held (6 of 10 pre-registered predictions):
| before (11v) | fitted mix, out of sample | one scale factor, out of sample | |
|---|---|---|---|
| take-off term, pooled level vs exact | 0.83 | 0.98 | 1.01 |
| take-off term, median log error per layout | 0.20 | 0.13 | 0.13 |
| take-off term, layouts within 5% | 5 of 24 | 3 of 24 | 8 of 24 |
| crossing term, pooled level vs exact | 0.92 | 0.98 | 1.00 |
| crossing term, median log error per layout | 0.068 | 0.093 | 0.096 |
- The best mix is about 89% jobs, 11% residents, and it is stable: held out one layout at a time it ranges only from 0.88 to 0.91 (0.80 to 0.87 for the crossing term).
- It fixes the census estimate's level: the take-off term goes from 0.83 to 0.98 of exact, the crossing term from 0.92 to 0.98.
- It does not fix the per-layout error. A single scale factor on the plain residents-plus-jobs sum does as well (median log error 0.129 against 0.128), and gets more layouts within 5%. Fitting separate mixes for flight paths and for delivery points does not help either.
- On the crossing term, fitting the level makes the typical layout worse (median log error 0.068 to 0.093). The census errors are skewed: a few layouts are 22% to 30% low, and pulling the average up overshoots the rest.
- Layer-split decisions barely move: a mean 2.56% worse than the split the closed form picks from real destination lists, against 2.68%. The mix fixes one bad split (layout 3, 17% to 0.2%), loses the take-off term's earlier fix (layout 5), and makes two small new misses.
What this means for a planner. Census maps can be calibrated to the right overall conflict level with one number per term, and that number barely changes from one layout to the next. Past that, mixing residents and jobs buys nothing. What is left is specific to each layout: where these operators' deliveries actually go. Earlier steps in the programme (11m to 11o) found the same for the crossing term, and found that asking one operator for its real destinations fixes most bad splits. The next step is a richer demand model (land use, floor area, or LODES origin-destination flows), checked the same out-of-sample way.
The mix was fitted against pickups in the same city that define the layouts, so a planner elsewhere would have to bring a from another dataset. The model has straight legs, Poisson launches, no avoidance and no wind. Census counts show where people live and work, not where parcels go. These results are preliminary. Code, tests and results: research/aerolog-scaling, milestone 11w.