aerolog-scaling asks a practical airspace question. When eight drone delivery operators share a city, which of them should share an altitude layer? Put the wrong two operators together and their drones meet more often. The programme has a closed form for the expected conflicts between every pair of operators, and this morning's milestone (11aa) checked it against the simulator's exact expectation: 0.26% apart, and the same best layer split in 24 of 24 layouts.
A planner will not always run that closed form. It costs about 100 CPU seconds per layout and needs every operator's full destination list. So the earlier milestones compared cheaper rules: hub distance (the rule you could apply with no data at all), service-disc overlap, census maps, and a "paper weight", a 100 m raster of where each operator actually delivers. All of those comparisons used the same 24 random layouts. A ranking found on 24 layouts, and a census blend fitted on them, could be tuned to those 24.
Today's step (11ab) draws 72 fresh layouts from the same generator (random hub sites within 7 km of a Manhattan hub, service radii 1.5 to 8 km, speeds 12 to 28 m/s, random market shares, 2014 New York pickup points as destinations). It scores 13 rules on them against the simulator-checked closed form. The blends are refitted on the old 24 layouts only. The nine predictions were committed before any new layout ran.
| rule (what the planner needs) | mean loss on 72 fresh layouts | best split found |
|---|---|---|
| destination raster + take-off term | 0.09% | 66 of 72 |
| destination raster (paper weight) | 0.15% | 65 of 72 |
| census maps (residents and jobs), 5 variants | 1.4% to 2.0% | 43 to 48 of 72 |
| uniform service discs | 5.8% | 32 of 72 |
| hub distance | 14.4% | 9 of 72 |
| demand overlap / density heuristics | 20% / 28% | 7 / 4 of 72 |
Loss is the extra en-route conflicts of the rule's split over the best of the 127 two-layer splits. For scale, the average split leaves twice the best split's conflicts (median; up to 7.5 times), so the choice matters.
Eight of nine predictions held. The tiers keep their order out of sample. Every paper-weight mean is within 0.02 of its value on the old layouts, and all 240 of the old layouts' cells reproduce to the last printed digit. The destination raster's seven misses sit mostly on near-ties, where the best two splits are within 0.7% of each other. Its two real misses (4.0% and 4.2%) are not. The take-off term fixes one of them.
The refuted one is a small, useful surprise. I predicted that 11w's census blend (a fitted mix of jobs and residents) would gain nothing over the plain census weight. Refitted on the old 24 layouts, it beats the plain weight by 0.54% on the new 72, just past the 0.5% I allowed. On the old layouts it had been slightly worse. The census rules also all do better on the new layouts than on the old ones, by 0.9% to 1.9%. The original 24 were a hard draw for census maps.
What a planner can take from it: if you have the operators' delivery destinations, a 100 m raster of them is as good as the full closed form for choosing layers, at about a thirtieth of the cost. With census maps alone, expect to lose 1% to 2%. With nothing but hub positions, expect 14% on average, and more than 10% on almost half of layouts.
Limits: the reference is the closed form, which was checked against the simulator on the old 24 layouts, not on these 72. The model has straight legs, no avoidance and no wind, in one city, with ride-hail pickups standing in for deliveries. Results and code: research/aerolog-scaling, milestone 11ab.