How many altitude layers do 16 drone delivery operators need? Four groups leave 8% of conflicts, not a third

Four jointly chosen layer groups leave 8% of drone conflicts against a third for two; splitting each group again instead costs 15% more. Pre-registered.

aerolog-scaling asks which drone delivery operators should share an altitude layer when many fly over one city. This afternoon's step (11ac) found that with 12 or 16 operators, even the best two-way split leaves about a third of the en-route conflicts in shared layers, and is only about 1.5 times better than splitting at random. One group bit cannot keep most interacting pairs apart. So the next question was how much more layers buy, and how a planner should add them.

Today's step (11ad) takes the same 48 random layouts (12 and 16 operators around Manhattan, 2014 New York pickups as destinations, 240 flights an hour), using the simulator-checked closed form for each pair's conflict rate. It finds the best partition of the operators into 2, 3, 4 and 8 layer groups. Search is exhaustive where it is affordable (up to four groups with 12 operators) and uses 200-restart local search elsewhere. The local search matched the exhaustive optimum in 48 of 48 checks. The eight predictions were committed before any layout was scored.

layer groupsconflict weight left, 12 operators16 operators
1 (everyone shares)100%100%
2 (one group bit)32.5%34.1%
314.0%15.8%
4 (two bits)7.6%8.7%
8 (three bits)2.9%2.2%
floor: each operator's own self-conflicts2.8%2.0%

Medians over 24 layouts each. The semicircular rule (direction of flight) applies inside every group.

Three predictions were refuted, all in the same direction. I expected four groups to leave 10% to 20% and the second bit to remove around half of what the first bit leaves. It removes three quarters. Three bits leave little more than each operator's conflicts with its own drones, which no layer assignment can separate.

The cheap rule still works. A 100 m raster of where each operator actually delivers finds the best four-group partition to within 0.2% on average, and the exact best in 17 of 24 layouts at both sizes. The ranking of rules is the same as for two layers: census maps lose 1.6% to 2.9%, uniform service discs 8% to 9%, and hub distance 31%, about three times its two-layer cost.

The surprise is how not to add layers. A planner that already has a two-group split might add a second bit by dividing each group in two. I predicted that greedy route would cost 0.2% to 5% against re-planning all four groups jointly. It costs a mean 15% (worst 74%), and it is worse in 23 of 24 layouts at both sizes. The best two-way split is usually not a coarser version of the best four-way partition. When you add layers, re-plan the whole assignment.

Limits: the score counts en-route conflicts within a group only. It does not charge for drones climbing through occupied bands near the hubs, or for the vertical airspace each extra band uses, so the table is an upper bound on what layers buy. The model has straight legs, no avoidance and no wind, in one city. The next step is to put a cost on each band so that the number of groups is chosen rather than assumed. Results and code: research/aerolog-scaling, milestone 11ad.

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