aerolog-scaling asks a practical airspace question: when eight drone delivery operators share a city, which of them should share an altitude layer? The programme answers it with a closed form, a formula for the expected number of losses of separation between every pair of operators. Until today, every check of that formula was against a sampled reference: 400,000 random flight pairs per layout, which milestone 11x found carry about 0.8% noise. That left one question open. How far is the formula from the simulator itself, once the noise is gone?
Today's step (11aa) removes the sampling. The simulator's expected conflict rate is an average over pairs of flights. For each pair, the exact window of launch offsets in which the two drones come within 50 m has a closed-form length (milestone 11r). So the expectation is a finite sum over every pair of destinations: about 1.6 billion interval computations per layout, vectorised in numpy. The eleven predictions were committed before any layout ran.
On 11g's 24 random eight-operator layouts over 2014 New York pickup points:
| result | |
|---|---|
| closed form / exact expectation (en-route, pooled) | 1.0026 |
| spread over layouts (sd) | 0.28% (it was 0.91% against the sampled reference) |
| layouts where the closed form picks the exact best layer split | 24 of 24 |
| largest change in any planner rule's mean loss vs the closed-form reference | 0.09% |
| cost of the exact reference | about 1,000 CPU s per layout (closed form: 50 s) |
Nine predictions held. Two were refuted: the cost (I predicted at most 800 s on average), and the bound I set on the zone-label quadrature. The 4-point sub-rules were up to 1.8% off the 8-point rule. A check written after seeing that showed the 8-point rule matches a 32-point rule to 2e-5, because the two shifted 4-point grids err in opposite directions. The bound was wrong; the numbers were fine.
One bug was worth the test suite. The shortcut "the return legs are the outbound legs run backwards" holds for interval lengths, but not for zone labels when two drones fly the same velocity, for example two flights of one operator to the same address. The estimator then labels the closest approach at the start of the overlap, which is a hub for outbound legs and a delivery point for return legs. The first kernel lost all same-operator en-route weight, and the brute-force test caught it before any real layout ran.
What the residual looks like: the closed form is 0.07% to 2.2% high over pairs of different operators and up to 4.4% low over one operator's own head-on pairs. Those same-operator pairs share a layer under every split, so they never change the decision.
For this question, then, the closed form is the simulator's answer, to well within anything a planner would notice, and the record from 11g to 11z holds on the simulator's own expectation. Limits: it is exact for the model, not for real flights. Straight legs, no avoidance and no wind, one city, ride-hail pickups standing in for deliveries. Results and code: research/aerolog-scaling, milestone 11aa.