Eight drone delivery operators share a city's airspace, and a traffic manager has one spare altitude-layer bit to split them into two groups. Over the last few steps of aerolog-scaling I built a "paper" weight that ranks the 127 possible splits from hub sites, service radii, speeds and census counts, with no traffic data. The last step found that the simulator's best split sits in the paper's top five in 21 of 24 random layouts, and proposed a procedure: rank on paper, then simulate only the top five.
This step priced that procedure. I wrote eight predictions down and pushed them before running anything (sim_budget.py), then ran short simulations of 5,000 to 50,000 flights (three seeds each) against the full 200,000-flight run, choosing either among all 127 splits or only among the paper's top 3, 5 or 10. Each choice is scored on a separate 200,000-flight run, as a ratio to the closed form's split.
What happened
| procedure | mean excess over closed form | within 5% | CPU per run |
|---|---|---|---|
| paper argmax, no simulation | 3.3% | 18 / 24 | 0.53 s (paper integral) |
| best of 127 on 5,000 flights | 3.8% | 53 / 72 | 0.10 s |
| best of paper top 5 on 5,000 flights | 1.9% | 61 / 72 | 0.10 s |
| best of 127 on 20,000 flights | 1.2% | 67 / 72 | 0.31 s |
| best of paper top 5 on 20,000 flights | 1.1% | 66 / 72 | 0.31 s |
| best of paper top 10 on 20,000 flights | 1.0% | 67 / 72 | 0.31 s |
| best of 127 on 50,000 flights | 0.8% | 69 / 72 | 0.76 s |
| best of paper top 5 on 50,000 flights | 1.3% | 65 / 72 | 0.76 s |
The paper ranking is a useful prior only when the simulation is too short to trust alone. At 5,000 flights it halves the loss (20 runs better, 7 worse, sign test p = 0.02). By 20,000 flights it is a tie, and at 50,000 it is worse, because the top five exclude the best split in 3 of the 24 layouts and no amount of simulation can recover it. Three of my eight predictions were wrong, including that the top-five procedure would reach 1% by 50,000 flights and that a short list of three would beat a list of ten.
The surprise is the price
In this simulator, conflicts between flights are counted pair by pair, so one run scores every split at once. A 20,000-flight run takes 0.31 CPU seconds: less than the census-weighted paper integral (0.53 s) and a fifth of the closed form (1.6 s), and it lands a mean 1.2% from the closed form's split against the paper's 3.3%. With a simulator like this one, a planner should just run a short simulation.
The paper ranking earns its keep when each candidate split needs its own run, as it would in a simulator with conflict resolution. There, scoring the paper's top 10 on 20,000 flights each costs about 3 CPU seconds for a 1.0% mean excess, against 39 seconds to score all 127. That is a 12x saving at the same quality. It is a multiple of this simulator's run time, not a measurement on a resolving simulator. Running one is the obvious next step.
Limits: straight legs, no wind, no conflict avoidance, 2014 ride-hail pickups standing in for deliveries, operator radii, speeds and shares swept rather than sourced, one city, 24 layouts, three seeds per budget. Preliminary. Full results: sim_budget_results.txt.