This is an earlier experiment. Read the amendments below before treating its findings as current.
The question
The margin for old news is a hand formula. Logged data could show how wrong the straight-line guess of a neighbour's position really is.
What was tested
We logged 15.2 million samples of guessed against true neighbour position from 420 cases. A model of the 95th-percentile error went into the simulator and was compared with the hand margin on 84 slow-link cases over four layouts.
What was found then
The learned margin fitted the errors more evenly and was smaller on average, but collisions rose from 44 to 75. Collisions come from the rare large errors, when a neighbour manoeuvred since its last message. That worst case is exactly what the hand formula assumes, and a fitted average smooths it away. Taking the larger of the two margins gave 45: better in turbulence (24 to 11), worse at speed (6 to 21). Both were rejected.
What changed
Nothing (the hand margin stays).
Corrections retained in the record
Earlier wording: Margin variant 2 ... seed 1 done, seeds 2–3 queued last
Record correction: All four seeds ran: 44 → 45, rejected.
README.md:792; git e11d68b
Numbers and source references
- 420
- logged casesREADME.md:792
- 15.2 million
- samplesREADME.md:792
- 84
- cliff casesFlight Plan Rev D-cont; README.md:659 (84-case cliff set)
- four seeds
- seedsREADME.md:792
- 44 → 75
- cliff set, learned marginREADME.md:792; git f11f42c
- 44 → 45
- cliff set, larger of the twoREADME.md:792; git e11d68b
- 24 → 11
- turbulent row, larger of the twoREADME.md:792
- 6 → 21
- speed-16 row, larger of the twoREADME.md:792
- 95th-percentile
- quantile fittedREADME.md:792