Unsensed spots · accuracy
–
Probability error · Brier
–
How much do sensors improve the map?
Your parking options official capacities · inferred availability
Data sources & assumptions
Cached official geometry and capacity. Spot positions estimated.
Occupancy is simulated with fixed seeds. Accuracy is measured only on unsensed spots; observations are excluded from that score.
Model: spots fill roughly in order of attractiveness, so a few sensed spots pin down one fill level per lot. Naive baseline: every unsensed spot gets the occupied share of the sensed ones.
Nearby lots share demand. Put sensors on one whole lot and its neighbors' estimates move too, with uncertainty that grows with distance.
Uncertainty about each lot's fill± fraction of the lot
Checked on real data: Birmingham, UKREAL
Compare simulated 12-spot sensor clusters. With the same observation budget, where sensors sit changes how much of the lot the model can infer.
Brier error on unsensed spotslower is better
Spots by the entrance are almost always taken, so sensors there keep saying "full" and tell you little about the rest. Sensors near the edge where the lot usually stops filling carry the most information.
Live location & data
GPS is processed here, never saved by Parkfly. Basemap providers receive map-area requests.
GPS updates estimated trip cost → sugar / bitter / looming → MN9 / MDN / Giant Fibre. Parking occupancy is simulated.
Finds a free spot—
Walk + full-lot penalty—
Fly decision accuracySIMULATED TEST
How accuracy was measured
Compare lot responses
The brain, firing
Recorded spikes from a full-brain run with tastes like this lot's (three recorded cases; the nearest one plays).
Does the fly park well?minutes lost, lower is better
Real fly-brain network
Simulator settings & libraries
Download the reproducible Colab notebook ↗ · Model report & measured results · Live demo, video & Q&A
Where the fly's answers come from