HOMODYNE ALGORITHM X-RAY · HMD-009
01 / 14 · 00:00

Your ETA is a forecast of traffic that doesn’t exist yet.

NOW · 07:30clear

the street in front of you

IN 10 MINqueue?

the junction you will reach

HOMODYNE ALGORITHM X-RAY · HMD-009
02 / 14 · 00:49

One number. One daily decision.

EXAM 08:00Leave now or have breakfast?
RIDE-HAILDriver 7 min · trip 22 min
MEETING 09:00Call ahead or not?

این ۲۲ دقیقه از کجا آمده؟

HOMODYNE ALGORITHM X-RAY · HMD-009
03 / 14 · 01:48 · GUESS FIRST
up to 50%more accurate real-time ETAs in places like Berlin · Jakarta · São Paulo · Sydney · Tokyo · Washington D.C.

Before: accurate for >97% of trips · CIKM 2021 paper: 40+% fewer negative ETA outcomes in cities like Sydney

2020 description of a historical system · today’s private system is not published

HOMODYNE ALGORITHM X-RAY · HMD-009
04 / 14 · 03:05

Two ingredients: history and now.

INGREDIENT 1 · HISTORYMon 08:00

what a normal day looks like

INGREDIENT 2 · LIVEright now

…but you reach segment 6 in 15 min

DeepMind: live data alone doesn’t show traffic 10, 20 or 50 minutes into a drive

HOMODYNE ALGORITHM X-RAY · HMD-009
05 / 14 · 04:12

Traffic jams travel backwards.

R1R2R3R4R5R6R7R8D

The queue grows from the bottleneck toward you. R4 in ten minutes is written in R5–R6 now.

اگر این‌جور باز کردن الگوریتم‌های روزمره به کارت می‌آید، هومودین هر هفته یکی را باز می‌کند.

HOMODYNE ALGORITHM X-RAY · HMD-009
06 / 14 · 05:19

Each segment listens to its neighbours.

R1R2R3R4R5R6R7R8S3S5S7S8DHOME →→ OFFICE (bottleneck D)

message passing · round 1: neighbours · round 2: neighbours of neighbours · supersegment = adjacent segments sharing significant traffic volume

HOMODYNE ALGORITHM X-RAY · HMD-009
07 / 14 · 06:34

Toy city: 12 km · 8 segments · 4 side streets

R1R2R3R4R5R6R7R8S3S5S7S8DHOME →→ OFFICE (bottleneck D)
DAYS 1–45TRAIN
DAYS 46–60TEST (no peeking at the future)

ILLUSTRATIVE · synthetic traffic · not Google data · not Tehran data

HOMODYNE ALGORITHM X-RAY · HMD-009
08 / 14 · 07:44 · GUESS THE RANKING
1 · CONSTANTone average speed for the whole route
2 · HISTORICALsegment average at this time of day
3 · LIVE-NOWfreeze today’s speeds at departure
4 · NEIGHBOUR-AWAREhistory + learned correction from own & neighbours’ live state (2 rounds)
HOMODYNE ALGORITHM X-RAY · HMD-009
09 / 14 · 08:46 · MEAN ABSOLUTE ERROR, MINUTES
constant3.43
historical1.72
live-now1.13
neighbour0.56

ILLUSTRATIVE · rush 07:15–08:45: 3.16 · 2.44 · 1.84 · 0.85 · neighbour best in 20/20 extra seeds

HOMODYNE ALGORITHM X-RAY · HMD-009
10 / 14 · 10:14
07:45 · JAM BUILDING24.0 min actual

live-now 18.1 · historical 21.1 · neighbour 20.6 · constant 16.6

08:30 · JAM CLEARING22.6 min actual

live-now 29.9 (worse than constant 16.6) · neighbour 24.8

ILLUSTRATIVE · “now” fails in both directions

HOMODYNE ALGORITHM X-RAY · HMD-009
11 / 14 · 11:42

We built a world where neighbours matter. Of course they won.

OUR TOY

shows why the idea can work

THE REAL SYSTEM

measured on real trips · GNNs sensitive to training curriculum · MetaGradients adapt the learning rate

HOMODYNE ALGORITHM X-RAY · HMD-009
12 / 14 · 12:44 · CLAIM BOUNDARY
THIS SHOWSAn ETA is a forecast. Live speed alone can mislead in both directions. In a synthetic world, neighbour information cut error.
THIS DOES NOT SHOWHow Google Maps works today (the description is from 2020). Anything about Tehran traffic. The “half the error” number outside our toy.
HOMODYNE ALGORITHM X-RAY · HMD-009
13 / 14 · 13:25 · ANSWER

ETA = history + live + neighbours → a forecast.

DAYS 1–5Note the app’s ETA when you leave
ON ARRIVALNote the real time
COMPAREJam building vs. jam clearing
HOMODYNE ALGORITHM X-RAY · HMD-009
14 / 14 · 14:13 · NEXT

A wrong ETA costs minutes. A wrong AI answer can cost more.

Konkur uses negative marking. Can a penalty teach an AI to say “I don’t know”?