Traffic-signal control in Midtown Manhattan
Sixth & Seventh Avenues, West 25th–30th Streets, 12 intersections
Experiments & control approaches
Both experiments use the same street network and compare against Tuned fixed-time: a repeating 90-second schedule with 52 seconds of avenue green, 30 seconds of cross-street green and eight seconds of clearance. Neighboring signals are timed to help vehicles progress along the avenues.
Original controllers
Can decisions at individual intersections improve on that fixed schedule? Morning and surge demand; 50 recorded runs.
- Actuated
- Switch when the other road is waiting and the current approach empties or its green reaches 40 seconds.
- Queue pressure
- Favor the approach with more queued demand, accounting for congestion on the roads it feeds.
- Jev (native)
- Every five seconds, Jev chooses whether each eligible signal holds its green or switches.
- Bounded Jev
- Jev can transfer five seconds of green between roads for one cycle. The cycle length and coordination stay fixed.
Coordinated plans
Can choosing one plan for all 12 intersections help? Plans change green splits while keeping the 90-second cycle and coordination. Morning, surge and changing demand; 60 fresh runs.
- Retuned fixed plan
- One schedule chosen offline, with different green splits at different intersections. It stays fixed throughout the run.
- Numerical plan selector
- Every 90 seconds, choose the plan with the lowest predicted queue cost from five candidates.
- Coordinated Jev
- Jev chooses from those same five plans using traffic observations and forecasts. Changes begin at each signal’s next eligible cycle.
Each table compares matched traffic within its experiment; the experiments use different random seeds. Actual Jev inference time is included. Cars travel straight and turn, with turn probabilities informed by the NYC counts; use 1× playback to see the turns more easily.
(a) Tuned fixed-time
90 s cycle1. Results
| Controller | Delay / trip (s) | Mean trip (s) | 95th percentile (s) | Mean queue (vehicles) | Completed (trips) | Unfinished (trips) |
|---|
Notes. Bold marks the best mean in each column, including ties: lower is better except for completed trips, where higher is better. Winners use unrounded values; bold does not imply statistical significance. Delay includes waiting to enter and unfinished trips. Travel-time statistics cover completed trips only. Completion counts include the three-minute drain. Each controller receives identical traffic inputs for each seed.
Technical checks
2. Data & scope
Weekday morning turning volumes cover all 12 intersections, from the Midtown South study’s June 2024 collection campaign. Internal traffic flows balance. Roads are schematic; lanes, speeds and signal plans are assumed.
Field validation remains outstanding. Pedestrians, bicycles, bus stops and curbside obstruction are not modeled. No comparison against independently observed NYC travel times or queues has been completed.
3. Experimental method
Five controllers share the same traffic inputs. The fixed plan was selected from 16 candidates on separate development seeds. Each run contains three minutes of warm-up, ten minutes of demand and three minutes of drain.
Native Jev chooses hold or switch. Code enforces green durations of 22–60 seconds, three seconds of yellow and one second of all-red. Requests have a two-second deadline; late decisions are discarded and a queue-pressure rule provides fallback.
Bounded Jev selects a green split once per cycle: 47/35, 52/30 or 57/25 seconds for avenue/cross street. The 90-second cycle, coordination offsets and clearance times stay fixed. Changes expire each cycle; late or failed requests retain the baseline split. The original Jev controller and its recorded runs are preserved.
Confidence intervals describe variation across five paired seeds. They do not include all uncertainty in traffic behavior or establish effectiveness on actual streets.
3. Experimental method
Four controllers share fresh traffic seeds 31–35. The original fixed plan remains the reference. A retuned fixed plan was selected from seven profiles on separate development seeds. The numerical selector and Jev choose from the same five-plan library.
Every 90 seconds, both selectors receive current traffic, two minutes of arrival history and a 180-second queue forecast for each plan. Jev evaluates the candidates without an automatic preference for the current plan. Neither selector sees future departures or the scheduled demand changes.
Each chosen split installs at the next eligible local cycle start. The 90-second cycle and original coordination offsets stay fixed; green durations can vary from 22 to 60 seconds. Yellow and all-red remain three and one seconds. Actual forecast preparation, API latency and installation waits are included.
Each run contains three minutes of warm-up, ten minutes of demand and three minutes of drain. The directional-demand case is synthetic. Five paired seeds provide a pilot comparison; confidence intervals omit uncertainty in the physical model. The retuned fixed and numerical arms help distinguish plan design from Jev’s contribution.