Asphalt or algorithms

Give a mid-sized city one million dollars to fight congestion. What buys more: three hundred metres of new road, one extra roundabout, or smarter software at every traffic light? It is a fair question, because the sums involved are absurdly mismatched. Road-building is priced in millions per kilometre, while re-timing a signal is priced in engineer-hours. This article checks what the software answer has actually delivered in Pittsburgh, Zurich, Manhattan and a dozen other cities, using field measurements wherever they exist, and saying so plainly where they don't.

The dumbest machines on the street

Most of the world's traffic signals still run fixed-time plans: pre-programmed green splits derived from historical averages, switched by time of day. A signal like that knows exactly three things:

• the clock,
• perhaps a push-button,
• perhaps an induction loop in the asphalt.

It does not know how long the queue is, what happened two intersections upstream, or that a bus with fifty passengers is about to arrive. So drivers idle at red lights facing empty cross streets, and a hiccup at one junction propagates unseen to the next. None of this is a hardware problem. The information needed to do better has existed for decades; the question has always been whether anything collects it and acts on it.

The generic fix is adaptive signal control: sensors feed actual demand to software that re-times the lights continuously. What has changed recently is not the idea but the sensing. And, as we will see, the most interesting new sensors are not bolted to poles at all. They are driving around the city in regular service.

Pittsburgh: the intersection that thinks for itself

The most-cited modern case is Surtrac in Pittsburgh, developed at Carnegie Mellon University. Its architecture is the interesting part. There is no central brain. Each intersection runs its own scheduler, treats arriving vehicles as an optimisation problem, computes a signal plan seconds ahead, and then tells downstream intersections what traffic it is about to send them. Coordination emerges from neighbours talking. That is why the system scales by adding intersections rather than upgrading a data centre, and why one failed junction degrades gracefully instead of taking the network down.

The famous numbers come from the 2012 pilot in East Liberty, nine intersections: travel times fell by about 25 per cent, waiting time at signals by about 40 per cent, stops by 30 per cent, with an estimated 21 per cent cut in emissions. Two honest footnotes belong next to those figures. First, they describe a nine-intersection pilot corridor, not a city. Second, they were measured by the research team itself: competently and transparently, but never independently audited. Pittsburgh has since grown the network to roughly fifty intersections and is expanding along its "Smart Spines" corridors, a $28.8 million programme covering 135 signals. That is itself a data point on cost: on the order of $210,000 per intersection, all-in, versus millions for any answer made of concrete.

Forty years of adaptive control — and what the field actually measured

Surtrac's press coverage can leave the impression that adaptive signals were invented in Pittsburgh. They were not. Sydney's SCATS has adapted signal timing since the 1970s and now runs in dozens of countries; London's SCOOT followed in the 1980s; Los Angeles built ATSAC for the 1984 Olympics and kept extending it until every signal in the city was adaptive. In this atlas alone, Bengaluru has put 169 junctions under adaptive control, Wiesbaden's DIGI-V rebuilt the sensing layer of an entire city, and Bangkok is rolling out AI-adaptive junctions. Adaptive control is not an experiment. It is mature municipal engineering with an unusually good publicity department.

Maturity also means the results are measurable, and the measured results are more sober than the pilots. The pattern across four decades: pilot corridors post spectacular numbers, city-scale systems post useful ones. The difference is not fraud but statistics. Pilots are placed where signals are worst, so they harvest the easiest gains first.

That third bar deserves its own paragraph, because nobody puts it in the brochure. Manhattan's Midtown in Motion, adaptive control across 110 and later 270 blocks, measured its roughly 10 per cent improvement with the E-ZPass readers already mounted over its avenues. The elegant trick came from outside: a peer-reviewed re-analysis of six years of taxi GPS data, the city's cab fleet as a floating measurement network, found the gains held for the first year or two and then dissolved as traffic grew and conditions drifted. Software infrastructure decays like asphalt. Faster, in fact, and more quietly: it decays in models rather than potholes.

The fleet is the sensor

Which raises the practical question every city asks first: where does the data come from? Cameras and radar on poles are the textbook answer, and the expensive one. The more elegant answer has been hiding in plain sight since the 1970s: the vehicles a city already operates on fixed routes, all day, every day.

Zurich wrote the template. After voters rejected an underground metro, a 1977 referendum committed 200 million francs to making surface trams and buses fast instead. The result, built out over decades: more than 3,000 detectors, every transit vehicle announcing itself to the network roughly 300 metres before each junction, and central computers that fold those reports into the signal plan of the whole network. The counter-intuitive finding, documented in the standard study of the system: a dynamically-timed tram crossing costs cross traffic only about 5–8 seconds of green, and car throughput at the junctions stayed roughly constant. The tram gets the light exactly when it arrives; the time is handed back the moment it has passed. Fifty years on, Zurich's fleet is still both the sensor and the customer of its signal network, and the city's transit punctuality is the quiet advertisement.

The international evidence on transit signal priority backs Zurich up. Across US field deployments synthesised by the federal transit research programme, buses on priority corridors run 2–18 per cent faster. Los Angeles measured 7.5 per cent on its Metro Rapid corridors and puts the operating savings at about $3.3 million a year on two corridors alone: money recovered not from fares but from schedules that need fewer buses to deliver the same headway.

Buses and trams also work as roving probes even where they get no priority. London's entire bus fleet reports its position through the iBus system, feeding control rooms, passenger apps and signal priority from one data stream. Dublin wired its 1,000 buses to send GPS fixes every twenty seconds and let analysts hunt congestion causes in the pattern. New York's taxis, as we saw, became the floating probes that let outside researchers audit Midtown's signals. And Kassel's VERONIKA project connected buses, trams and fifteen intersections by direct radio so the signal could plan for the exact second a tram would arrive. Technically successful by every published account, though, in a pattern this atlas keeps encountering, the project closed in 2019 without publishing a single effect number. A pilot that ends unmeasured is a story, not evidence.

The newest twist outsources the fleet entirely. Google's Green Light analyses aggregated Maps driving data and sends signal re-timing suggestions to city engineers in twenty cities. Google reports up to 30 per cent fewer stops at tuned intersections. Hold that number with tongs: no peer-reviewed evaluation exists, and the closest thing to an outside check is a city commissioning its own. Boston had the analytics firm INRIX assess 114 re-timed intersections and reports an average 13.5 per cent cut in delay and 20 per cent fewer unnecessary stops: encouraging, city-measured, still not an audit. And the two most candid partner cities complicate the story. Seattle rolled a recommendation back after finding no net benefit, and Manchester's engineers ignored suggestions that conflicted with the thing their signals were deliberately doing, namely prioritising buses. Which is the deeper lesson: a signal network encodes a city's values. "Fewer car stops" is one value. It is not the only one, and in Zurich it loses.

The honest cost math

So does the million dollars beat the roundabout? On the field evidence: usually, with caveats. Pittsburgh's expansion prices adaptive control at very roughly $200,000 per intersection. A mid-sized city's worst twenty junctions come to a few million dollars, delivered in months without a single road closure, against road-building sums that start in the tens of millions and take a decade. Los Angeles' $3.3-million-a-year operating savings on two bus corridors is the kind of return no asphalt project posts. But the Manhattan bar chart above is the fine print: the software answer is cheap to buy and impossible to stop paying for. Signal timings drift, sensors fail, models go stale, the engineer who understood the system retires. Budget for the optimisation team, not just the installation. Or budget for the third bar.

The catch

Three failure modes, all documented in this atlas, keep the enthusiasm honest:

1. Installed is not operated. Dhaka automated its signals and the police kept directing traffic by hand; Algiers wired 22 of a planned 500 intersections and froze. The failure pattern is institutional, not technical — the same lesson as our survey of how smart cities die.
2. Gains fade without tenure. Manhattan's own data shows what happens when a system is launched and then merely owned rather than operated.
3. Adaptive signals cannot repeal geometry. They harvest the waste in bad timing, which is real but finite. A city that treats smarter signals as an alternative to managing car volumes, rather than a complement to transit priority as Zurich does, will spend its million dollars buying back one rush hour of induced demand.

Verdict

Software first, concrete second: the evidence supports the slogan, as long as the slogan comes with a maintenance contract. Adaptive signal control is mature, measured, and cheap by infrastructure standards. Its field results are single-digit to mid-teens improvements, sustained only where someone keeps tuning. And its best data source is increasingly the city's own fleet, which turns every bus and tram into both a sensor and a beneficiary.

Who operates it in year five — and who answers when it stops working?

— the one test this atlas applies to every entry, and the one adaptive signals must pass too

Zurich has had an answer since 1977. That, more than any algorithm, is why it works.

What cities should do

1. Budget the tuning, not just the installation. The field results hold only where someone keeps re-timing the network — Manhattan's own data shows what a system that is owned rather than operated settles back to. Fund the optimisation work for the life of the system, or expect the gains to fade.
2. Answer the year-five question before signing. Dhaka automated its signals and the police went on directing traffic by hand; Algiers wired 22 of a planned 500 junctions and stopped. Both failures were institutional, not technical. Name the team that operates the system in year five — and who answers when it stops working.
3. Use the fleet you already run. Buses and trams are both the best available sensor and the clearest beneficiary of priority at the junction, which is why Zurich's answer has held since 1977. Treat adaptive control as a complement to transit priority, not as a substitute for managing car volumes: signals harvest the waste in bad timing, and that waste is real but finite.

Frequently asked