Traffic along some of Seattle’s stop-and-go streets is running a little smoother after Google tested out a new machine-learning system to optimize stoplight timing at five intersections. The company launched this test as part of its Green Light pilot program in 2023 in Seattle and a dozen other cities, including some notoriously congested places such as Rio de Janeiro, Brazil, and Kolkata, India. Across these test sites, local traffic engineers use Green Light’s suggestions—based on artificial intelligence and Google Maps data—to adjust stoplight timing. Google intends for these changes to curb waiting at lights while increasing vehicle flow across busy throughways and intersections—and, ultimately, to reduce greenhouse gases.
“We have seen positive results,” says Mariam Ali, a Seattle Department of Transportation spokesperson. Green Light has provided “specific, actionable recommendations,” she adds, and it has identified bottlenecks (and confirmed known ones) within the traffic system.
Managing the movement of vehicles through urban streets requires lots of time, money and consideration of factors such as pedestrian safety and truck routes. Google’s foray into the field is one of many ongoing attempts to modernize traffic engineering by incorporating GPS app data, connected cars and artificial intelligence.
On supporting science journalism
If you’re enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.
Preliminary data suggest the system could reduce stops by up to 30 percent and emissions at intersections by up to 10 percent as a result of reduced idling, according to Google’s 2024 Environmental Report. The company plans to expand to more cities soon.
The newfangled stoplight system doesn’t come close to replacing human decision-making in traffic engineering, however, and it may not be the sustainability solution Google claims it is.
All Systems Go
Electric stoplights are controlled in one of three main ways. The oldest use “fixed-time” lights that run on set schedules based on manually collected car counts. Newer lights may be “vehicle-actuated,” with detectors, usually installed under the road surface, that sense the presence or absence of cars and can adjust timing accordingly. Finally, adaptive or responsive traffic signals rely on algorithms in addition to sensors like cameras to keep tabs on vehicle flows across multiple intersections.
“Only around 4 to 5 percent of traffic signals in the United States are working in that adaptive mode,” says Aleksandar Stevanovic, a civil and environmental engineer at the University of Pittsburgh, who studies traffic control. Though effective, adaptive stoplights are costly to install and maintain. The most widely used adaptive control systems cost tens of thousands of dollars in initial investment per intersection, according to 2014 data from the U.S. Department of Transportation.
Google’s Project Green Light doesn’t require pricey fixed sensors, nor does it need on-the-ground observation. Instead it aggregates existing traffic data from Google Maps, gleaned from vehicles that essentially act as “mobile sensors,” says Henry Liu, a civil and environmental engineer at the University of Michigan, where he leads the Transportation Research Institute.
Google first builds a computer model of each intersection based on anonymized driving tracks from its Maps program, using machine learning to process the vast collection of information. Where cars repeatedly slow and stop, the model infers an intersection and calculates the exact timing of lights. The company also uses machine learning to identify potential adjustments. This could include shaving a few seconds off a red light and shifting that delay to another side of the intersection, says Matheus Vervloet, a Google Research project manager. Or it could include more quickly speeding through the entire light cycle. Tapping into the existing digital data on vehicle movements, Vervloet adds, could prove to be more affordable and efficient than building new sensor networks to monitor traffic.
Liu is working on a similar optimization system with digitally connected General Motors (GM) vehicles. The exact economics of his tool will depend on what GM charges for the data, but he expects that both his and Google’s systems would be “a fraction” of the cost of other options.
Google is currently offering its program to the participating cities at no charge. Vervloet declined to answer Scientific American’s questions about Google’s current financial investment in Green Light, the number of people working on the project or any potential plans to charge cities to participate in the future.
An early test of Liu’s tool in Birmingham, Mich., decreased the time spent and the number of stops at intersections by as much as 20 percent and 30 percent, respectively. Yet Liu doesn’t put much stock in any those numbers.
“It’s all dependent on the baseline you’re comparing to,” Liu says. In the case of Birmingham, the town has only fixed time lights based on car counts that haven’t been recently updated. “That’s why we see significant improvement,” Liu adds. Google’s reported results are based on assessment of 70 intersections, Vervloet says, most of which also aren’t currently running adaptive systems—so it’s difficult to compare Green Light’s outcomes with other, newer technologies, too.
Google’s approach is also much narrower than many modern traffic-control systems. So far Green Light only addresses optimization for one variable: fewer personal vehicle stops at lights. Vervloet says the test intersections were selected specifically to avoid complicating factors such as intersecting bus and bike lanes, trolley cars and high-use pedestrian crossings.
Even within those bounds, Green Light’s suggestions don’t always hit the mark. In one instance, Ali says, Seattle DOT reverted a Green Light–recommended shift in stop light timing because the adjustment “did not result in a net benefit.” And in Manchester, England—another of the Green Light pilot cities—traffic engineers often opted to ignore Google’s recommendations, Transport for Greater Manchester, a local government body, wrote in a statement e-mailed to Scientific American. In many instances, the city’s traffic engineers have intentionally set stoplights to prioritize bus routes or encourage commuters to steer clear of passing through residential areas, the statement noted. As a result, Google’s single-minded suggestions to minimize intersection stops were often moot. Decision-making by skilled humans remains key.
When orchestrating stoplight timing and coordination, traffic engineers must consider not just cars but also pedestrians, cyclists and public transit. And lights can help slow traffic down in school zones or discourage certain routes as throughways, Stevanovic says. Solving the problem of urban traffic “is not rocket science. It’s more difficult,” he says—half in jest but also seriously. “Traffic has so many uncertainties. In one hour, you can have five different goals that you want to achieve.”
Turning Green?
Google has figured out an easy-to-implement system for minimizing delays and frustration at some stoplights. But Green Light’s larger mission is to reduce traffic-associated carbon emissions and help cities meet sustainability goals, Vervloet says. Based on the results Google has made public, it’s unclear whether the project can achieve that aim.
On a company webpage describing Green Light, Google says that pollution can be 29 times higher at urban intersections than on open roads, according to a 2015 study published in Atmospheric Environment. Idling cars burn fuel to go nowhere, and reducing congestion can reduce local pollution. Yet it’s an open question whether cutting down stop-and-go traffic leads to lower overall greenhouse gas emissions in the long run.
Only about 2 percent of all U.S. transportation-related emissions are the result of congestion, according to a 2022 report from the Congressional Budget Office. Driving at higher speeds burns more fuel. And faster car travel may ultimately translate to people willing to take on longer commutes more regularly in a process known as induced demand. Stevanovic cautions that, the more car traffic is prioritized over improving infrastructure for public transit, bikes and pedestrians, the more people are likely to drive. The idea of minimizing stoplight delays to cut emissions is similar to the logic that leads legislators to recommend expanding highways to reduce pollution.
Still, traffic remains a legitimate quality of life issue for many, and the involvement of a company as influential as Google raises the profile of traffic engineering. Perhaps more interest and solutions will follow. “It’s great that Google is working on this problem,” says Stevanovic. “The contribution they can make is huge,” he says—as long as the company keeps its eyes on the road.