Here Technologies, which supplies location and other transportation data, orchestrated a traffic forecasting contest that concluded this month with what company researchers said were surprisingly precise results.
While the foibles and arbitrary decisions of drivers could still easily jam up would-be traffic predictions, undercutting their usefulness as a commute or planning aid today,the results of Here’s contest show how complex computational setups known as neural networks can discern hard-to-find patterns from vast stores of data.
And that could have significance, for the ways cars get from one place to another, and the way researchers and planners analyze the interplay of transportation and environmental concerns, company officials said.
“It ultimately goes to, How do you help mitigate traffic?” said Jordan Stark, who studied urban planning and worked for former senator Chuck Hagel (R-Neb.) before heading global communications for Here.
A range of key policy questions startbecoming answerable with such technological advancements, he said. Take the high occupancy or toll (HOT) lanes on the Capital Beltway around Washington.
“If there are 10 percent more electric vehicles on the road, what is that impact, not only on C02 emissions, but also if you provide them a HOT lane? How does that impact the transportation system of Northern Virginia?” Stark asked. “You’re getting to that point” in which computers will be able to make predictions in such areas, he said.
Michael Kopp, head of research at Here and a founding co-director of the Austria-based Institute of Advanced Research in Artificial Intelligence, said Here gave researchers huge volumes of traffic data from the major German, Turkish and Russian cities. It amounted to months’ worth of color-coded traffic data, with speed in green, direction in blue, and volume in red.
The data, reported from vehicles, was precise, and the researchers set about building software tools to glean patterns from what amounted to torturously long and information-laden movies. They were judged based on their ability to predict the way traffic would look after the data they were provided ended, specifically five, 10 and 15 minutes later.
The top teams did so with incredible accuracy, organizers said. The error rate of the winner, Sungbin Choi, an independent researcher from Seoul, was “really impressively small. It’s less than 1 percent,” Kopp said.
“There are no theories how to build a successful neural network,” Kopp said, referring to a type of machine-learning that is meant to mimic the way researchers once thought brains functioned. Computers are fed vast amounts of data and “learn” to find certain patterns.
Another approach might be to try to “predict the traffic on let’s say a Monday morning at 10, maybe I look at other Monday mornings at 10. I average, and I submit that. People could have just done that. It’s an algorithm, if you like. Not a very powerful one,” Kopp said.
“It turns out that’s much worse than what people actually managed to achieve with neural networks,” Kopp added. “How do we know there are patterns? Well, these things are the ultimate pattern-finding tools. . . . We don’t know why it works. We know how it works.”
