Crash course
by Tai Viinikka
Secretly, most of us think we’re great drivers. Whether we are aggressive
or cautious, we assume the roads would flow smoothly, and traffic jams would
be a rare marvel, if only everyone drove like, well, like me. Like many secret
vanities, this one is just not true, and traffic researchers have built computer
models that prove it.
A new approach to modeling, in which individual simulated drivers make individual
simulated decisions, began in the 1990s. These driver-behavior models are reaching
maturity now, and their proponents say that even the most vexing and perplexing
aspects of traffic can be predicted and even understood.
Models that follow individual vehicles are already beginning to make useful
predictions. Moshe Ben-Akiva, the director of MIT’s Intelligent Transportation
Systems program has helped design more efficient toll plazas and flag unneeded
on-ramps before they ever leave the planning stage. But “our most important
emphasis is on predicting traffic and giving good alternatives,” says
Ben-Akiva, a professor of civil and environmental engineering. He predicts that
the MIT traffic simulation software, DynaMIT, will be used to advise drivers,
with radio traffic updates, phone numbers that give traffic conditions, and
computer-controlled highway signs. These predictions are based on making sure
each simulated driver in the model has his own artificial personality - without
that individuality, the forecasts can be surprisingly wrong.
Traffic in Boston has gotten worse recently, even relative to other American
cities. Bostonian's travel time index (a measure of congestion that takes into
account typical commute distances, not just drive time) was 11th in the nation
in 1991. But by 2002, the city ranked 5th, behind Los Angeles, San Francisco,
Chicago, and Washington DC, according to the Texas Transportation Institute,
which publishes an Urban Mobility Report each year. The Institute says the average
American city dweller wastes 58 hours per year stuck in traffic. Governments
attack this problem by building roads, funding public transit, raising or lowering
speed limits, closing on-ramps during certain hours, or metering on-ramps to
only allow a certain amount of traffic onto the freeway. But without some kind
of traffic modeling, all these efforts are essentially blind fiddling.
For longer than there have been digital computers, physicists and engineers
have labored over traffic equations. By analyzing traffic and putting its variables
into equations and graphs, scientists hoped to quantify traffic flow to predict
the point when traffic turns viscous and then to stop-and-go. The first traffic
mathematics were published in the late 1930s, about 30 years after the first
Ford Model T was produced. The topic became urgent with the traffic jams of
the late 1950s and only became more so as traffic became an increasingly annoying
part of everyday life.
Most of these models -- and the ones that followed in the digital era -- saw
traffic as a flowing liquid. A traffic jam occurs when more traffic can get
into an area than can leave it, like a hose filling a bucket with only a small
drain-hole in the bottom. Yet those models can’t predict the waves of
stop-and-go motion that ripple through a real highway. So while they may get
the average speed of the cars correct, they’ll be wrong about factors
like driving behavior and acceleration. A step in the right direction came in
the 1950’s, when experts at General Motors did race track testing to quantify
braking and acceleration behaviors. But they still didn’t count for individual
decisions.
Ben-Akiva’s view was that the models needed more room for individuality.
So he and his colleagues created computer programs to imitate individual drivers.
Every one of the thousands of drivers on the virtual roads in the DynaMIT model
has a home, a destination , and a route in mind. They also have pre-determined
“driving behaviors,” such as reaction time, a tendency to jump out
into a faster lane; to speed or dawdle; to ignore traffic signs; or to bypass
a waiting line of cars and try to muscle back in. “Drivers are strategic,”
Ben-Akiva says, “They anticipate turns, they see stoppages and react.
But they don’t all anticipate [equally], so you need to model human differences.”
To model lane changes, for example, the engineers digitized video footage showing
vehicles merging. They tagged each car that traveled through the frame, then
had a computer vision program track the cars through later frames. Creating
tracks through space and time for each vehicle meant they could observe the
details of how drivers change lanes. As the program puts itself into each driver’s
place, one after another. By making such simple decisions based on different
driver profiles for thousands of drivers each second, the model predicts how
a whole highway will behave.
It turns out that if everyone were the same, traffic would be more chaotic —
all drivers would see that their lane is closing and try to merge out of it
at the same time, for instance. So naive models get more congested quicker because
drivers all act alike, and because they all act dull.
There are other realities beyond driver personality that a good model can grapple
with. University of Leed’s DRACULA traffic model features some of the
same individual driving differences as DynaMIT, and also lets drivers plan their
own route to the destination. The simulated drivers learn, remembering results
from one trip to the next, and experiment to find better routes. H. Michael
Zhang at University of California Davis has a model in which every driver behaves
the same, but the identical automatons will drive differently depending on whether
the surrounding traffic is accelerating, coasting, or braking. Basically, Zhang's
simulated drivers can get nervous, or impatient. So the importance of human
frailty is a key for his model, too.
Such models can do more than just satisfy scientific curiosity about why traffic
behaves as it does – they can help planners design roads, and maybe help
workers plan their commutes. In the summer of 2004, a team from University of
Duisburg-Essen put their traffic model to work, creating a predictive map of
metropolitan Cologne. The model highlights congested areas up to an hour in
advance. It's on the web, and more than a hundred thousand people look at it
each day to find out what their drive to or from work will be like. This may
be what the future looks like, in more ways than one.
So will Boston ever have traffic forecasts, instead of these traffic “nowcasts”
that only tell us what has already gone wrong? That’s up to the city and
state traffic authorities who run the traffic monitoring and operations center
in South Boston. Testing of DynaMIT in Irvine and Los Angeles is under way,
and the results may give Boston’s traffic managers enough confidence to
start forecasting your morning drive.
For now, DynaMIT, with its “behavioral realism,” shows that we should
be grateful that we’re all different. Although the rudenesses or pokey
behaviors will continue to annoy us, we can take solace in that fact that if
everyone drove exactly like me, it would be much worse.