- Jan 4, 2017
There are nearly 14,000 taxis in New York City, and sometimes that doesn’t seem like enough. But in a new study from MIT, researchers suggest that just 3,000 ride-sharing vehicles—be it a traditional taxi, an Uber/Lyft car, or a future autonomous robo-cab—could do the same job if each accepted up to four passengers. And if all passengers were willing to share their rides with nine other strangers in return for less traffic and lower cost, the city would need just 2,000 of such vehicles.
That would be welcome news for cities like NYC, where gridlock has gotten so bad the city has proposed to implement congestion fees. It could also be a boon for cities like Los Angeles or London that are trying to cut down on urban smog from vehicle emissions. (Of course, whether ride-hailing services and ride-sharing apps actually ease traffic or make it worse is still a matter of vigorous debate, hinging in large part on whether taking certain fleets off the road encourages more people to drive their own car.)
MIT’s “secret sauce,” as lead researcher Daniela Rus puts it in the Proceedings of the National Academy of Sciences, is an algorithm her team developed to find the most efficient routes for carpooling vehicles to pick up and ferry multiple passengers to their destinations in a single trip. It first lays out all the requests and available vehicles on a map in real time. Then it analyzes all possible trip combinations to find the best one before assigning passengers each vehicle. If a new request appears during the initial trip, the system calculates whether a cab should pick up that new party. It will also send idling cars to places with high demand based on historical information.
“One of the challenges is to create a solution that can deal with thousands of requests that appear at [once] in real time,” says Rus, director of the Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT. “We call our algorithm ‘anytime optimal,’ meaning our algorithm will be refined over time, ultimately converging to an optimal solution.”
To test the algorithm, the team ran simulations using public GPS data of 3 million taxi rides in Manhattan from one week in 2013. They found that if each cab carried up to four separate passengers, the wait time per passenger would average 2.7 minutes, and each rider’s trip would be delayed by roughly 2.3 minutes. With cabs or minibuses that can shuttle 10 passengers, that wait time would increase to approximately 2.8 minutes, while trip delays would average around 3.5 minutes. The algorithm would work with any ride-sharing service, and even with autonomous cars, according to the researchers.