It’s mesmerizing video animation on YouTube showing simulated autonomous vehicle traffic streaming through a six-lane, four-way intersection. Dozens of cars flow through the streets, stopping, turning, slowing down and accelerating to avoid hitting their neighbors. And not a single car stopped. What if at least one of these vehicles is not autonomous? What if there was only one?
In the coming decades, autonomous vehicles will play an increasingly essential role in society by keeping drivers secure, making deliveries, or increasing accessibility and mobility for elderly or disabled passengers.
However, MIT assistant professor Cathy Wu argues that autonomous vehicles are just one part of a complicated transportation system that may include individual autonomous cars, delivery fleets, human drivers, and an array of last-mile solutions designed to get passengers home, not to mention road infrastructure , such as motorways, roundabouts and, yes, intersections.
Transportation currently accounts for about one-third of U.S. energy consumption. The decisions we make today about autonomous vehicles could have a substantial impact on this number, from a 40 percent reduction in energy consumption to a doubling of energy consumption.
So how can we better understand the problem of integrating autonomous vehicles with the transport system? Just as importantly, how can we exploit this understanding to guide us towards better functioning systems?
Wu, who joined the Laboratory for Information and Decision Systems (LIDS) and MIT in 2019, is the Gilbert W. Winslow Postdoctoral Fellow in Civil and Environmental Engineering and a core faculty member of MIT’s Institute for Data, Systems and Society. Growing up in a family of electrical engineers in the Philadelphia area, Wu sought a field that would allow her to apply her engineering skills to solve social problems.
While an undergraduate at MIT, she contacted Professor Seth Teller of the Computer Science and Artificial Intelligence Laboratory to discuss her interest in autonomous cars.
Teller, who died in 2014, answered her questions with heartfelt advice, Wu says. “He told me, ‘If you have any idea what your passion in life is, you have to pursue it as tough as you can. Only then can you hope to find your true passion.
“Anyone can tell you to follow your dreams, but his insight was that dreams and ambitions are not always clear from the beginning. Finding and pursuing your passion takes hard work.”
Following this passion, Wu continued to work at Teller, as well as in Professor Daniela Rus’s Distributed Robotics Laboratory, and finally as a graduate student at the University of California, Berkeley, where she won the 2019 IEEE Knowledgeable Transportation Systems Society’s Outstanding Ph.D.
While in college, Wu had an epiphany: she realized that for autonomous vehicles to deliver on their promise of fewer accidents, time savings, lower emissions, and greater socioeconomic and physical accessibility, these goals had to be explicitly designed—either as physical infrastructure or algorithms used by vehicles and sensors or deliberate policy decisions.
At LIDS, Wu uses a type of machine learning called reinforcement learning to study how traffic systems behave and how autonomous vehicles in those systems should behave to achieve the best possible results.
Reinforcement learning, which was most eminent in AlphaGo, DeepMind’s Go-beating program, is a powerful class of methods that captures the idea of trial and error – given a goal, a learning agent repeatedly tries to achieve it, but fails. and learning from mistakes made in the process.
In a traffic system, the goal may be to maximize the overall average speed of vehicles, minimize travel time, minimize energy consumption, and so on.
By examining common elements of traffic networks, such as mesh roads, bottlenecks, and on- and off-ramps, Wu and her colleagues found that reinforcement learning can match and, in some cases, exceed the effectiveness of current traffic control strategies. More importantly, reinforcement learning can shed fresh featherlight on understanding complicated network systems that have long eluded classical control techniques. For example, if just 5-10 percent of vehicles on the road were autonomous and used reinforcement learning, it could eliminate congestion and enhance vehicle speeds by 30-140 percent. And the lessons learned from one scenario often translate well to others. These insights could soon inform public policy and business decisions.
During this research, Wu and her colleagues helped improve a class of reinforcement learning methods called policy gradient methods. Their advances proved to be an overall improvement on most existing deep reinforcement learning methods.
However, reinforcement learning techniques will need to be continually refined to keep up with scale and changes in infrastructure and changing behavior patterns. City planners, car manufacturers and other organizations will need to translate research findings into action.
Wu currently works with public agencies in Taiwan and Indonesia to exploit lessons from his work to drive better dialogues and make better decisions. By changing traffic lights or using incentives to change driver behavior, are there other ways to achieve lower emissions or smoother traffic?
“I’m surprised by this job every day,” Wu says. “We set out to answer the question about autonomous cars, and it turns out that lessons can be learned and applied in a different way, and that leads to new and exciting questions to answer.”
Wu is content to have found her intellectual home at LIDS. According to her, it is “a very deep, intellectual, friendly and hospitable place.” His research inspirations also include the MIT 6.003 (Signals and Systems) course – which he encourages everyone to take – conducted in the tradition of professors Alan Oppenheim (Research Laboratory of Electronics) and Alan Willsky (LIDS). “The course taught me that so much in this world can be fruitfully explored through the prism of signals and systems, whether electronics, institutions, or society,” he says. “As I say this, I just realize that LIDS thinking has been empowering me all this time!”
Researching and teaching during the pandemic hasn’t been basic, but Wu is coping best with a arduous first year as a faculty member. (“I work from home in Cambridge – my short walking commute doesn’t matter at the moment,” she says wryly.) To relax, she likes running, listening to podcasts on topics ranging from science to history, and reverse engineering her favorites Trader Joe’s frozen foods.
She also worked on two Covid-related projects that originated at MIT: One examines how environmental data, such as data collected by IoT-connected thermometers, can assist identify emerging epidemics in communities. Another project asked whether it was possible to determine how contagious the virus was on public transport and how various factors could reduce the risk of transmission.
Wu says both are in the early stages. “We hope to contribute a little to increasing the pool of knowledge that can help policymakers somewhere. It has been very informative and rewarding to do this and see all the other efforts going on at MIT.”