Transportation is a public system that almost everyone depends on, and almost everyone can see failing in some way. For Cathy Wu ’12, MNG ’13, that problem is both personal and technical: traffic shaped her family life growing up, and computer games helped point her toward systems thinking.
Now an associate professor in the MIT Department of Civil and Environmental Engineering (CEE) and the Institute for Data, Systems, and Society (IDSS), Wu studies how machine learning and reinforcement learning can support better decisions in complex systems. Transportation is a central focus, but the larger question is broader: how can computational tools help society reason through choices that are too complicated for today’s methods to handle easily?
Why transportation is such a hard systems problem
Wu’s interest in transportation is rooted in the way it touches everyday life. Her father had a long commute, and her family lived on a street too busy for outdoor play. Those early experiences, along with games like “SimCity,” helped shape her interest in designing safer and more efficient transportation systems.
The difficulty is that transportation design is rarely about one simple choice. A researcher or planner may need to model and compare dozens, hundreds or thousands of variants. That makes a fully evidence-driven approach hard to reach with conventional tools.
Wu sees reinforcement learning, or RL, as a way to expand what researchers can test. In her view, the goal is not just automation for its own sake. It is to give transportation researchers and their practitioner partners a better way to explore the systems they want to build.
“Designing transportation systems consists of modeling and analyzing dozens, if not hundreds or thousands, of variants, which means that an evidence-driven approach to designing those systems is simply not within reach of today’s tools,” Wu says.
From autonomous vehicles to hard optimization problems
Wu’s path into artificial intelligence and transportation began while she was an undergraduate at MIT. A lecture on autonomous vehicles by the late professor Seth Teller, delivered during an Independent Activities Period robotics competition that Wu won, sharpened her approach to the field.
She worked with Teller, then moved to Professor Daniela Rus after Teller encouraged her to do so. Rus had done research on robotaxis, and that connection helped Wu continue exploring the link between AI and transportation. Wu also credits “my friends at Dropbox,” who invited her to do a second internship focused on transportation issues.
After earning her master’s degree at MIT, Wu completed her PhD at the University of California at Berkeley. There, she saw transportation researchers spending years developing optimization methods to model and analyze a single new variant of a system. Her work aimed to change that pace by developing reinforcement learning and optimization methods that could make the process far more efficient.
In 2018, during the last year of her PhD at UC Berkeley, Wu applied RL to a traffic problem: automatically analyzing the possible traffic flow impact of autonomous vehicles across different traffic networks. The research went viral, but it did not turn RL into an easy answer.
The setback that reshaped the research
After a postdoc at Microsoft focused on RL theory, Wu returned to MIT as faculty. She was drawn by the sustainability focus of CEE and by IDSS’s emphasis on bringing data science into other disciplines.
Then came a difficult period. Over the next two years, Wu’s attempts to apply RL to traffic problems failed. The earlier success had shown that RL could be used in transportation systems, but the method proved highly sensitive.
In 2022, Wu and her students identified a core issue: an RL algorithm that works on one problem may fail on another, even when the two problems are closely related. That finding helped explain why success in one traffic setting did not automatically carry over to another.
The next step came in 2023. Wu and her team found a way to work around RL’s sensitivity. The key was to recognize that, within a group of related problems, RL may not train well on 90 percent of them, but may train quite well on 10 percent.
By training on the problems that solve and generalize well, the resulting models could perform well across related problems, including problems that would not have been solved through direct training. The researchers also created an algorithm to identify which problems should be used for RL training.
That algorithm improved training efficiency by up to 30 times. In practical terms, a task that would normally require 100 training models may only require three models.
What RL could mean for transportation policy
Wu’s more recent work applies reinforcement learning to a transportation optimization problem with direct policy implications: eco-driving. In this setting, vehicle speeds are intelligently controlled to reduce excessive stopping and starting.
The research shows that eco-driving measures could reduce vehicle emissions by between 11 and 22 percent. It also provides evidence that policies using those measures could significantly improve system efficiency.
For Wu, that result matters because it connects computational research to evidence-based policy. Transportation debates can be difficult because the systems are complex and the tradeoffs are not always obvious. Data-driven tools can help identify which questions can be analyzed systematically and where objective answers may be possible.
Her work also extends beyond transportation. Wu says much of her group’s recent research has produced algorithms that streamline the development of solvers for hard optimization problems, including logistics, supply chains, manufacturing, and resource allocation.
A use-inspired approach to complex systems
Wu describes her preferred style of work as “use-inspired basic research.” That means beginning with a practical problem while developing fundamental knowledge that may transfer to other important problems.
In her group, students start with consequential issues such as safety, congestion, and accessibility. They examine where current methods break down, then let those shortcomings shape the research direction.
That approach keeps the work grounded. Reinforcement learning is not treated as a universal solution. It is tested against hard transportation problems, refined when it fails, and judged by whether it can help researchers and policymakers understand complex systems more clearly.
Wu’s teaching reflects the same personal motivation. She says she loves working with students in both the classroom and research mentoring, especially when a concept clicks. The broader goal remains consistent: use computation to improve systems that affect people’s lives.