MIT Associate Professor Cathy Wu is developing reinforcement learning methods to solve complex transportation problems that conventional optimization tools struggle to address. Her research aims to help planners analyze many possible system configurations efficiently and use computational evidence to guide practical decisions, tells MIT News.
Wu’s interest in transportation began with personal experience. Her father faced a long daily commute, while computer games such as SimCity introduced her to systems thinking. At MIT, exposure to autonomous vehicle research led her toward artificial intelligence and transportation.
During her doctoral work at the University of California at Berkeley, Wu saw researchers spend years developing optimization techniques for individual transportation scenarios. In 2018, she demonstrated reinforcement learning’s potential by using it to analyze the effects autonomous vehicles could have on traffic flow across different networks.
Progress was not straightforward. Wu later found that reinforcement learning algorithms could be extremely sensitive, succeeding on one problem but failing on a closely related one. Her team eventually developed a method that trains models on a carefully selected subset of problems that generalize effectively to others. The approach improved training efficiency by as much as 30 times.
This work now supports contextual reinforcement learning, which seeks solutions across families of related problems rather than optimizing one case at a time.
Wu’s team has also explored eco-driving, in which vehicle speeds are intelligently controlled to reduce unnecessary stopping and acceleration. The research indicates that such strategies could cut vehicle emissions by 11–22%, providing quantitative evidence that could inform transportation policy.
The methods extend beyond mobility. Wu’s group is developing algorithms for difficult optimization problems in logistics, supply chains, manufacturing, and resource allocation.
Her broader goal is to combine fundamental research with real-world needs. Rather than starting with algorithms and searching for applications, Wu encourages researchers to begin with consequential problems and develop computational tools capable of addressing them.
