MIT Professor Speeds Up AI Training by 30 Times

Growing up, Cathy Wu watched her dad disappear into long commutes while her family stayed home on a street too dangerous for play. Today, the MIT professor is using artificial intelligence to redesign the transportation systems that shaped her childhood.

Wu specializes in reinforcement learning, a type of AI that learns through trial and error. For years, she dreamed of using it to help transportation researchers test hundreds of design ideas without building physical models or running expensive simulations.

In 2018, she had her first breakthrough. Wu successfully trained AI to analyze how self-driving cars might affect traffic flow across different road networks. The research went viral and seemed to validate her entire approach.

Then everything stopped working. For two years, Wu and her students tried to apply the same methods to other traffic problems. Every attempt failed.

"That was stressful," Wu admits. She wasn't sure if the problem was her teaching, her students' work, the complexity of traffic itself, or whether reinforcement learning just couldn't handle these challenges.

But Wu's team kept digging. In 2022, they discovered something crucial. The AI algorithms were extremely sensitive. A method that worked perfectly on one problem would completely fail on a nearly identical one.

That discovery led to the real breakthrough in 2023. Wu's team found that while AI might fail to train on 90 percent of related problems, it could train beautifully on the remaining 10 percent. Better yet, models trained on those easier problems could then solve the harder ones they'd originally failed on.

They designed a new algorithm that identifies which problems to use for training. The result? A 30 times improvement in efficiency. Tasks that once required training 100 separate AI models now need only three.

The Ripple Effect

Wu's work reaches far beyond fixing traffic jams. Her methods apply to any complex system where researchers need to test countless variations, from energy grids to public health networks. Transportation researchers who once spent years developing a single model can now analyze hundreds of scenarios in the same timeframe.

The breakthrough also restored something equally important: confidence that AI can tackle society's toughest optimization problems. After years of failure, Wu proved that the right approach can overcome even the most stubborn technical barriers.

Wu now leads a team at MIT focused on "contextual reinforcement learning," continuing to refine how AI solves interconnected problems. She tells students that societal impact takes time. "It will take years to really understand what's going on and where the real problems are," she says. "In the meantime, try to be helpful."

For Wu, that means staying curious and asking many questions, just like she did playing SimCity as a kid dreaming of better streets.