Updated

Updated · MIT News · Oct 2

MIT's Cathy Wu Lifts Reinforcement Learning Efficiency 30-Fold by Training on the Best 10%

Updated

Updated · MIT News · Oct 2

MIT's Cathy Wu Lifts Reinforcement Learning Efficiency 30-Fold by Training on the Best 10%

1 articles · Updated · MIT News · Oct 2

Wu and her team built an algorithm that identifies the subset of related problems on which reinforcement learning trains reliably, cutting training needs by up to 30 times.
The method addresses a core weakness in RL sensitivity: models may fail on 90% of similar tasks but generalize well when trained on the 10% that solve cleanly.
Wu said the advance followed years of setbacks after an earlier 2018 traffic-study success, with her group only pinpointing the sensitivity problem in 2022 and devising the workaround in 2023.
Recent transportation tests suggest the approach can inform policy as well as theory, with RL-based eco-driving controls reducing vehicle emissions by 11% to 22%.
Wu's broader aim is to turn use-inspired AI research into practical tools for transportation, logistics, supply chains, manufacturing and other complex systems.

Sources

How did an MIT professor turn years of AI failures into a breakthrough that cuts traffic emissions by 22 percent?

Could combining classical algorithms with reinforcement learning finally solve the unpredictable chaos of modern global supply chains?

Will neural simulators running 10,000 times faster than government standards ultimately hand control of our urban traffic grids to AI?

1 total