Cathy Wu
@cathywu.bsky.social
AI & Transportation | MIT Associate Professor
Interests: AI for good, sociotechnical systems, machine learning, optimization, reinforcement learning, public policy, gov tech, open science.
Science is messy and beautiful.
http://www.wucathy.com
Interests: AI for good, sociotechnical systems, machine learning, optimization, reinforcement learning, public policy, gov tech, open science.
Science is messy and beautiful.
http://www.wucathy.com
This project was 4 years in the making and it's finally out!
We found that controlling vehicle speeds to mitigate traffic across a city can cut carbon emissions between 11 and 22 percent. To do this, we used deep reinforcement learning to optimize one million eco-driving scenarios. 🚗🤖🧠
We found that controlling vehicle speeds to mitigate traffic across a city can cut carbon emissions between 11 and 22 percent. To do this, we used deep reinforcement learning to optimize one million eco-driving scenarios. 🚗🤖🧠
August 11, 2025 at 6:47 PM
This project was 4 years in the making and it's finally out!
We found that controlling vehicle speeds to mitigate traffic across a city can cut carbon emissions between 11 and 22 percent. To do this, we used deep reinforcement learning to optimize one million eco-driving scenarios. 🚗🤖🧠
We found that controlling vehicle speeds to mitigate traffic across a city can cut carbon emissions between 11 and 22 percent. To do this, we used deep reinforcement learning to optimize one million eco-driving scenarios. 🚗🤖🧠
9/n We like multi-agent traffic tasks, so we tried those out too. Up to 25x sample efficiency improvements! 🚦🚗 Btw, 25x means training 4 tasks to match the performance of training 100 tasks. I cannot express how excited we are to try this out on more applications.
December 8, 2024 at 6:13 PM
9/n We like multi-agent traffic tasks, so we tried those out too. Up to 25x sample efficiency improvements! 🚦🚗 Btw, 25x means training 4 tasks to match the performance of training 100 tasks. I cannot express how excited we are to try this out on more applications.
8/n It works surprisingly well! Up to 50x sample efficiency improvements over typical training (brown vs orange, green)! Sequential oracle (pink) chooses the single best training task at each step, given full knowledge. And some heuristic baselines (random, greedy, equidistant).
December 8, 2024 at 6:13 PM
8/n It works surprisingly well! Up to 50x sample efficiency improvements over typical training (brown vs orange, green)! Sequential oracle (pink) chooses the single best training task at each step, given full knowledge. And some heuristic baselines (random, greedy, equidistant).
7/n Using this model, we choose training tasks––one at a time––that best improve overall generalization performance across contexts. We leverage ideas of optimism in the face of uncertainty (UCB) and Bayesian optimization to design a suitable acquisition function. Then iterate!
December 8, 2024 at 6:13 PM
7/n Using this model, we choose training tasks––one at a time––that best improve overall generalization performance across contexts. We leverage ideas of optimism in the face of uncertainty (UCB) and Bayesian optimization to design a suitable acquisition function. Then iterate!
👼 As an applied RL researcher, this is the most optimistic I have been about RL in years. It feels like seeing the light at the end of the tunnel when RL training starts working reliably. Without a ton of compute or tuning. Very excited for what is to come. Here is what we did👇
December 8, 2024 at 6:13 PM
👼 As an applied RL researcher, this is the most optimistic I have been about RL in years. It feels like seeing the light at the end of the tunnel when RL training starts working reliably. Without a ton of compute or tuning. Very excited for what is to come. Here is what we did👇