School of Civil and Environmental Engineering
Present
Building behavioral models at scale with reinforcement learning and imitation learning
Speaker:
Dr. Eugene Vinitsky
Assistant Professor, Civil and Urban Engineering, NYU
Microsimulation is a key tool used to study transportation impacts. These microsimulators rely on low-dimensional behavioral models that do not always correctly reflect the potential impact of a new intelligent traffic control strategy. This can show up as a large simulation-to-reality gap that slows down the deployment of new intelligent congestion control strategies. We propose to improve on existing behavioral models by using abundant behavioral data to learn models of human behavior. Unfortunately these imitation-based models, while reproducing qualitative behaviors, tend to have high crash rates that prevent their easy use in micro-simulators. We investigate whether the combination of imitation and reinforcement learning at scale can give us the best of both worlds: human-like behavior and human-level crash rates. We design a new simulator, GPUDrive, that allows us to scale RL training to billions of steps and show promising evidence that this approach can scale.
Monday, September 23, 2024
12:20 – 1:10 PM
MCHU 202
Bio: Eugene Vinitsky is an assistant professor of Transportation Engineering at NYU and member of the C2SMARTER consortium on congestion reduction. He received his PhD in controls engineering from UC Berkeley with Alexandre Bayen where he developed RL algorithms used to control over a hundred traffic-smoothing cruise controllers. During his PhD he spent time at Tesla, Deepmind, Facebook AI Research, and was a research scientist in the Apple Special Projects Group.
For more information, contact: Monika Filipovska at monika.filipovska@uconn.edu