Mengdi Wang
Princeton University, NJ·Computer Science
Reinforcement learning (RL) lets a computer learn by trial and error which actions earn the most reward. It is attractive for controlling robots, self-driving cars and power grids.
- Fields
Computer Science
Engineering
Mathematics & Statistics
- Topics
reinforcement learning
distributionally robust optimization
risk-averse learning
- Grants
- 1 known grant, through Aug 2027
What they work on
- Develop distributionally robust training methods for model-based reinforcement learning so that policies avoid unsafe states even when the real environment differs from the training data.
- Bridge Bayesian and safe reinforcement learning with Bayesian risk-averse learning that keeps improving models and policies after deployment while holding risk below a set level.
- Design near-optimal policy learning algorithms that use online change point detection to adapt when the environment changes in sudden, piecewise steps.
- Build a rigorous simulation approach using space-filling experiment design to evaluate a policy across many environment settings and find unexpected unsafe behavior before deployment.
More about the research
In these settings, however, a policy, meaning the learned rule for choosing actions, must avoid dangerous states, including failures no one anticipated. Policies can break when the real world differs from training data or changes over time. This collaborative project develops model-based RL methods, which learn a model of how the environment behaves, with safety built in at every stage: training, deployment, improvement and evaluation.
Why it matters
Learning-based control is increasingly considered for robots, autonomous vehicles and power systems, where unsafe actions can cause real harm. Methods that hold up when conditions shift, that keep improving without taking excess risk, and that expose hidden failures in simulation could make reinforcement learning safer to deploy in these high-stakes domains.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Pan Xu
Duke University, NC·Computer Science·Early-career
Develop distributionally robust reinforcement learning methods that use minimax optimization over a set of possible environment dynamics, so that policies trained in a simulator perform reliably in an unknown target domain, with finite-sample guarantees.
Vijay Subramanian
University of Michigan, MI·Engineering
Develop robust reinforcement learning algorithms that use model-class knowledge, future predictions and off-policy methods to cope with mismatches and limited data from simulators or offline records.
Navid Azizan Ruhi
Massachusetts Institute of Technology, MA·Engineering·Early-career
Develop Hard-Constrained Neural Networks (HardNet), which use a differentiable projection layer to guarantee input-dependent constraints, and embed them in reinforcement learning to produce certifiably safe policies.
The school
Princeton University: 26 labs on PI Pages, $559M a year on research, 403 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.