Vijay Subramanian
University of Michigan, MI·Engineering
Reinforcement learning (RL) lets a computer learn how to make decisions by trial and error. It could run real-world systems such as traffic lights, but in the real world mistakes can be costly.
- Fields
Engineering
Computer Science
Mathematics & Statistics
- Topics
reinforcement learning
off-policy learning
partially observed systems
- Grants
- 1 known grant, through May 2027
What they work on
- 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.
- Design efficient, safe and constrained RL algorithms with provable guarantees, using model-free and function-approximated methods, including approaches for partially observed systems.
- Evaluate the robust and safe RL algorithms for traffic signal control through comprehensive simulations based on models of two instrumented real-world sites.
More about the research
A learned controller must stay safe, respect hard constraints and keep working when conditions differ from what it was trained on. This collaborative cyber-physical systems project, which links computing with physical infrastructure, develops RL methods with provable guarantees of robustness and safety. It uses smart traffic signal control as its motivating application and test bed.
Why it matters
Traffic signal systems in smart cities could run more efficiently with learning-based control, but only if that control is reliably safe. RL algorithms with proven safety and robustness guarantees could help transportation agencies and other operators of cyber-physical systems adopt data-driven control without risking unsafe behavior when real conditions differ from training data.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
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.
Mengdi Wang
Princeton University, NJ·Computer Science
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.
Ioannis Kantaros
Washington University in St Louis, MO·Engineering·Early-career
Train multi-agent reinforcement learning (MARL) control policies for robot teams and mathematically certify their robustness against adversarial, high-dimensional sensor inputs, testing them on aerial and ground robot testbeds and photorealistic simulators.
The school
University of Michigan: 124 labs on PI Pages, $2.1B a year on research, 963 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.