Nan Jiang
University of Illinois Urbana-Champaign, IL·Computer Science·Early-career
Reinforcement learning (RL) is a branch of artificial intelligence that learns to make good decisions by trial and error. It has worked well in simulations, where mistakes cost nothing, but in the real world trial and error is often expensive, dangerous or unethical.
- Fields
Computer Science
Mathematics & Statistics
- Topics
reinforcement learning
model selection
hyperparameter tuning
- Grants
- 1 known grant, through Apr 2027
What they work on
- Develop model selection methods for offline reinforcement learning that choose among candidate policies using a holdout dataset, building on theoretical work on value-function selection.
- Make model selection methods for offline RL work in practice by handling poorly fitted candidate value functions and historical data with insufficient coverage.
- Establish the theoretical conditions under which offline RL training, learning only from historical data, is guaranteed to succeed, and derive new algorithms from these insights.
More about the research
In personalized medicine, for example, an RL system cannot try a possibly harmful treatment just to learn from it. Offline RL avoids this by learning only from historical data, but that creates hard technical problems.
Why it matters
Many real decisions, such as adjusting a patient's treatment, controlling a robot or recommending content, cannot safely be learned by trial and error. Reliable offline reinforcement learning, backed by theory, could let AI systems learn these decisions from existing records instead. Better model selection methods could make it practical to tune and trust such systems before they are used.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Chi Jin
Princeton University, NJ·Computer Science·Early-career
Identify the general structures of function classes, such as those represented by neural networks, under which model-free reinforcement learning is tractable, and design provably efficient algorithms that also handle model misspecification.
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.
The school
University of Illinois Urbana-Champaign: 41 labs on PI Pages, $907M a year on research, 870 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.