Pan Xu
Duke University, NC·Computer Science·Early-career
Many important decisions are made step by step, while learning from limited and imperfect information. Examples include choosing medical treatments, recommending content and controlling robots.
- Fields
Computer Science
Mathematics & Statistics
- Topics
reinforcement learning
distributionally robust optimization
transfer learning
- Grants
- 2 known grants, through Sep 2031
What they work on
- 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.
- Design off-dynamics reinforcement learning algorithms that use partial target-domain data through data augmentation and cross-domain learning, and test them on benchmarks and real-world healthcare datasets.
- Build learning frameworks for settings with limited interaction and high costs for switching policies, focusing on stable and efficient real-world deployment.
- Develop fast approximate sampling-based exploration strategies for contextual bandits, decision problems that use side information, with provable performance guarantees.
- Extend sampling-based exploration to deep neural network applications and to full reinforcement learning problems, producing provably efficient algorithms released as open-source software.
More about the research
Reinforcement learning, a branch of machine learning in which an agent learns by trying actions and receiving rewards, is a natural tool for these problems. However, it often needs a great deal of trial and error, which can be unsafe or costly in the real world. The two projects address this problem from two directions. One asks how an agent can explore efficiently. The other asks how an agent trained in one setting can be trusted in a different one. Both combine mathematical guarantees with practical algorithms released as open-source software.
Why it matters
Decision-making systems that can learn safely from simulators and limited real data could support settings where trial and error is costly or unethical, such as choosing medical treatments, managing public health responses and controlling robots. Faster, provably reliable exploration methods could improve recommendation systems, robotics and personalized healthcare. The open-source tools are intended for broad use.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
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.
Haipeng Chen
William & Mary, VA·Computer Science
Develop physics-inspired, task-aware graph neural networks that build problem structure into reinforcement learning, so it can handle badly shaped or implicit objectives in combinatorial optimization.
Yuxin Chen
University of Pennsylvania, PA·Mathematics & Statistics
Characterize the fundamental limits on sample and computational complexity of reinforcement learning, meaning how much data and computing an agent needs to learn good decisions in an unknown environment.
The school
Duke University: 61 labs on PI Pages, $1.6B a year on research, 417 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.