Chi Jin
Princeton University, NJ·Computer Science·Early-career
How can a computer program learn to make good decisions one step after another, especially when the world is huge or only partly visible? This is the core question of reinforcement learning (RL), in which an agent learns by trial and error which actions lead to good outcomes.
- Fields
Computer Science
Engineering
Mathematics & Statistics
- Topics
reinforcement learning
machine learning theory
partial observability
- Grants
- 3 known grants, through Jul 2028
What they work on
- 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.
- Develop a general-purpose model-based reinforcement learning method using value-targeted system identification, along with stochastic-approximation versions for tractable computation and feature-learning methods for previously collected data.
- Determine which structural conditions make learning in Partially Observable Markov Decision Processes (POMDPs), models of decisions with hidden states, statistically or computationally efficient, addressing hidden-state inference and exploration.
- Develop efficient learning results for large partially observable problems using Predictive State Representations (PSRs), which model a system through predictions of future observations, combined with function approximation.
- Design decentralized algorithms that learn equilibria in Partially Observable Markov Games, where several agents each act with incomplete information.
- Build hpcGPT, a support assistant for supercomputer users, by combining large language model fine-tuning with Retrieval Augmented Generation over support history, documentation and real-time system status.
More about the research
RL drives robots, self-driving cars and game-playing systems. Yet the mathematical theory explaining when and why these methods work lags far behind practice. The lab's main projects aim to build that theory and to design algorithms with proven guarantees. A separate collaborative project applies large language models to help users of supercomputing centers.
Why it matters
Decision-making systems for robotics, autonomous driving, resource allocation and medical diagnostics often act with incomplete information. Provable guarantees could make such systems more reliable and data-efficient, because their behavior would rest on theory rather than trial alone. Separately, hpcGPT could shorten the wait for help among the tens of thousands of scientists and engineers who use academic supercomputing centers. That benefit would matter most for computing centers and research groups with limited support staff.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
CAREER: Foundations of Reinforcement Learning under Partial Observability
NSF · Aug 2023 – Jul 2028
Collaborative Research: Frameworks: hpcGPT: Enhancing Computing Center User Support with HPC-enriched Generative AI
NSF · Aug 2024 – Jul 2027
RI: Medium: Provable Reinforcement Learning with Function Approximation and Neural Networks
NSF · Oct 2021 – Sep 2026
Closest labs
Xia Wang
Ohio State University, OH·Computer Science
Build hpcGPT, a question-answering service for academic high-performance computing centers, by combining large language model fine-tuning with retrieval augmented generation so answers reflect center-specific documentation and past support experience.
Nan Jiang
University of Illinois Urbana-Champaign, IL·Computer Science·Early-career
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.
Miguel Lejeune
George Washington University, DC·Engineering
Advance real options analysis for infrastructure protection investment under multiple sources of uncertainty, building on optimal stopping theory and the Black-Scholes model from finance.
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.