Offline Reinforcement Learning 

Offline Reinforcement Learning is a paradigm that learns policies exclusively from static datasets of previously collected interactions, making it particularly appealing for real-world applications, such as healthcare, finance, and robotics, etc. The two biggest challenges are the distributional shift and the restrictive conditions on the attributes of function approximation. I'm interested in tackling these challenges, developing and analyzing sample-efficient (also known as PAC-learnable) algorithms using statistical learning theory. Most recently, I have been considering the problems under the learning paradigm using pair-wise preference-based human feedback (like in ChatGPT), rather than explicit reward signals.

AI for Science

I am increasingly interested in developing artificial intelligence tools for scientific discovery and understanding. In particular, my research in this direction mainly focuses on three aspects: 

(i) AI-guided generation of scientific hypotheses

(ii) integration of AI with scientific experiments and simulation

(iii) learning meaningful representation of scientific data

Currently, I closely collaborate with researchers in neurobiology and behavior on large-scale neurophysiologic studies. We have made significant progress in understanding the neuroscience underlying neurophysiologic choices and investigating the type of goal-directed characteristics of planning. I have been involved and actively seeking opportunities in other scientific domains.

Precision Medicine

The demand to address patients' heterogeneous responses has led to the development of tailored treatment rules to enhance individual results. Methodological advancements in this field have transitioned from 

(i) binary treatment (like treatment vs. control) to multicategory and continuous treatment (like dose-finding)

(ii) single treatment stage to multiple stages as seen in dynamic treatment regimes, even more recently, an infinite number of stages in mobile health studies

I am broadly engaged in developing statistical methods for personalized medicine in various settings.   

Graph (Deep) Representation Learning 

Graph neural networks (GNNs) have led to new developments in various domains, including the analysis of information processing in the brain, chemical synthesis, recommender systems, and modeling social networks. Among the methods for deep graph representation, the family of GNNs has achieved remarkable success in real-world graph/node-focused tasks. Despite their success, existing neural architectures suffer from a lack of interpretability in results due to their black-box nature, and an inability to learn geometric representations of varying orders. I am particularly focusing on developing novel neural architecture to solve the limitations with theoretical foundations.

Accountability 

Accountability is a crucial and desirable characteristic in many statistical problems. Many real-world scenarios it is desirable to avoid making overconfident predictions/decisions that could result in costly and/or irreversible consequences. That is, rather than solely relying on a point estimate or greedy-optimal solution, having a confidence interval estimation or near-optimal decision set with convergence guarantees would be more promising. In particular, the traditional statistical analysis tools may fail when the universal approximators, e.g., deep neural networks, are utilized. I am devoted to proposing solutions to ensure accountability via borrowing ideas from optimization and information theory.