Abstract:

Reinforcement learning has become an essential and powerful tool for modeling sequential decision data and inferring the optimal decision rule. Although enjoying enormous success in various fields, it still faces critical challenges in applications in medicine and human behavior studies, where the sample size can be small, noise is large, and unobserved confounders could be present. In this talk, we introduce two recent works that separately address two issues. One is a regularized framework that leads to more conservative and potentially safer treatment rules. This method is applied to an insulin dose-finding problem for diabetic management. Another is proposed to address the unobserved confounding issue in a partially observed Markov decision process setting. We utilize the proximal causal framework to estimate the value function of any potential treatment strategy. This approach is applied to a family relationship study that aims to understand the strategies for improving romantic relationships.

报告人简介：

Dr. Ruoqing Zhu is an Associate Professor of Statistics at the University of Illinois at Urbana Champaign. Dr. Zhu received his Ph.D. of Biostatistics in 2013 from the University of North Carolina at Chapel Hill and worked as a postdoctoral research associate at Yale University from 2013 to 2015 before he joined the department of Statistics at UIUC. His research interest lies in personalized medicine, reinforcement learning, random forests, survival analysis, sufficient dimension reduction, and machine learning applications in biomedical problems such as infectious diseases, nutrition health, diagnostics and cancer. He is affiliated with the Carle Illinois College of Medicine, the National Center for Supercomputing Applications and the Carl R. Woese Institute for Genomic Biology. He is currently serving as an Associate Editor at Journal of the American Statistical Association and Statistical Analysis and Data Mining.