Abstract:

Mobile health (mHealth) provides effective ongoing support in everyday life to help users sustain a healthy lifestyle. In the first work, we aim to prepare for a new mHealth clinical trial, where the reward to be improved is users' commitment to physical activity (PA). A personalized policy decides whether and when notifications are pushed to mobile devices to prompt short bouts of activity, a mediator of the reward. However, the reward is sparsely observed through weekly or monthly surveys, creating challenges for learning the policy efficiently. We develop a reinforcement learning algorithm that sequentially updates the personalized policy based on these sparsely observed rewards, leveraging causal information provided by domain experts to speed up learning.
Following an adaptive clinical trial that sequentially updates the data collection policy, the data can be used to learn a personalized treatment policy for future patients. However, there is a tradeoff between the benefits for patients in the clinical trial and the benefits for future patients if the learned policy is deployed. In the second work, we explore the tradeoff between the training and test values of the estimated policy in single-stage decision problems.

报告人简介：

Daiqi Gao is a postdoctoral fellow in the Department of Statistics at Harvard University, working with Professor Susan Murphy. She received her Ph.D. from the Department of Statistics and Operations Research at the University of North Carolina at Chapel Hill, where she was advised by Professors Yufeng Liu and Donglin Zeng. Previously, she earned her B.S. in Industrial Engineering and Statistics from Tsinghua University. Her primary research interests are in statistical reinforcement learning and machine learning, with applications in mobile health and personalized medicine.