PhD Thesis Defense: Karine Karine, Methods and Tools for Learning from Uncertain and Incomplete Longitudinal Data
Content
Speaker:
Abstract:
In this thesis, we study methods and tools for learning from uncertain and incomplete longitudinal data. (1) First, we introduce a toolbox for Bayesian probabilistic modeling of longitudinal data that can help researchers in other domains, such as behavioral domain, do modeling. (2) We explore the application of RL methods to the problem of learning intervention option selection policies. We study the effect of context inference error and partial observability on the ability to learn effective policies. We also create a simulation environment that mimics the complex dynamics of the participant's behaviors in a mobile health app study, which can be very useful for testing RL algorithms for the behavioral domain. (3) We introduce a practical method using batch Bayesian optimization that learns the delayed effect of actions when using Thompson sampling (TS) and shows good performance in a low number of iterations. (4) Finally, we propose a novel method that uses large language models (LLMs) to update the behavior of an RL policy to accelerate personalization in adaptive intervention. We extend our new simulation environment to add support for LLMs.
Advisor:
Benjamin Marlin