Statistics 711: Multi-Stage Decision Problems & Dynamic Treatment Regimes.

Instructor: Susan A. Murphy, samurphy@umich.edu

Time/Place: Monday – 3:00-5:00 p.m. – 245 Dennison.

What do Dynamic Treatment Regimes and Reinforcement Learning have in common? Reinforcement learning is body of methods for solving multi-stage decision problems. And dynamic treatment regimes are the policies constructed in reinforcement learning. Dynamic treatment regimes are time-varying, individually tailored, treatments for medical disorders. To tailor the treatment to the individual we use decision rules that assign treatment based on evolving individual information. The goal is to make these decisions so as to maximize selected individual responses. In this course we will be interested in combining statistical techniques with reinforcement learning. We will be particularly interested in methods that can be used with a training set (e.g. sample) of finite horizon trajectories collected in clinical trials. Students taking this class will have the opportunity to discuss papers ranging from the mathematical to the applied.

Requirements:

(1) Ability/desire to read papers across a variety of disciplines.

(2) Interest in learning theory (e.g. assessment of generalization error).

(3) Graduate statistical/biostatistical/engineering background.