‐Learning

Q‐learning is a regression‐based approximate dynamic programming algorithm that is commonly used to estimate sequences of decision rules that maximizes mean utility when applied to the population of interest. Because Q‐learning is based on fitting a series of regression models, it is (i) highly configurable in the sense that these regression models can be chosen to be parametric, semiparametric, or nonparametric; and (ii) extensible to settings with censored, high‐dimensional, or missing data. This entry reviews Q‐learning for decision problems evolving over a finite time horizon including a brief review of open questions and active lines of research.

Duke Scholars

Author Marie Davidian Biostatistics & Bioinformatics, Division of Biostatistics