An Experimental Analysis of the Two-Armed Bandit Program

Abstract

We investigate, in an experimental setting, the behavior of single decision makers who at discrete time intervals over an "infinite" horizon may choose one action from a set of possible actions where this set is constant over time, i.e. a bandit problem. Two bandit environments are examined, one in which the predicted behavior should always be myopic (the two-armed bandit) and the other in which the predicted behavior should never be myopic (the one-armed bandit). We also investigate the comparative static predictions as the underlying parameter of the bandit environments are changed. The aggregate results show that the cutpoint behavior in the two bandit environments are quantitatively different and in the direction of the theoretical predictions. Furthermore, while a significant number of individual cutpoints exhibit nonstationarity (contrary to the theory), the most likely, i.e. maximum likelihood estimates, collection of decision rules that best explain overall behavior are those that are consistent with the underlying theory.