Bayesian Learning in UnstableSettings: Experimental Evidence Based on the Bandit Problem

We study learning in a bandit problem where the outcome probabilities of six arms switch (jump) over time a restless bandit. In the experiment, optimal Bayesian learning tracks the jumps through learning of the probability of a jump or direct jump detection and, once a jump has occurred, re-learns the outcome probabilities. Such Bayesian learning is much more complex than the natural alternative which learns through trial-and-error (adaptive expectations). Yet, when combined with a partially myopic decision rule, Bayesian learning better matches the behavior observed in the lab. This result suggests that agents may be less limited in their computational capacities than previously thought, and that complexity does not always hamper fully rational learning.