Nonparametric learning rules from bandit experiments: The eyes have it!

How do people learn? We assess, in a distribution-free manner, subjects' learning and choice rules in dynamic two-armed bandit (probabilistic reversal learning) experiments. To aid in identification and estimation, we use auxiliary measures of subjects' beliefs, in the form of their eye-movements during the experiment. Our estimated choice probabilities and learning rules have some distinctive features; notably that subjects tend to update in a non-smooth manner following choices made in accordance with current beliefs. Moreover, the beliefs implied by our nonparametric learning rules are closer to those from a (non-Bayesian) reinforcement learning model, than a Bayesian learning model.