"Preferences-dependent learning in the centipede game"

Astrid Gamba and Tobias Regner

When agents do not have access to public statistics about past interactions, they may form expectations about opponents' behavior based on their own past experiences. Agents endowed with heterogeneous preferences behave differently and acquire different pieces of information, thus forming heterogeneous and possibly wrong conjectures about the opponent's behavior. We provide experimental evidence that heterogeneous play of the Centipede game is the result of a learning process conducted by more and less altruistic agents receiving partial information about opponents' past plays. We manipulate the quality of information feedbacks after each round of play. If subjects rely only on their own past experiences, long run behavior resembles a self-confirming equilibrium, whereby more selfish subjects take at earlier nodes due to their lack of trust. Aggregate information release decreases heterogeneity of behavior by increasing the passing rates of selfish subjects and play moves towards Bayes Nash equilibrium.

Preferences-dependent learning in the centipede game