Learning to cooperate in the Prisoner's Dilemma: Robustness of Predictions of an Instance-Based Learning Model

The dynamics of cooperation in repeated Prisoner's Dilemma (PD) interactions are captured by an instance-based learning model that assumes dynamic adjustment of expected outcomes (IBL-PD model). This research presents this model's predictions across a large number of PD payoff matrices, in the absence of human data. Rapoport and Chammah (1965) test three hypotheses in a large set of PD payoff matrices: (1) as reward of cooperation increases, more cooperation is observed; (2) as the temptation to defect increases, less cooperation is observed; and (3) as punishment for defection increases, more cooperation is observed. We demonstrate that the same IBL-PD model that was found to predict the dynamics of cooperation in one particular payoff matrix of the PD produces accurate predictions of human cooperation behavior in six additional games. We also make detailed predictions of the dynamics of cooperation that support these three hypotheses.