Strategy Inference in a Prisoner's Dilemma Game

We experimentally test the ability of subjects to identify repeated-game strategies from prisoner's dilemma choice data. In the experiments, subjects use a finite state grammar to build models to fit game histories. The histories are designed to distinguish between strategies with empirical and theoretical validity. We find that subjects successfully identify unconditional, punishment, and counting (or opportunistic) strategies. When data are observationally equivalent among different strategies, punishment strategies tend to be inferred. When the data are noisy, subjects' strategies overfit. The inferred strategies that do not fit the data tend to be more complex than necessary, and contain the behaviors of positive reciprocity and forgiveness, though forgiveness is not part of any relevant strategy that fits the data. Our experiment provides an empirical basis for repeated-game strategies