The rationality of prejudices.

Abstract

We model an N-player repeated prisoner's dilemma in which players are given traits (e.g., height, age, wealth) which, we assume, affect their behavior. The relationship between traits and behavior is unknown to other players. We then analyze the performance of "prejudiced" strategies--strategies that draw inferences based on the observation of some or all of these traits, and extrapolate the inferred behavior to other carriers of these traits. Such prejudiced strategies have the advantage of learning rapidly, and hence of being well adapted to rapidly changing conditions that might result, for example, from high migration or birth rates. We find that they perform remarkably well, and even systematically outperform both Tit-For-Tat and ALLD when the population changes rapidly.