A value-relativistic decision theory predicts known biases in human preferences

Traditional models of decision-making assume the existence of an external frame of reference for measuring the value of possible outcomes. We show that this fundamental assumption prevents classical decision models from predicting realistic decision-making behavior. Making an alternative relativistic assumption about the nature of reward leads us to formalize a rational agent as one which minimizes its internal decision-computational costs while retaining satisfactorily predictive models of its external environment. In computational evaluation, our model replicates previously unexplained ‘irrational’ behavior of human subjects.