An Exploration of Differential Utility in Iterated Prisoner's Dilemma

While prisoner's dilemma has frequently been used in studies of animal behavior, past work in the area fails to address the question of differential utility. As animals needs change, their behavior also changes. A hungry animal may be less likely to cooperate than a full one. In this study an abstract species is modelled using an energy balance computation that assigns a state of hungry, full, or neither to a given animal together with a modified finite state representation for a prisoner's dilemma strategy that implements distinct but linked strategies for each possible state. The strategy used is conditioned on the animal's current hunger state. An evolutionary algorithm is used to generate effective strategies for this deterministic form of differential utility prisoner's dilemma. A tool called fingerprinting is used to document the strategies that arise. It is shown that the strategies that arise for the three possible hunger states have different distributions among known strategies and also differ from a baseline experiment in which agents are evolved to play prisoner's dilemma without any form of differential utility