Satisficing and Learning Cooperation in the Prisoner s Dilemma

Jeff L. Stimpson, Michael A. Goodrich, Lawrence C. Walters

Abstract: The prisoner's dilemma is a useful model for studying the balance between self-interest and group-interest in multi-agent systems. Although many stratergies have been developed that perform well, most of these stratergies make strong assumptions about the information available to the agent. It is in this context that we describe a satisficing learning stratergy for the prisoner's dilemma and present evidance that stable outcomes other than the Nash equilibrium are possible. In addition, we offer emperical evidence that under typical circumstances, mutual cooperation is the most likely outcome and identify conditions under which two satisficing agents will learn to cooperate.