Varying Decision Inputs in Prisoner's Dilemma
Submitted to CIBCB 2005
Daniel Ashlock,
with LeeAnn Barlow

Abstract PDF eprint

This study continues an investigation into factors that can modify the emergence of cooperation in the iterated Prisoner’s Dilemma. It is part of a project to construct agents that play the Prisoner’s Dilemma in a manner similar to biological agents; in this study a representation called a binary decision automata is used. Binary decision automata are finite state machines that are given a selection of Boolean inputs that describe features of the game. Each state both specifies one of the variables and generates transitions and actions based on the value of the variable. The software permits the automata to see a subset of the possible decision variables and the decision variables made available have a strong impact on the way agents trained with an evolutionary algorithm behave. A collection of twenty-two game descriptors are used; six are based on recent information about past play, sixteen are features that are based on long-term information about play. The level of cooperation and other measures of behaviour all vary strongly with the set of variables made available to the agent. In this study the level of cooperation is assessed at different evolutionary epochs to permit the evaluation of how cooperation emerges over time. It is found that the most importance variable for the emergence of cooperation was the variable that checks whether a player’s opponent cooperated last time; an unexpected development is that the most used variable in evolved automata was one that checks to see if the automata is being exploited.