Understanding representational sensitivity in the iterated prisoner's dilemma with fingerprints

The iterated prisoner's dilemma is a widely used computational model of cooperation and conflict. Many studies report emergent cooperation in populations of agents trained to play prisoner's dilemma with an evolutionary algorithm. This study varies the representation of the evolving agents resulting in levels of emergent cooperation ranging from 0% to over 90%. The representations used in this study are directly encoded finite-state machines, cellularly encoded finite-state machines, feedforward neural networks, if-skip-action lists, parse trees storing two types of Boolean functions, lookup tables, Boolean function stacks, and Markov chains. An analytic tool for rapidly identifying agent strategies and comparing across representations called a fingerprint is used to compare the more complex representations. Fingerprints provide functional signatures of an agent's strategy in a manner that is independent of the agent's representation. This study demonstrates conclusively that choice of a representation dominates agent behavior in evolutionary prisoner's dilemma. This in turn suggests that any soft computing system intended to simulate behavior must be concerned with the representation issue.