Multiple Agent Genetic Networks for Iterated Prisoner's Dilemma

Abstract

The well known game Iterated Prisoner's Dilemma (IPD) is examined as a test case for a new algorithm of genetic search known as Multiple Agent Genetic Networks (MAGnet). MAGnet facilitates the movement of not just the agents, but also the problem instances which a population of agents is working to solve in parallel. This allows for simultaneous classification of problem instances and search for solution to those problems. As this is an initial study, there is a focus on the ability of MAGnet to classify problem instances of IPD playing agents. A problem instance of IPD is a single opponent. A good classification method, called fingerprinting, for IPD exists and allows for verification of the comparison. Results found by MAGnet are shown to be logical classifications of the problems based upon player strategy. A subpopulation collapse effect is shown which allows the location of both difficult problem instances and the existence of general solutions to a problem.