Examination of graphs in Multiple Agent Genetic Networks for Iterated Prisoner's Dilemma

Abstract

Multiple Agent Genetic Networks (MAGnet) are spatially structured evolutionary algorithms which move both evolving agents as well as instances of a problem about a combinatorial graph. Previous work has examined their use on the Iterated Prisoner's Dilemma, a well known non-zero sum game, in order for classification of agent types based on behaviours. Only a small complete graph was examined. In this study, a larger set of graphs with thirty-two nodes are examined. The graphs examined are: a cycle graph, two Peterson graphs with differing internal rings, a hypercube in five dimensions, and the complete graph. These graphs and properties are examined for a number of canonical agents, as well as a few interesting types which involve handshaking. It was found that the MAGnet system produces a similar classification as the smaller graph when the connectivity within the graph is high. Lower graph connectivity leads to a process by which disjoint subgraphs can be formed; this is based on the method of evolution causing a subpopulation collapse in which the number of problems on a node tends to zero and the node is removed.