The Impact of Elite Fraction and Population Size on Evolved Iterated Prisoner's Dilemma Agents

Daniel Ashlock and Eun-Youn Kim
Submitted to CEC 2016

Abstract PDF eprint

The iterated prisoner's dilemma is a simultaneous two-player game widely used in studies on cooperation and conflict. Past work has shown that the choice of representation or available resources such as the number of states or neurons of evolving agents has a large impact on the behavior of evolved agents. This study revisits three qualities of the agent training algorithm for finite state agents to examine their impact on agent behavior: population size, elite fraction, and the number of states the agent is permitted. All three of these algorithm parameters are shown to have an impact on the character of evolved agents. Assessment of agent behavior is performed using three tools. The first is play profiles which bin the ranges of score space. The total score assessment, a global characterization of the type of play that occurs over the course of evolution, is the second assessment used. For the third assessment, an analysis of the ability of agents with different numbers of states to compete with one another is performed. High elite fractions in the training algorithm are found to encourage cooperation. Larger populations increase cooperation for populations of agents with small and intermediate numbers of states but have little effect for agents with large numbers of states. As in past studies with fixed population size and elite fraction, agents with large numbers of states are found to have more diverse and less cooperative behavior. Having more states is also found to grant a competative advantage.