Discussion papers

CPM-02-96 - 2 May 2002

Understanding Our Creations: Using Machine Learning to Understand Machine models.

David Hales


Artificial Society and Artificial Life computer models often contain many exogenously defined parameters. Such parameters may take a number of possible values. Alternative values represent alternative model assumptions. The space of all possible values comprises a parameter space of model variants. Results of experimentation with such models are often based on execution of the model for a fixed set of (arbitrary) parameter values. Given some (possibly) emergent phenomena of interest, for which the model has been constructed to investigate, such a method gives a simple “existence proof”. An existence proof demonstrates that at least some parameter values are sufficient to produce the phenomena of interest. However, a deeper understanding of the model can be achieved by exploring the parameter space.

In this paper an artificial society model (the StereoLab) is outlined which contains cooperating interacting agents. Cooperation between agents is mediated by stereotypes that are held by agents. The stereotypes evolve memetically via cultural interaction and mutation. The model specifies a large number of exogenously defined parameters that determine various aspects of agent and environmental behaviour. Two methods of exploring this very large parameter space are applied in the search for regions which produce high levels of agent cooperation. By using these techniques a region is located in which a novel cooperation forming process is identified based on the formation of agent groups.