Three-dimensional architectures grown by simple stigmergic agents

A simple model of multi-agent three-dimensional construction is presented. The properties of this model are investigated. Based on these properties, a fitness function is defined to characterize the structured patterns that can be generated by the model. The fitness function assigns a value to each pattern. The choice of the fitness function is validated by the fact that human observers tend to view patterns with high (resp. low) fitness as structured (resp. unstructured). A genetic algorithm based on this fitness function is used to explore the space of possible patterns. The genetic algorithm is able to make use of sub-modules of existing patterns and recombine them to produce novel patterns, but strong epistatic interactions among genes make the fitness landscape rugged and prevent more complex patterns from being produced.