Fidelity and complexity of standing group conversation simulations: A framework for the evolution of Multi Agent Systems through bootstrapping human aesthetic judgments

Simple rule based Multi Agent Systems are widely used in the fields of social simulations and game artificial intelligence in order to incorporate the complexity and richness of action and interaction into the characters in the virtual environments while keeping computational cost low. This paper presents an approach to synthesize the spatio-temporal dynamics of groups in standing conversation: four simple spatial rules form the building-blocks and a framework to automatically evolve rule and the parameter space by bootstrapping a-priori human judgment on the aesthetic quality of the simulations is introduced. The framework consists of a Genetic Algorithm and a scorer (fitness function) developed based on a machine learning system trained using human evaluations. The results of the study suggest that the framework is capable of deriving optimal rule and parameter combinations utilizing only a relatively small set of human scored training data. Further, the relationship between rule-complexity and visual fidelity is explored.