The Impact of Subcultures in Cultural Algorithm Problem Solving

Abstract

Cultural Algorithms are computational models of social evolution based upon principle of Cultural Evolution. A Cultural Algorithm consists of a Belief Space consisting of a network of active and passive knowledge sources and a Population Space of agents. The agents are connected via a social fabric over which information used in agent problem solving is passed. The knowledge sources in the Belief Space compete with each other in order to influence the decision making of agents in the Population Space. Likewise, the problem solving experiences of agents in the Population Space are sent back to the Belief Space and used to update the knowledge sources there. It is a dual inheritance system in which both the Population and Belief spaces evolve in parallel. In this paper we compare three different social fabrics (homogeneous, heterogeneous and Sub-Cultures) over a wide range of problem complexities. The performances of these three different evolutionary approaches are compared relative to a variety of benchmark landscapes of varying entropy, from static to chaotic. We show that as the number of independent processes that are involved in the production of a landscape increases, the more advantageous subcultures are in directing the population to a solution. Such landscapes are often characteristic of deep learning problems in which patterns are generated by the interaction of many simple interactions. While sub-cultured approaches can emerge in a given problem, they do not have to. It is shown that for single layer generators for a landscape or image, sub-cultures do not effectively emerge since they are not needed to solve such problems.