Improving the Scalability of Generative Representations for Openended Design

Abstract

With the recent examples of the human-competitivenessof evolutionary design systems, it is not of interestto scale them up to produce more sophisticated designs.Here we argue that for computer-automated designsystems to scale to producing more sophisticatedresults they must be able to produce designs withgreater structure and organisation. By structure andorganization we mean the characteristics of modularity,reuse and hierarchy (MR&H), characteristics thatare found both in man-made and natural designs. Weclaim that these characteristics are enabled byimplementing the attributes of combination,control-flow and abstraction in the representation, anddefine metrics for measuring MR&H and define twomeasures of overall structure and organisation bycombining the measures of MR&H. To demonstrate themerit of our complexity measures, we use anevolutionary algorithm to evolve solutions to differentsizes for a table design problem, and compare thestructure and organisation scores of the best tablesagainst existing complexity measures. We find that ourmeasures better correlate with the complexity of gooddesigns than do others, which supports our claim thatMR&H are important components of complexity. Wealso compare evolution using five representations withdifferent combinations of MR&H, and find that thebest designs are achieved when all three of theseattributes are present. The results of this second setof experiments demonstrate that implementingrepresentations with MR&H can greatly improvesearch performance.