Learning Genetic Representations as Alternative to Hand-Coded Shape Grammars

Abstract

Shape grammars have been used to analyze and describe designs, and to create new designs that are similar in style to the designs the grammar is based on. The grammars are created by hand, involving a large amount of research about the designs and the design process. This paper proposes a different approach, where a system is given design examples, and in a bottom-up process learns stylistic features of the examples. This is achieved by using an evolutionary system that is able to change the representation it is using. With the creation of a more and more complex evolved representation, the search space of the evolutionary process is transformed so that the search for new designs is biased towards designs similar to the design examples.