Creating continuous design spaces for interactive genetic algorithms with layered, correlated, pattern functions.

摘要

Interactive evolutionary design (IED) is a design paradigm that can be used to generate computer graphics content by means of artificial evolution. Traditional evolutionary design research relies on objectively computable fitness functions to evaluate the quality of individuals in a population of potential solutions to a design problem. IED systems rely instead on subjective judgment to determine fitness.; Most implementations of IED systems demonstrate significant signature. The term signature refers to the lack of visual diversity in the populations and individuals generated by IED systems. Signature is primarily the result of the solution space representation. Frequently, primitives and techniques are used which are not sufficiently general, or are biased towards specific visual qualities. In practice, such systems are only able to access a small region of a problem domain's ideal potential solution space. Alternatively, too general a representation is used, resulting in the need to search too large a region of solution space. This makes it impractical for an interactive system to be used to find fit individuals.; One of the most common uses of IED is the generation of nonrepresentational images, usually either for artistic purposes or for use as textures or surface shaders. Focusing on this problem domain, continuous pattern functions are introduced and used as a new genetic primitive in an evolutionary design context. Abstracted from pattern based procedural texturing techniques, continuous pattern functions are defined in order to produce a wide range of patterns and forms for generating images and surfaces. Their flexibility enables greater visual diversity and control of visual attributes than has previously been demonstrated in IED image systems. Formal graphic design knowledge is integrated into continuous pattern functions to further increase the visual diversity of generated populations. Finally, layer-based cloning methods are introduced to address the “synchronization problem” of smoothly facilitating feature correlation.