Procedural Texture Evolution Using Multiobjective Optimization [PDF]
Brian J Ross and Han Zhu, July 2002.
This paper investigates the application of evolutionary multiobjective optimization to two-dimensional procedural texture evolution. Genetic programming is used to evolve procedural texture formulae. Earlier work used multiple feature tests during fitness evaluation to rate how closely a candidate texture matches visual characteristics of a target texture image. These feature test scores were combined into an overall fitness score using a weighted sum. This paper extends this research by replacing the weighted sum with a Pareto ranking scheme, which preserves the independence of feature tests during fitness evaluation. Three experiments were performed: a pure Pareto ranking scheme, and two Pareto experiments enhanced with population divergence strategies. One divergence strategy scores individuals using their nearest-neighbour distance in feature-space. Another scheme uses a normalized, ranked measurement of nearest neighbour distance. A result of this work is that acceptable textures can be evolved much more efficiently with MOP evolution than compared to the weighted sum approach done earlier. The ability to create a diverse selection of Pareto solutions is advantageous in this problem domain, since the acceptability of a final texture is ultimately a subjective, aesthetic decision by the user.