Multiple-process procedural texture

Abstract

Our newly developed generation of procedural textures (GPT) system automatically generates procedural textures for the computer graphics industry. The system makes use of hybrid parallel Monte Carlo tree search and gender-based genetic algorithm modules that share a common multiple-generation population of procedural textures and a knowledge database. It also uses a multi-objective fitness function. The parallel Monte Carlo tree search module was inspired by gaming algorithms. To speed up the search, this module is enhanced with knowledge from previous successfully created procedural textures or tree node analyses. The gender-based genetic algorithm module automatically simulates several key features in natural selection and uses a multiple-generation breeding population, the notion of gender, and the concept of aging. This maintains diversity while providing many breeding opportunities for highly successful offspring. A third module selects generated shaders from the multiple-generation population and mutates them by replacing nodes with subtrees using the knowledge database. We evaluated the fitness quality of each module and compared the fitness quality of the system running in both single- and multiple-process mode. The optimal fitness quality was achieved by executing the system in multiple-process mode using a hybrid of these modules. We give examples of the GPT running in interactive mode, where a user directs the search towards the desired look using an esthetic evaluation.