The effect of primitive sets on the expression of evolved images

Genetic programming for evolutionary art often focuses on improving fitness functions to improve image quality. We take the opposite approach, to determine how influential the internal representation is on the expression of images. We define four primitive sets based on common ideas from the literature and compare the resulting images based on a constant fitness function. Although it is obvious that changing the primitive set has an effect on the resulting images and their fitness, it has not been thoroughly investigated. This paper explores the effect of changing primitive sets in genetic programming for evolutionary art. We find that different primitive sets have different effects on how the final image looks as well as how it is affected by genetic operators. We find that geometric primitive sets are better for creating recognisable images, and are able to make small localised changes to the image over generations. In contrast the mathematical primitive sets result in intricate patterns over the whole image, and a small change in the tree can result in a change across the whole image rather than in a localised area.