In this paper we present a new technique for automatically approximating the aesthetic Fitness of evolutionary art. Instead of assigning fitness values to images interactively, we use the Universal Similarity Metric to predict how interesting new images are to the observer based on a library of aesthetic images. In order to approximate the Information Distance, and find the images most similar to the training set, we use a combination of zip-compression, for genomes, and jpeg-compression of the final images. We evaluated the prediction accuracy of our system by letting the user label a new set of images and then compare that to what our system classifies as the most aesthetically pleasing images. Our experiments indicate that the Universal Similarity Metric can successfully be used to classify what images and genomes are aesthetically pleasing, and that it can clearly distinguish between "ugly" and "pretty" images with an accuracy better than the random baseline.
展开▼
