Learning personalized exploration in evolutionary design using aesthetic descriptors

Abstract

We describe an aesthetic learning approach to one of the most challenging problems in interactive evolutionary design: modeling user preference for lessening the burden placed on users in hundreds of loops. In the approach, two aesthetic descriptors, high-level and low-level descriptors, are proposed based on pixel distribution and aesthetic criteria respectively. Starting with a collection of evaluated images, we apply both descriptors to the images, and then use decision tree learning algorithm to obtain the computational learning model. The model is adopted to automatically guide the subsequent generations. Classification and evolutionary results are shown in our experiments to evaluate the learning model and compare the two descriptors’ learning ability in the evolutionary runs. The reported results indicate that high-level descriptors are more appropriate in approximating user’s implicit aesthetic intentions for solving the problem considered.