Aesthetics as a Criterion: Navigating Solution Spaces Utilizing Computer Vision, the Aesthetic Measure, and Artificial Neural Networks

A computational framework is proposed to quantify the aesthetic experience of perceiving architecture. This framework is used as the criterion for navigating a design solution space. A parametric model’s solution space is represented as a set of computer-generated images. We utilize computer vision to recognize parts and the relations between parts and the whole. These values are inserted into an adaptation of Birkhoff’s aesthetic measure formula, calibrated via crowdsourced hedonic response. An artificial neural network (ANN) model is trained to predict the hedonic response of images using (A) parts relations as inputs and (B) the average hedonic responses as target outputs. The prediction of the ANN is used as a fitness function for optimization. The same public evaluates the parametric model’s output to compare the ANN model’s effectiveness in predicting their response. The method is useful in translating formal qualities into quantities and navigating solution spaces, especially with surrogate models.