Computational aesthetic evaluation: steps towards machine creativity

Programmer and artists have invented a broad range of generative systems that create art and music. These powerful systems sometimes produce results that surprise their human collaborators, but the surprises are not always welcome or useful. Machine creativity needs a computational self-critical function that can guide generative systems toward valuable creative output.

This course provides a fast-moving, state-of-the-art overview of computational aesthetic evaluation. Some notable limited successes aside, computational aesthetic evaluation is far from a solved problem, and a "how to" course is not possible at this time. The intent of this course is to identify all of the significant trail heads, to share what previous explorers have found, and to encourage future journeys by artists and researchers along the paths that seem most promising.

The course begins with a brief summary of terminology, then reviews classic formulaic and geometric theories of aesthetics that are possibly amenable to digital exploitation, including Birkhoff's "aesthetic measure", the golden ratio, Zipf's law, fractal dimension, basic gestalt design principles, and the rule of thirds. A section on evolutionary art systems focuses on aesthetic evaluation in fitness functions, including interactive systems, strategies for automated evaluation such as performance goals, error measures, complexity measures, multi-objective and Pareto optimization, and biologically inspired methods that produce emergent aesthetic fitness functions such as coevolution, niche construction, swarm behavior, and curious agents. The course concludes with a review of the future of computational aesthetic evaluation, recent developments in the empirical study and psychological modeling of aesthetics, and the nascent field of neuroaesthetics.