Evolutionary art is a form of generative art in which works of art from the visual arts, music and performing arts are created using evolutionary algorithms. Evolutionary algorithms are methods for solving optimization problems using the principles of natural evolution. By conceiving of artistic processes as an optimization, can be created objects which have an aesthetic effect on people . For time reasons, the implementation of this class of algorithms is mandatory with the computer, but in principle could also be calculated by hand. Evolutionary art is therefore part of digital art .
The basis of evolutionary art, like all evolutionary algorithms, is a population of individuals, each representing a visual structure. This representation can be either indirect, as individuals, as in genetic programming, each containing a program that creates a visual structure, so that the biological distinction between genotype and phenotype is maintained. However, representation can also be direct, as in evolutionary strategy , by considering an individual only as a phenotype to which evolutionary operations are applied. In this case, an individual includes a picture, drawing, moving picture or the like in the sense of a image file or video file .
Almost all applications of evolutionary art using indirect representations produce non-representational visual works. Regardless of whether direct or indirect representation, there are only a few approaches to representational evolutionary art.
In the evolutionary art process, a starting population of individuals is first defined. In an indirect representation – as usual in genetic programming – random programs and thus random visual structures are generated. In a direct representation, non-random visual structures are usually selected by the artist, e.g. images from previous evolutionary runs.
There follows a reproduction phase in which the present individuals are propagated according to a reproduction strategy by applying recombination and mutation operations to the representational structures. The nature of these operations depends on the nature of the programs or direct visual structures, just as in evolutionary algorithms, for example, linear and hierarchical structures of individuals require customized recombination and mutation operations.
Part of the reproduction strategy is the way individuals are selected for recombination (selection for reproduction). If the reproduction strategy is based on genetic algorithms , fitness values must be available for each individual beforehand . The frequency of selection for reproduction is a strictly monotonous function of this fitness, e.g. the higher the fitness, the greater the probability of selection. If the reproduction strategy is based on evolutionary strategies, the selection is equally distributed randomly.
After the reproduction phase, there will be a population of offspring, each of whom will need to determine a fitness score that will somehow reflect the aesthetics of the visual structures. An algorithmic determination of these values would require a formal aesthetic model that is not or only partially available in previous methods for evolutionary art. Therefore, algorithmic methods are limited to the determination of simple properties of the image analysis and models based thereon. The definition of fitness by a human or a group of people is widespread (interactive evolution). Usually this is the artist who sets the ratings according to his subjective aesthetic criteria. For example, alternative methods of empirical fitness estimation is the time a viewer views a visual structure presented to them. There are also preconscious methods in which an attempt is made to derive a correlation between physiologically measurable characteristics of an observer and his aesthetic evaluations (eg, pupil reactions ). The most innovative approaches are offered here by neuroesthetics which identifies brain regions involved in aesthetic assessments that are supposed to produce correlations between the activities of these regions and aesthetic assessments (analogous methods such as neuromarketing). However, as these approaches require complex and still very expensive medical imaging equipment , their use in evolutionary art has so far been limited to a few small studies.
If parents and offspring each have a fitness value, a selection strategy is used to determine which individual may continue to exist in the next generation and possibly reproduce. This selection strategy considers either only the offspring or the union of parents and offspring. Furthermore, if no abort criterion, such as reaching a predetermined maximum number of generations, is reached, the next iteration of the evolutionary art process is started with a new reproduction phase.
One application of evolutionary art is non-photorealistic rendering , a field of computer graphics in which graphics are deliberately not faithfully represented in their physical image. An example is the generation of an artificial painting from a photograph . The British scientists Collomosse and Hall developed in 2005 an algorithm that produces paintings based on photographs. A painting is understood as a sequence of brushstrokes, with brushstrokes defined by attributes such as position, direction, color, etc. A genetic algorithmis used to search the space of all possible paintings in this way. The fitness function , which assigns a quality to each solution candidate , compares the edge image of a candidate to an initially calculated salience image . The salience of a picture detail indicates how obvious it is to a human viewer. In the algorithm of Collomosse and Hall, the salience of image details is made up of three factors: rarity level, degree of visibility, and a third factor that first learns the taste of users in subregions to distinguish important artifacts from unimportant ones.
The salience calculation is based on the idea that works of art represent “not a mirror” ( Ernst Gombrich) of reality, but rather an interpretation of the artist.