EVOGENIO^® - Evolutionäre Kunst

Dr. Günter Bachelier -

The first Evolutionary Art processes (1995--2003)

As previously stated, in this context evolutionary art is seen as art created by a method that resorts to evolutionary concepts. Since I use the help of computers in my evolutionary art method, it can be explained by the description of a generic evolutionary algorithm with four main components: population, evaluation, selection and reproduction.  The goal was to transfer these components to an artistic context (here: visual art) and  determine what could be the meaning of these central concepts in the scope of visual art. There is no obvious or single answer to this question, since a huge variety of possible evolutionary art processes are possible.

 

Population and individuals

Although most conventional evolutionary art approaches resort to an expression-based image representation,  all the individuals in my evolutionary art processes are bitmap images. The population is simply a set of such individuals i.e. it has no other internal structure such as a graph. In  contrast to expression-based representations, I have named this approach "data-based" evolutionary art. It is possible to use the whole range of concepts that have been implemented in graphic file format, such as alpha channels and layers. File formats using  meta-information (commentary, IPTC-Header, XML tags, ...) can  be used directly  as image individuals, since the fitness value can be saved as a kind of meta-information in the file.

The seeding operation, i.e. the creation of the first population, is an important part of my approach. Typically, in expression-based approaches, the individuals of the first population are created through the random generation of the corresponding expression trees. A random initialization in the case of data-based evolutionary art is not reasonable because such an image would not have any structures. It is more efficient and effective to start with images from outside an evolutionary process which have already satisfied to some extent the aesthetic preferences of the artist.

Evaluation and Selection

After the initialization and after the generation of offspring, the individuals are evaluated. In most evolutionary algorithms this is done with a fitness function. However, in the case of interactive evolutionary art, the quality of an individual reflects the aesthetic preference of the artist who is evaluating the image individuals. In my evolutionary art processes there is no explicit fitness function, therefore this is an interactive evolution approach, requiring input from an external source, the artist. A binary evaluation is applied as strategy to select the individuals and parents for the next generation. My evolutionary art process depends solely on my aesthetic judgments, images that do not satisfy my aesthetic needs are not selected and are deleted at a later stage.

 

Reproduction

Reproduction  is the operation by which new individuals (offspring) are produced from the genetic code of one or more  mature individuals (parents). It is necessary to define recombination and mutation operators suitable for  data-based evolutionary art. If we use bitmap images as individuals, the interpretation of reproduction with the two components (recombination and mutation) is not that obvious, but it is clear that the interpretation is different from expression-based evolutionary art. My solution, in 1995, was to apply some image processing functions as reproduction operators and to use random but constrained parameters of such functions so as to introduce variations in the next
generation.


Fig. 1: Recombination and mutation in the first evolutionary art  processes

Recombination

One possible definition of recombination, used in my art process, is an analogy to the crossover operations in genetic algorithms, or to the discrete recombination operator in evolutionary strategies. Segments of one parent image are selected which, in turn, build an offspring image with the complementary parts of the second image. The concept through which this is achieved is called Regions-of-interest (ROIs), i.e., possible overlapping segments of an image defined by the artist.

A simple reproduction strategy that was used first selects at random two parents from the image population (see fig. 1). One of them is selected randomly as the primary parent who is copied to one layer of the offspring individual. Then ROIs of the secondary parent are selected randomly and they are masked with non-sharp edges and later copied to a higher layer of the offspring.

This reproduction strategy can be generalized to multi-sexual reproduction if the offspring obtain their genetic material (image components) from more than two parents, i.e. a second, third, ... parent inserts their selected ROIs in the copied primary image.

Mutation

Copying images and image parts in an offspring individual is the recombination part of the reproduction process. Additional variation (mutation) is introduced by transforming the transmitted regions by means of image  processing operations with randomized parameters within certain constraints. The  primary image is not just copied but also undergoes transformations. In most cases, a RST-transformation (Rotation-Scaling-Translation) was used i.e., the image of the primary parent and the ROIs are rotated by an angle between 0 and 360 degrees, they are scaled by a scaling factor, e.g. between 0.8 and 2.4, and the result is moved in the x and y direction.

After recombination with mutation, the parents and offspring build an interim population  and the elements from this population are evaluated, i.e., the  artist decides if the images are compatible with his aesthetic preferences. The images that survive this evaluation are selected for the next generation.