EVOGENIO® - Evolutionäre Kunst
Dr. Günter Bachelier -
The second Evolutionary Art processes (2004-2007)
The second evolutionary art process was developed in 2004 using some additional concepts, such as a global image pool, image templates (as an analogy to the genome), multi-sexual recombination and specific types of meme reproduction, as a translation of the ontogenetic concept of spores or fruits.Overview
The process begins with the definition of a template image that
consists of several masks, each on one layer (see point 2 in fig. 1). A
multi-sexual reproduction process (see point 3 in fig. 1) exchanges
those masks randomly by images from a global image pool (see
point
1 in fig. 1) that consists of several classes of images. Recombination
strategies define which mask is exchanged by an image from which
class. This reproduction is multi-sexual because more than two layers
and image parents are always selected.Fig. 1: Overview of the second Evolutionary Art Process
After generating a population of about 100 individuals, the evaluation is done by the artist putting each image in one of three classes:
- "non-reproduction" class for images that do not match the aesthetic preferences and which will be deleted;
- "optimization" class for images that have a good overall impression but with local faults or defects which will be optimized by hand;
- "direct reproduction" class for images that match the preferences.
Additionally, they undergo a second reproduction phase where mathematical symmetry operators with random but constrained parameters are applied to them (see point 7 in fig. 1). Forty to eighty meme images were generated for each symmetry operator. These images were directly inserted into the global image pool, where every symmetry operator used has its own image class. Images from the first and second reproduction processes are then available for selection in the next iteration or generation, where the same or a different image template is used.
Global image pool
A huge global image pool consisting of pixel images is provided. The
images come from non-evolutionary art processes such as self-organizing
painting and from previous evolutionary runs. The image pool is
structured into different classes, (see imagetyp-T10, T14, T17, M, and
S
in fig. 1) depending on the origin and method of generation. Many of
the
image individuals from the first evolutionary art processes are
included
in a specific image type S, called "Spaces". These individuals are
often
used as background (first layer).The idea of a predominant image or meme pool is derived from the idea of genetic load {Born1978}: Preserved genes from the evolutionary history of a population or a species to be used now or later. These genes were originally included in the gene structure of the individuals, but in my evolutionary art process all memes were externalized in the image pool where they are used for recombination purposes. Such a predominant image pool has no close biological analogy because it represents not only all the genes of a present ecosystem, but also the whole history of such an ecosystem.
Initialization by determination of an image template
An important building block is the use of an image template in analogy
to a genome. A genome is defined by a certain number and position of
genes. A certain number of masks and their position to each other
exemplify the analogy in this evolutionary art process. Each mask is
located on a specific layer of the image template. The masks and their
composition in the template are generated by the artist and they
reflect
my interest in the dependencies between symmetry and symmetry breaking.
In most cases, bilateral symmetry is used for the masks.Recombination of image individuals
A population is defined by a number of image individuals. Those image
individuals are generated by using a multi-sexual recombination process
that randomly takes images of the image pool and exchanges the content
of masks by the corresponding content of the selected images (see
fig. 2). This recombination process has a similarity to the
multi-sexual
recombination of viruses where more than two individuals recombine to
build offspring. The gene pool is the set of all individuals that ever
existed which corresponds to the image pool.Fig. 2: Recombination with a template and masks
There are lots of possible recombination strategies which determine what type of image should be used for one of the masks. It is part of the artist's task to specify such strategies. An actually used recombination strategy specifies that for the background layer (layer 1) a random image is chosen from the class "Spaces from the year 2004", for layer 2 a random image is chosen from imagetyp-T14 (plane group p3m1), for layer 3 a random image of imagetyp-T14 or imagetyp-T10 (plane group p4m) is chosen, for layer 4 a random image is chosen from imagetyp-T17 (plane group p6m), and for layer 5 a random image is chosen from imagetyp-M (image individuals that were directly transferred to the image pool, see point 5 in fig. 1):
- layer 1: "Spaces from the year 2004"
- layer 2: image type T14
- layer 3: image type T10 or T14
- layer 4: image type T17
- layer 5: image type M
Evaluation
After having generated a population of roughly 100 image individuals,
the images are evaluated by the artist by putting them into one of the
following classes "non-reproduction", "optimization" or "direct
reproduction". Manual optimizing
The option of manually optimizing image individuals is specially
important when paths are used, given that the probability of images
with
local faults can be high in this instance depending of the
RST-parameters. A fault mask or stencil is built when the
RST-transformed image does not fully overlap its corresponding path.
Image individuals with one or more of such faults are called "non
valid". If the overall impression of such an image is good, then it is
selected for the "optimization" class and the fault is reversed by hand
by moving the RST-transformed image on the layer in such a way, that it
fully overlaps the path, followed by the merging of the image and path.The detection of non valid image individuals is done manually by the artist during the evaluation, however a script was developed which can automatically detect such individuals if a specific background color is used that is not intentionally used in any of the images. One of the drawbacks of this procedure is the production of "false positive" detections, i.e. images are marked as non valid because there are pixels of the specific color, though they all have valid masks. A threshold of pixels was introduced in order to minimize the false positive detections: the image classified as non valid only when the number of pixels with such color is above the threshold
Selection for Meme Reproduction
After the optimization of image individuals, the classes "direct
reproduction" and "optimization" are merged and then directly copied
into the image pool in the imagetyp-M class. One of the unique features
of my new evolutionary art process is that these images or a selected
subset of them undergo a second stage of reproduction, the meme
reproduction. In the present implementation, all images from the merged
classes are selected for this second reproduction.Image meme reproduction
This second stage of reproduction shares similarities with species
that produce fruits or spores. Those fruits or spores have none
or
few morphological similarities with the parents but they are able to
yield offspring by themselves.Usually, parts of image individuals are selected and a whole new image can be generated from this part using a system of simple image operations. Given that symmetry is one of my main interests, special mathematical operations (two dimensional symmetry groups or plane groups {GrünbaumShephard1986}) are used to perform this task. There are 17 plane groups and four of them generate seamless images (pmm, p4m, p3m1, p6m). After experimenting in 2004 with all plane groups, three (p4m, p3m1, p6m) of the four seamless plane groups are now used in my evolutionary art process. Every symmetry group defines its own class in the global image pool. The plane group p3m1 fits my aesthetic preferences very well and it is used the most in the meme reproduction process. From every selected image individual from the first reproduction process, 80 different meme images were generated with p3m1, where p4m and p6m generate 40 each.
The generation of a meme image with p3m1 will be described next in some detail. The process begins with the selection of an image part from the source image, e. g. M04-05-08b-1-026 (rotated 90° counter clockwise) in fig. 3).
Fig. 3: Building blocks of a seamless plane covering with p3m1
The selection with a polygon selection operation is an equilateral triangle. Its side length is a predefined constant or a function of the side lengths of the source image (e.g. 0.8 of the infimum of the two side lengths). The position of the triangle in the image M04-05-08b-1-026 is random and different in every of the meme reproduction processes. The selected image part, serves as a basic building block, G1, that can be copied and altered with some simple geometric procedures to generate five other building blocks, G2-G6, in the
following way (see fig. 3):
- G2 is generated if G1 is flipped right-left and than rotated by 60° clockwise (rotate[60](flip[rl](G1)))
- G3 is generated if G2 is flipped up-down (flip[ud](G2))
- G4 is generated if G1 is flipped up-down (flip[ud](G1))
- G6 is generated if G1 is flipped right-left and than rotated by 60° counter clockwise (rotate[-60](flip[rl](G1)))
- G5 is generated if G6 is flipped up-down (flip[ud](G6))
Fig. 4: Selection of a meme image from the seamless plane covering
Optional selection for insertion in the image pool
The images generated by the meme reproduction process can be copied
directly in the global image pool or they can be evaluated. The direct
copy option was implemented since evaluation by a human is not
reasonable, due to the large amount of meme images and because of the
unavailability of an aesthetic preference model that could evaluate
these images with machine learning methods. This approach is also
reasonable because experience has shown that meme images generated by
the selected plane groups fit, in most cases, the aesthetic
preferences,
provided the source image was selected according to the same
preferences.Selection for physical transformation
The selection for physical transformation is not part of the
evolutionary process in a strict sense, but it is part of my art
process
in general. It is not only selected here which image should be
physically obtained but also all other aspects related to this
decision:
Size, materials, kind of printing, number of copies printed, etc.The social sculpture Health Art
Since 2003/04 my evolutionary art process is embedded in my social
sculpture "Health Art", therefore the
selection for physical transformation is constrained by specific
techniques and materials which are compatible with this ecological and
health-conscious concept. First, the selected image individuals were
printed with acrylic on canvas and then sealed with shellack to avoid
toxic emissions from the paint. Then the canvas is integrated in the
picture frame with the
absorber
material that neutralizes and/or binds a large number of gaseous air
pollutants.Change of generations
In evolutionary art processes which are more closely related to the
biological processes, there is a change of generation after the
reproduction and selection of individuals. The selected offspring or a
mixture of parents and offspring build the next generation, followed on
by further reproduction and selection processes. In my evolutionary art
process there is no change of generation in the this sense, because
none
of the individuals are transferred directly to the next
generation. Selected individuals and the meme images are transferred
instead into the global image pool. This has a biological correlate to
species that produce fruits or spores, e.g., plants that live only one
year.Resulting Images
Fig. 5 shows two prototypical image individuals that were generated
with this evolutionary art process. They include most of the
factors of my current contrast aesthetics: Concrete art shapes and
Informel content, biomorphic {Gielis2003} and
hard-edge shapes, shapes with global bilateral symmetry, local and
global symmetries and symmetry breaking in the content.Fig. 5: Two examples of prototypical new image individuals: M04-05-08b-1-026 and M05-06-11-2-020
References
- J. Born. Evolutionsstrategien zur numerischen Lösung von Adaptionsaufgaben. PhD thesis, Humboldt-Universität, Berlin, 1978.
- Johan Gielis, editor. Inventing the Circle - The Geometry of Nature. Geniaal Press, Antwerpen, November 2003.
- Branko Grünbaum and Georey C. Shephard, editors. Tilings and Patterns. Freeman, New York, 1986.