Activities/Kandid: Difference between revisions

Kandid is a tool for breeding graphics. In Kandid a population consists of images looking like abstract

art or geometric patterns. These graphical forms are not drawn by hand.

Instead new forms can be found using an interactive genetic algorithm.

The user tells the program which images are 'cool' or never seen

Darwinian evolution consists of a simple cycle. The most fit entities in a population reproduce. The resulting offspring are a combination of their parents with some random mutations. Some offspring may be improvements over their parents. The most fit of offspring continue to reproduce. In Kandid the definition of fitness comes from the users aesthetic preferences. The program is a working horse presenting the user proposals for new images letting the user the decision which one is preferable. The concept behind Kandid was introduced by Karl Sims 20 years ago. It is a combination of image rendering algorithms, genetic programming and interactive aesthetic selection. Back in 2003 I published a different version of Kandid based on fractal geometry at http://kandid.sourceforge.net. Kandid for the Sugar desktop is more than a simple port. It is based on different graphic algorithms and allows the user more introspection of how the images are composed.

You can consider Kandid as a simulation of evolution. But compared to the evolution of live there are some differences. The first is how fitness is handled. In Kandid a human selects the proposed images. Next issue is the limited range of output. The program can produce lots of color combinations and geometrical shapes. But after using Kandid for a while the style of the images is predictable.

You can use Kandid to improve you aesthetic feeling. Most of the random images are too indifferent. But sometimes you find a surprisingly interesting composition of shapes or a god looking color combination. Maybe you see things you never noticed before.

You can use Kandid to get in touch with art made by machines. Or more precisely: Art made by humans trying to control a generative random process. You can see the success and the limits of this algorithmic approach for making art.

Kandid is also about complexity. Every single image produced by Kandid is made of simple shapes and color schemata. But combining these parts will result in a complex looking image. The result is more than simply summing up the details.

Always mark the image you find most appealing with a high ranking

and set the images, that should not appear again in the next generation,

=== Tweak the controls ===

You can send an image to the journal and later open it with an image manipulation program. The images are exported in Portable Network Graphics (PNG) format. The RGB colors and the alpha channel is included. In the population view all images are painted on a black background. If the image is partial transparent you see this black background shining through. After loading an exported image to an image manipulation program it may look different. Depending on your image manipulation program a white background or a checkerboard is looking through. Internally Kandid stores the parameters for recreating an image not an pixel per pixel representation. Bitmaps are only used for post processing the images.

The image rendering is made from simple geometrical shapes. For example select a letter from the alphabet or a mathematical sign and draw it repeatedly on a drawing canvas. Or subdivide the drawing area in four smaller squares, pick some of these little squares by random and repeat this process for several times.

Or select a bulk of points at random positions and paint a smooth curve using these points.

Each of this rendering algorithm draws its output to a separate layer. The final image is a combination of approximately less than dozens of these layers. The layers are organized as a tree. The finally visible layer is composed from two secondary layers. The secondary layers are combined from layers below. Until you reach one of the leafs of this image rendering tree. There are different strategies for combining two layers. A simple example is adding the color values of two layers to build the resulting layer. A more complex composing strategy is using one of the layer as an mask channel to clip the other layer.

Kandid can display all the details an image is made from. Select "Explain details" and wait a few seconds. Then the program shows a scrollable view with all image layers and how they are combined. The final image is made of nothing more than the displayed details but the effects during composing are often hard to understand.

 

Select "Show ancestor" to display a family tree of an newly breeded image. On the top of the page you see the youngest image and below its ancestors. Sometimes the offspring looks different form its parents. This can have two reasons. One is mutation. In this case the impact of the mutated parameters take precedence over the inherit genome. The other reason is genome information that having no visual effect in the parent, but now in the offspring it is visible. Don't confuse the details view and the ancestor view. In the details view you see all the elements an image is made of. In the ancestor view you see the final images from the current generation and earlier generations.

In traditional genetic programming a function is needed to calculate an indication how optimal a genome is. This is called fitness function or fitness landscape. The parents with the best results are preferred producing offspring. Unfortunately some problems, like producing interesting digital images, don't have mathematicall defined fitness functions. Therefore Kandid has no build in fitness function. No assumption is made how meaningfully an image is for the user. Kasimir Malewitsch 'black square on a white field' can be reinvented with Kandid. But most users would prefer more colorful images.

There are different paradigms for implementing a genetic algorithm. ''Evolution Strategies'', invited by Ingo Rechenberg, are suitable for evolving real valued parameters to optimizing the result of a fixed formula. ''Genetic Algorithms'', introduced by John Holland are also for optimizing parameters in hard coded functions. Where the parameters are binary coded instead of floating point numbers. Genetic operations in Kandid can be better understand as 'Evolution Strategy'. But this abstract dispute.

Merging two genes is called crossing over. It is only used in sexual reproduction. Genetic algorithm can be implemented only with mutation operations. The advantage of crossing over is that the successor gene can take profit from both parent genes. Crossing over is a list operation. The new merged gene list is created by copying elements from both lists step by step. At every step a random generator decides between continuing with the current list or changing to the other list as a source for the copying. This leads to multiple crossing over points per gene sequence.

Users can report bugs and suggest an enhancement to [http://bugs.sugarlabs.org/ http://bugs.sugarlabs.org/]. Use component name 'Kandid' to tag your ticket.

* Added a simple ancestors view. The ancestors of newly generated images will be displayed as a tree. This is only a temporary solution and should be replaced later.