The Painting Fool: Stories from Building an Automated Painter - Computers and Creativity - page 24

Fig. 1.7 Aflower
arrangement piece from the
“Pencils, Pastels and Paint”
gallery by The Painting Fool
the scene should be negatively correlated with the rectangle's height, width and sat-
uration, so that buildings on the left and right of the scene were smaller and less
saturated, leading to a depth effect. The genome of the individuals were the list of
rectangles making up the scene. Crossover was achieved by swapping contiguous
sublists, i.e. splitting the genomes of parents into two at the same point and produc-
ing a child by taking the left hand sublist from one parent and the right hand sublist
from the other parent (and vice-versa for another child). Mutation was achieved by
randomly choosing an individual with a particular probability, the mutation rate,
for alteration. This alteration involved changing one aspect of its nature, such as
position, shape or colour.
We experimented with one-point and two-point crossover, and with various muta-
tion rates, population sizes and number of generations, until we found an evolution-
ary setup which efficiently produced scenes that looked like the tip of Manhattan
(Colton 2008a ). We turned each rectangle into a segment of a segmentation, and
The Painting Fool was able to use these invented scenes as the subject of some pic-
tures. Moreover, we used the same techniques to evolve the placement of flowers
in a wreath effect, with the rectangle position holders replaced by segmentations
of flowers. When rendered with pencil and pastel effects, these arrangements be-
came two of the pieces in the “Pencils, Pastels and Paint” permanent exhibition, as
described at www.thepaintingfool.com , with an example given in Fig. 1.7 .
In an attempt to climb the meta-mountain somewhat, we realised that in defin-
ing the fitness function, we had ultimately performed mathematical theory forma-
tion. This suggested that we could employ our HR mathematical discovery system
(Colton 2002 ), to invent fitness functions in our place. Using the same parameters
required to define the original correlations (rectangle width, height, hue, saturation,
brightness, and co-ordinates) as background information, and by implementing a
new concept formation technique involving correlations, we enabled HR to invent
new fitness functions as weighted sums of correlations over the parameters. For each
fitness function, we calculated the fitness of 100 randomly generated scenes. If the