Information Technology Reference
In-Depth Information
Fig. 1.8
Ten scenes generated for different invented fitness functions and two randomly generated
scenes
average fitness was greater than 0.8, then it was likely that optimal fitness was too
easy to achieve, and if it was less than 0.4, then it was likely that there were some
contradictions in the fitness function. Hence, we only accepted fitness functions with
an average for the 100 random scenes of between 0.4 and 0.8.
For each of ten acceptable invented fitness functions, we evolved a scene to max-
imise the fitness, and on each occasion, the scenes exhibited visually discernible
properties. Moreover, two of the scenes genuinely surprised us, because the fitness
functions had driven the search towards scenes which we didn't expect. In particular,
for one fitness function, the fittest scene involved clumping together the rectangles
in three separate centres (scene G in Fig. 1.8 ), and for another fitness function, the
fittest scene had buildings placed on top of each other (scene C), which was not ex-
pected at all. The ten scenes arising from the fitness functions are given in Fig. 1.8 ,
along with two randomly generated scenes, for comparison (R1 and R2). This ap-
proach to the invention and deployment of fitness functions is described fully in
Colton ( 2008a ). It raises the issue of software defining, employing and defending
its own aesthetic considerations, something we will come back to in future work. It
also highlights one of the accepted tenets of Computational Creativity research—
that creative software should surprise its programmers.
Specifying correlation-based fitness functions for evolutionary scene generation
worked well, but it had two main drawbacks: (i) for artistic purposes, sometimes
the scene must fully adhere to some constraints, yet there is no guarantee that it
will be possible to evolve a scene scoring 100 % for fitness, (ii) specifying a fitness
function is not always a particularly natural thing to do and it would be better if
someone using The Painting Fool's teaching interface were able to express their
desires for a scene in a visual manner. To address these issues, we investigated the
usage of constraint solving, whereby the requirements for a scene, or an element
within a scene, are expressed by dragging, scaling and changing the colour of a set
of rectangles. Following this, the constraints expressed in the scene are induced and
translated into a constraint satisfaction problem (CSP, as described by Abdennadher
Search WWH ::




Custom Search