Information Technology Reference
In-Depth Information
the image) of the chosen parent's parameters, or by crossing-over the parent's pa-
rameters with those of another randomly selected member of the current, or any
previous population. Crossover is an operation that samples parameters from one
image, complementary parameters from another, and recombines these to produce
a complete set of parameters for generating, in this instance, a new image. The new
image typically exhibits a combination of traits of its parents.
For a particular user the final offspring image of the 16 was consistently gener-
ated either completely at random; or in such a way as to maximise its distance from
the parent (according to the statistical image properties applied to the biomorph im-
ages above); or, by locating an individual that maximised the creativity-lite measure
discussed in relation to the biomorphs. In the latter two cases a sample of images
was generated from the parent by mutation and the one that maximised the distance
or creativity-lite measure as appropriate was chosen to fill the final slot. The off-
spring were always positioned in random order on the screen so as to avoid user
selection biases introduced by image placement.
Following their engagement with the software users were presented with a survey
asking them to rank how appealing, novel, interesting or creative they found the
intermediate images and final result.
The distance technique for generating offspring, generally, appeared to be an
impediment to the software's success. This technique was ranked somewhat worse
than the random generation technique with regard to the novelty of the images it sug-
gested, and significantly worse than the creativity technique in three of six scales we
recorded. This performance matched our intuition that maximising distance pushes
populations away from the direction selected by users, undoing the effects of evo-
lution. The creativity technique significantly outperformed the distance technique
and rated better than the random technique in novelty of the final image, and the
creativity of the suggested intermediate images. The research also found that the
mean score for all responses was best for the creativity technique; and that the pro-
portion of users who answered positively to the question “Did you feel that you
could control the quality of the images through your selections?” was higher for
creativity (55 %) than for the other two techniques (31 % for random and 40 % for
distance). Hence, we can conclude tentatively that the use of the creativity-lite mea-
sure improved the performance of the interactive algorithm with respect to natural
language notions of novelty and creativity.
13.9 Discussion
Where to from here?
EvoEco
is not the be-all and end-all of automated creativity.
For starters, the software will never be able to step outside of its specification to
generate the kind of novelty we would expect a human to devise. For instance, it
can never generate a three-dimensional model of a tree. Is this significant? A hard-
coded range of representations limits all systems we devise. This is inescapable.
Can a digital machine exceed the expectations of its developers nevertheless? We
