Information Technology Reference
In-Depth Information
Fig. 11.3
Examples of images created using an Evolutionary Engine and an adaptive AJS in the
1st (
upper row
) and 11th (
lower row
) iteration of the experiment
Table 11.4
Percentage of
images classified as external
by the ANNs used to guide
evolution in iterations 1 and
11, and the difference
between them
Set
Iteration 1
Iteration 11
Difference
Painting masterpieces
99.68 %
96.88 %
−
2
.
80 %
User-guided evolution
17.99 %
10.07 %
−
7
.
91 %
differences between runs are also clear, indicating the ability of the approach to pro-
mote a search for novelty. They also indicate that the aesthetic reference provided
by the external set manages to fulfil its goal, making it possible for AJSs to differen-
tiate between those images that may be classified as paintings and those generated
by the GP system (Machado et al.
2007
).
A set of experiments was carried out to compare the performance of the AJS from
the 1st and 11th iteration, using datasets made up of images that were not employed
in the runs. The experimental results are presented in Table
11.4
and show that the
AJS of the 11th generation performs worse than the one of the 1st iteration at clas-
sifying external imagery (a difference of 2.8 %), and better at classifying evolution
generated images (a difference of 7.91 %). These results suggest that the iterations
performed with the evolutionary engine promote the generalisation abilities of the
AJS, leading to an overall improvement in classification performance.
The integration of an AJS within a bootstrapping evolutionary system of this kind
is extremely valuable. As the results indicate, it allows the generation of images that
explore the potential weaknesses of the classifier system and the subsequent use of
these images as training instances, leading to an overall increase in performance.
Additionally, if the evolutionary system is able to generate images that the AJS is
unable to classify correctly (even after re-training it) and that a human can classify,
it shows that the set of features is not sufficient for the task at hand. Additionally, it
gives indications about the type of analysis that should be added in order to improve
the performance of the AJS.
