Information Technology Reference
In-Depth Information
Figure 10. Automatic feature extraction using genetic programming
any case, a chain delivers one value. Incorporat-
ing windowings leads to the concept of method
trees:
Definition 15
A
method
tree
is a general win-
dowing whose children build a
chain. If the chain
entails a windowing, this becomes the root of a
new, embedded
method tree.
within the search space. Its four children are ex-
ponential smoothing, a filtering, another method
tree consisting of the chain just described (Fourier
transformation with peaks applied to windows),
and the average of the peaks. This last child
returns the desired features. Before the genetic
programming approach is technically described,
Figure 9 presents the process of automatically
extracting features for a given classification task
and data set.
The search space within which the best method
tree is to be found is called the universe of method
trees. Evolutionary algorithms in general search
in such a universe by maintaining populations of
elements in this universe. Hence, a population is
a set of (different) method trees. The navigation
The methods, which are performed on each
window, can be seen as children of the window-
ing operator. Together they output a value series.
The tree structure emerges from the nesting of
windowing operators.
An example of a method tree is shown in
Figure 8, where the root identifies the element
Figure 11. XML method tree representation for YALE.
<operatorname=”Root”class=”ValueSeriesPreprocessing”>
<operatorname=”Chain1”class=”OperatorChain”>
<operatorname=”ExpSm”class=”ExponentialSmoothing”/>
<operatorname=”Filter”class=”FilterTransformation”/>
<operatorname=”Windowing”class=”Windowing”>
<parameterkey=”overlap”value=”2”/>
<operatorname=”Chain2”class=”OperatorChain”>
<operatorname=”FFT”class=”FastFourierTransform”/>
<operatorname=”MaxIndex”class=”MaxIndexPoint”/>
</operator>
</operator>
<operatorname=”Avg”class=”AverageFunction”/>
</operator>
</operator>
Search WWH ::
Custom Search