Database Reference
In-Depth Information
In this paper, we suggest the following taxonomy. Note however that
the components of this taxonomy are not mutually exclusive, namely, there
are a few algorithms which combine two of them.
(1) Manipulating the Inducer — We manipulate the way in which the base
inducer is used. More specifically, each ensemble member is trained with
an inducer that is differently manipulated.
(2) Manipulating the Training Sample — We vary the input that is used
by the inducer for training. Each member is trained from a different
training set.
(3) Changing the target attribute representation — Each classifier in the
ensemble solve a different target concept.
(4) Partitioning the search space — Each member is trained on a different
search sub-space.
(5) Hybridization — Diversity is obtained by using various base inducers
or ensemble strategies.
9.5.1
Manipulating the Inducer
A simple method for gaining diversity is to manipulate the inducer used
for creating the classifiers. Below we survey several strategies to gain this
diversity.
9.5.1.1
Manipulation of the Inducer's Parameters
The base inducer usually can be controlled by a set of parameters. For
example, the well-known decision tree inducer C4.5 has the confidence level
parameter that greatly affect learning.
In the neural network community, there were several attempts to gain
diversity by using different number of nodes. Nevertheless, these researches
concludes that variation in numbers of hidden nodes is not effective method
of creating diversity in neural network ensembles. Nevertheless the CNNE
algorithm which simultaneously determines the ensemble size along with
the number of hidden nodes in individual NNs, has shown encouraging
results.
Another effective approach for ANNs is to use several network topolo-
gies. For instance, the Addemup algorithm
[
Opitz and Shavlik (1996)
]
uses
genetic algorithm to select the network topologies composing the ensemble.
Addemup trains with standard backpropagation, then selects groups of
Search WWH ::
Custom Search