Biology Reference
In-Depth Information
(partitioning in the parent node). When further partitioning is no longer pos-
sible, the node becomes the terminal node.
This method is attractive since it is easy to interpret the branching deci-
sions. In addition, the CART method supports wide ranges of data formats
that can be sparse in nature. It is best used as an exploratory tool, for exam-
ple, to determine important independent variables, but it does not have good
predictive power (Gaudart et al. 2005).
3.5.1.3 Neural Network
The neural network (NN) method is an information processing method
analogous to the functioning of the brain. It has remarkable capabilities in
many applications, including pattern recognitions. As described earlier in
this chapter (Section 3.4.3), the NN method is a powerful tool for ground
cover classifications. This method is suitable when little is known about the
underlying process of the disease transmission. Generally, NN have a good
predictive power. Assessing exactly how prediction is made using the pre-
dictors, however, may be less obvious.
Typical NNs consist of interconnected nodes arranged in layers, with the
independent variables acting as input and the dependent variables as out-
put. Each node carries a threshold function, while each connection carries
weight. For the nodes in the first layer, they will take the NN inputs through
its weighted connections and evaluate them using the threshold functions.
The outputs from these nodes, in turn, will be fed to the nodes in the next
layer—as determined by the topology—for further evaluation, and so forth.
The weights that connect the nodes are determined during the training
phase by minimizing the error between the predicted output and the out-
come data.
We have applied NN to model malaria transmission in Thailand using pre-
cipitation, temperature, relative humidity, and vegetation index data as the
independent variables (Kiang et al. 2006). Our study shows a distinct depen-
dency of malaria incidences with meteorological and environmental vari-
ables. Using various NN approaches, we also showed the trade-off between
training and testing accuracy as the architecture complexity increases.
3.5.1.4 Genetic Algorithms
Genetic algorithm (GA) is a search technique inspired by Darwin's evolu-
tion theory commonly used in optimization problems. For a given prob-
lem, GAs search for a solution in each iteration by evolving the current
solution. The search typically begins with a set of candidate solutions
called
populations
, which are subject to further evaluation on their “fitness.”
The objective function typically serves as the fitness function. Based on
this fitness function, the current population is chosen randomly and sub-
sequently modified (mutated or crossed over) to form a child generation of
