Evolving Neural and Fuzzy Systems - Computational Intelligence in Time Series Forecasting - page 301

evolve the real vectors is to use the evolutionary strategies or evolutionary
programming (Fogel et al. , 1990; Yao, 1993) rather than genetic algorithms.
8.1.1.2 Evolving the Network Architecture
The first action in evolving the network architecture is to lay down the network
topological structure, i.e. the proper number of nodes, the interconnection pattern
of the nodes, the activation function to be assigned to each node, etc . This activity,
if properly carried out, is very promising in leading, through the process of
evolution, to a final network architecture, optimally shaped for the given problem
to be solved.
The adequacy of the selected topology generally depends on the network task.
For example, if the network to be evolved is to be used for identification of
nonlinear interdependencies between the collected data of a time series and to
process them, then it must be a multilayer network because a single-layer network
is not capable of doing this. Similarly, if the network has to be able to discover and
to handle the temporal dependencies in the environment, then it must be a
recurrent network because the feed-forward networks are not capable of doing
this.
A further important decision to be made when evolving neural networks is to
select the appropriate initial network topology size. For example, if the selected
network topology size is too small, then the evolved network might fail to learn the
desired input-output mapping. In contrast, if it is too large, then the generalization
capability of the network will be very poor (Sietsma and Dow, 1991).
All this indicates that, for adequate selection of initial network topology, much
expert knowledge and practical experience is needed, because here also we are
short of a well-paved way for systematic topology selection. Therefore, for the less
experienced network developer, the only way left is to select different initial
network topologies and, using a trial-and-error strategy, to find the most
appropriate one.
The next critical issue of an evolving neural network architecture is the
decision to be made about the encoding strategy to be used. Encoding strategies
help in transforming the network structure into specific representations, called
genotypes , on which the evolutionary operators (mainly mutation and
recombination) act during the process of network evolution. Both the selected
genotypes and the evolutionary operators to be used belong to the crucial issues to
be resolved before the evolving process is initiated. This is needed because the
application success of a neural network in solving the problem for which it is
evolved depends predominantly on the selected genotypic representation and on
the evolutionary operations.
For genotypic representation, two alternative encoding strategies are available:
x direct encoding strategies , in which all architectural aspects of the network
are encoded by direct transformation of genotypes to phenotypes, for
instance through building a connection matrix
x indirect encoding strategies , in which grammatical or morphological
encoding is used, based on a compressed description of the network to be
evolved.