Evolving Neural and Fuzzy Systems - Computational Intelligence in Time Series Forecasting - page 297

8
Evolving Neural and Fuzzy Systems
8.1 Introduction
One of the main application fields of evolutionary computation, especially of
genetic algorithms and evolutionary programming, has for a long time been the
design or evolving of intelligent computational structures, such as neural networks,
fuzzy logic systems, neuro-fuzzy systems and of their combination to implement
intelligent controllers. In the following, evolving of neural networks and fuzzy
logic systems using evolutionary algorithms will be presented.
8.1.1 Evolving Neural Networks
In evolving of neural networks for specific applications, the user is faced with the
following two key issues:
x what network architecture ( i.e. how many hidden layers, number of
neurons in each layer, and what interconnections between them) should be
selected as the most adequate
x what specific weight values should the interconnecting elements have for
optimal network performance.
No standard guidelines are available for resolving the above selection problems, at
best only some recommendations and some hints could be found in some
publications. In this chapter we will take a closer look at these selection difficulties
and we will describe some approaches that have been used successfully in evolving
of optimally design neural networks.
In the past, most very frequently a trial-and-error approach has been used in
developing the neural network structures, which have afterwards been optimized
by simulation or by some optimization methodologies. For the process of network
development, two basic approaches have been used.