Evolving Neural and Fuzzy Systems - Computational Intelligence in Time Series Forecasting - page 308

genetic algorithm. The completed evolutionary fuzzy system was written in C++
code and was compiled with the Borland C++4.5 compiler. The benefits of the
developed system have been demonstrated on examples of iris data classification,
but the system could be successful for a large range of similar problems.
Several researchers have focused their attention on integration of fuzzy logic
and evolutionary approaches in optimal tuning of the parameters of a fuzzy logic
controller by adapting the fuzzy membership functions by learning the IF-THEN
rules (Varsek et al . 1993; Mohammadian and Stonier, 1994; Herrera et al . 1995).
For instance, Zeng and He (1994) evolved a fuzzy controller with a self-learning
feature for approaching the optimality conditions of the given control task.
Thereafter, the integrated genetic algorithm took over the initiative to tune the
controller parameters optimally. The modified fuzzy controller was successfully
applied to control an unstable nonlinear system that demonstrated high accuracy
and robustness of the evolved fuzzy controller.
Wong and Chen (2000) elaborated a genetic-algorithm-based approach to fuzzy
systems construction directly from collected input-output data. The basic idea of
the approach is that each individual in the population determines the number of
fuzzy rules and that the consequent part of the evolved fuzzy system is determined
by a recursive least-squares method. The effectiveness of the approach was
demonstrated on construction of some nonlinear systems.
In the recent past, some reports have been published on evolving and/or tuning
a fuzzy controller implemented using neural networks. Kim et al . (1995)
introduced a genetic-algorithm-based computationally aided design methodology
for rapid prototyping of control systems. As an example, they designed a fuzzy net
controller (FNC), with the intention to use genetic algorithms for optimizing the
fuzzy membership functions capable of meeting various operational specifications.
Seng et al . (1999) described a genetic-algorithm-based strategy for simultaneously
tuning the parameters of a fuzzy logic controller implemented on an RBF network,
named NFLC (neuro-fuzzy logic controller). Belarbi and Titel (2000) presented an
alternative approach to designing all parameters of fuzzy logic controllers ( i.e . the
parameters of the membership functions of both the input and the output variables,
and the rule base) using genetic algorithms. The fuzzy logic controller designed
was implemented in neuro-technology. The application of binary-coded genetic
algorithm was reported by Palit and Popovic (2000) in order to train a fixed
structure neuro-fuzzy network that used the singleton type of rules consequent.
Thereafter, the genetic-algorithm-trained neuro-fuzzy network was applied to
forecast the future values of a chaotic time series. In addition to the above
applications Setnes and Roubos (2000), Roubos and Setnes (2001) also applied the
a genetic-algorithm-based fuzzy logic system for identification and modeling of a
nonlinear plant. In their method, a real-coded genetic algorithm was mainly used to
fine tune the fuzzy antecedent memberships (triangular) that were obtained by
similarity-based fuzzy set merging. It was reported that the genetic-algorithm-
tuned fuzzy model was transparent, accurate but compact. Similar applications of
real-coded genetic algorithms were reported by Panchariya et al. (2003, 2004) for
improving the fuzzy model transparency. In the last case, using a distance
(entropy)-based fuzzy clustering algorithm, an initial Takagi-Sugeno fuzzy model
with high accuracy was obtained. However, the initial fuzzy model was not