|
Neural Networks are currently used extensively to find
solutions to certain kinds of problems that can not be
efficiently solved by means of conventional algorithms. Neural
Networks widely applied are known as backpropagation networks.
However, backpropagation networks suffer from lengthy training
time. Furthermore, it is difficult to physically interpret the
results obtained from trained networks. This thesis proposes a
neuro-fuzzy system which can overcome these limitations. The
neuro-fuzzy system under consideration is implemented as a two-
layer Fuzzy Hyperrectangular Composite Neural Network (FHRCNN).
A special hybrid training algorithm is developed to find a set
of appropriate initial weights in order to speed up the
learning process. First we divide the output space into fuzzy
regions, and then transform function approximation into a
pattern recognition problem. In this step, we use the
supervised decision directed learning (SDDL) algorithm to find
the information imbedded in the training data. The hidden nodes
of the FHRCNN are then initialized according to the extracted
information. We may use the least mean squared error (LMS)
algorithm or the backpropagation algorithm to minimize the
error to an acceptable value. After sufficient training, the
fuzzy neural network can evolve automatically to acquire a set
of fuzzy if-then rules. Based on the experimental results we
conclude that the proposed neuro-fuzzy approach is an
attractive alternative to traditional techniques as a tool for
system identification and time series prediction.
|