RECURSIVE LEARNING ALGORITHMS ON RBF NETWORKS FOR NONLINEAR SYSTEM IDENTIFICATION

Science and technology development has the tendency of learning from nature
where human also try to develop artificial intelligent by imitating biological neuron
network which is popularly termed Artificial Neural Network (ANN). It represents an
interconnection among neuron which consists of several adjustable parameters which
are tuned using a set of learning examples to obtain the desired function represent the
actual system.
Radial Basis Function (RBF) networks, one of feed forward artificial neural
network architecture, have recently been given much attention due to its good
generalization ability. The RBF network is popular among scientist and engineer and
used in a number of wide ranging signal and control applications which includes the
area of system identification or estimation.
The learning approach, a process which updated the parameters of RBF networks,
will be the most important issue in neural computing research communities. The
learning method will determine the performance’s capability of the networks for the
system identification process which will be one of the key issues to be discussed in
the thesis.
This thesis proposes derivative free learning, using finite difference, methods for
fixed size RBF network in comparison to gradient based learning for the application
of system identification. The thesis also try to investigate the influence of
initialization of RBF weights parameters on the overall learning performance using
random method and advanced unsupervised learning, such as clustering techniques,
as a comparison. By taking advantage of localized Gaussian basis function of RBF
network, a decomposed version of learning method using finite difference (or derivative free) gradient estimate has been proposed in order to reduce memory
requirement for the computation of the weight updates.
The proposed training algorithms discussed in this thesis are derived for fixed size
RBF network and being compared with Extreme Learning Machine (ELM) as the
ELM technique just randomly assigned centers and width of the hidden neurons and
update the output connected weights. The proposed methods are tested using well
known nonlinear benchmark problems and also evaluated for system with irregular
sample time or known as lost packets. The finite difference based gradient estimate,
proposed in this thesis, provides a viable solution only for identifying a system with
irregular sample time.