
Digital Library of the
European Council for Modelling and Simulation 
Title: 
Growing RBF Network Models For Solving Nonlinear Approximation And
Classification Problems 
Authors: 
Gancho Vachkov, Valentin
Stoyanov, Nikolinka Christova 
Published in: 
(2015).ECMS 2015 Proceedings edited
by: Valeri M. Mladenov, Grisha Spasov, Petia Georgieva, Galidiya Petrova, European
Council for Modeling and Simulation. doi:10.7148/2015 ISBN:
9780993244001 29^{th}
European Conference on Modelling and Simulation, Albena (Varna), Bulgaria,
May 26^{th} – 29^{th},
2015 
Citation
format: 
Gancho Vachkov, Valentin Stoyanov, Nikolinka Christova (2015). Growing
RBF Network Models For Solving Nonlinear Approximation And Classification
Problems, ECMS 2015 Proceedings edited by: Valeri
M. Mladenov, Petia Georgieva, Grisha Spasov, Galidiya Petrova
European Council for Modeling and Simulation. doi:10.7148/20150481 
DOI: 
http://dx.doi.org/10.7148/20150481 
Abstract: 
In this paper a multistep learning
algorithm for creating a Growing Radial Basis Function Network (RBFN) Model
is presented and analyzed. The main concept of the algorithm is to gradually
increase by one the number of the Radial Basis Function (RBF) units at each
learning step thus gradually improving the total model quality. The algorithm
stops automatically, when a predetermined (desired) approximation error is
achieved. An essential part in the whole algorithm plays the optimization
procedure that is run at each step for increasing the number of the RBFs, but only for the parameters of the newly added RBF
unit. The parameters of all previously RBFs are
kept at their optimized values at the previous learning steps. Such
optimization strategy, while suboptimal in nature, leads to significant
reduction in the number of the parameters that have to be optimized at each
learning step. A modified constraint version of the particle swarm
optimization (PSO) algorithm with inertia weight is developed and used in the
paper. It allows obtaining conditional optimum solutions within the range of
preliminary given boundaries, which have real practical meaning. A synthetic
nonlinear test example is used in the paper to estimate the performance of
the proposed growing algorithm for nonlinear approximation. Another
application of the Growing RBFN model is illustrated on two examples for data
classification, one of them from the publicly available red wine quality data
set. 
Full
text: 