Constructive modular neural networks for system modeling and control
Loading...
Date
2004-04
Authors
Chu Kiong, Loo
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This thesis has its main focus on the development of novel neural networks for
system modeling and control based on constructive learning. Two aspects of neural
networks, viz. the modular structure design and the constructive algorithm are dealt with
comprehensively.
In line with the modular paradigm of decomposition strategy in particular and
'Divide and conquer' in general, the proposed network is strengthened with 'expertise
level' gating module, in addition to 'expertise domain' gating function which provides
crucial information to remove redundant local experts.
In addition, the importance of network structure as part of the learning process
is highlighted. This leads to the ~evelopment of a constructive learning modular neural
network called Qrowing Multi-Experts Network (GMN).
A whole gamut of local experts insertion method is further investigated which
includes, perception based insertion, errcr-driven based insertion and novelty-based
insertion. GMN employs hybrid unsupervised and supervised learning J?I'Ocess in which
the novel unsupervised algorithm called Growing Neural Gas i.l-implemented. The
network is built constructively based on perception-based insertion. In addition, a
confidence interval estimator is embedded into GMN to quantify the reliability of
GMN. Furthermore, the Self-regulating Growing Multi-Experts Network (SGMN) that
utilizes error-driven criteria for local experts insertion is also devised. The error-driven
insertion of SGMN is known as batch-mode insertion method and is not suitable for
xix
incremental learning that requires insertion based on the novelty of an input pattern.
Therefore, the Incremental GMN (IGMN) that utilizes novelty-based insertion is
developed and tested with benchmark problems. IGMN employs supervised learning
and uses certainty factor as the local expert insertion criterion. Besides, IGMN is
equipped with a robust learning algorithm to alleviate the effects of outliers.
A variety of system modeling and control tasks has been investigated to assess
the practical applicability ofthe SGMN and IGMN models. The experiments reveal the
potential of SGMN and IGMN as building blocks for system identification and control
applications. In contrast with th~ other existing modular neural network, GMN, SGMN
and IGMN allow interpolation among the local experts in a variety of ways instead of a
unique way. The network approximation capability is thus enhanced considerably as a
result of this generalization.
Description
Keywords
Modular neural networks , System modeling