Abstract:
This thesis presents a new constructive algorithm called multilayered constructive
architecture (MCA) for designing and training multiple hidden layered artificial neural
networks (ANNs). Unlike most previous constructive algorithms, MCA puts emphasis
on both simplicity and generalization ability of designed ANNs. In order to maintain
simplicity, MeA uses a minimum number of user specified parameters in designing
ANNs. The use of both layered and cascaded architecture in MCA increases the
generalization ability of designed ANNs. MCA has been tested extensively on a number
of benchmark problems in machine learning and neural networks, including Australian
credit card assessment, breast cancer, diabetes, glass and heart disease. The experimental
results show that MCA can produce compact ANNs with good generalization ability.
