Abstract
This paper presents a new constructive algorithm (NCA) in
designing artificial neural networks (ANNs). Unlike most previous studies on designing ANNs, NCA puts emphasis
on architectural adaptation as well as functional adaptation in its design process. This algorithm uses a constructive approach to determine both the number of hidden layers in an ANN and of neurons in each hidden
layer automatically. To achieve functional adaptation, NCA trains hidden neurons using different training sets created by employing a similar concept used in the boosting
algorithm. Eight classification problems were used to evaluate the performance of the proposed approach. The experimental results show that NCA can produce compact ANNs with good generalization ability in comparison with other algorithms.
