This paper describes the cascade neural network design algorithm (CNNDA), a
new algorithm for designing compact, two-hidden-layer artificial neural ne
tworks (ANNs). This algorithm determines an ANN's architecture with connect
ion weights automatically. The design strategy used in the CNNDA was intend
ed to optimize both the generalization ability and the training time of ANN
s. In order to improve the generalization ability, the CNDDA uses a combina
tion of constructive and pruning algorithms and bounded fan-ins of the hidd
en nodes. A new training approach, by which the input weights of a hidden n
ode are temporarily frozen when its output does not change much after a few
successive training cycles, was used in the CNNDA for reducing the computa
tional cost and the training time. The CNNDA was tested on several benchmar
ks including the cancer, diabetes and character-recognition problems in ANN
s. The experimental results show that the CNNDA can produce compact ANNs wi
th good generalization ability and short training time in comparison with o
ther algorithms. (C) 2001 Elsevier Science Ltd. All rights reserved.