A novel algorithm based on successive approximation training for feedforwar
d neural networks is presented in this paper. The convergence of the algori
thm is analysed theoretically and the training error is estimated. Theoreti
cal analysis shows that the novel training algorithm is able to overcome th
e stalemate problem in the later training stage of the traditional algorith
ms. Numerical experiments show that the proposed algorithm increases the ra
te of convergence and improves the generalization performance by avoiding l
ocal minima. (C) 2002 Elsevier Science B.V. All rights reserved.