Two-dimensional learning strategy for multilayer feedforward neural network

In the learning process of a mulitlayer feedforward neural network (MFNN), there are two independent processes: a forward process and an iterative learning process. The two processes of learning MFNN are not described entirely in general iterative learning methods, therefore it is difficult to derive an effective learning algorithm. This paper introduces a completely new concept for the design of (MFNNs) learning algorithm. A two-dimensional (2-D) general learning model is established for MFNN. The 2-D model has two independent dynamics: one reflects the feedforward process of the network and another reflects the learning process of the network. The learning of MFNN is treated from the 2-D system point of view. The learning algorithm is proposed by the convergence analysis of system error rather than minimizing the cost function. The conditions for network error convergence are stated. Experiment results indicate that the performance of the algorithm is very efficient (C) 2000 Elsevier Science B.V. All rights reserved.