An initialization method for multilayer neural networks as function approximators

A multilayer neural network and its initialization method, which takes the distribution of given training patterns into consideration, are proposed. The network is composed of four layers, and the role of each layer in the network is specialized by analyzing the internal information processing of the network. The size of the network, initial values of weights, and parameters defining the characteristics of the nonlinearities of processing units in hidden layer are determined from a selected portion of the given training patterns. With these initial conditions, the performance of the network is further improved by the general error backpropagation learning process. The proposed model and method give any desired mapping performance with smaller network size and faster learning speed than obtained with conventional multilayer neural networks and a random initialization technique. To show the usefulness of the proposed network, an illustrative mapping is realized by the proposed algorithm