Regression Hyper Surfaces with a Modified Neural Network Architecture
L. F. Mingo, P. Gisbert, J. D. Carrillo, C. Hernández · 2003
This paper describes the application of Enhanced Neural Networks ( ENN )to the regression analysis problem. This architecture permits to have different weights for each pattern once the network has been trained, in such a way that the Mean Square Error MSE is lower than classical Multilayer Perceptrons. These nets are employed as universal approximator of continuous mappings. The training algorithm is introduced same way as the mathematical proof of the aproximation capabilities of this architecture. Finally, several results are shown, among them, the two-spirals problem, with a comparison with other methods that have tried to solve this problem. In short, ENN have a better behaviour than the backpropagation networks, and they improveb the MSE ratio.