Geometrical Synthesis of MLP Neural Networks

This paper presents a new constructive algorithm to design multilayer perceptron networks used as classifiers. The resulting networks are able to classify patterns defined in a real domain. The proposed procedure allows us to automatically determine both the number of neurons and the synaptic weights of networks with a single hidden layer. The approach is based on linear programming. It avoids the typical local minima problems of error back propagation and assures convergence of the method. Furthermore, it will be shown in this paper that the presented procedure leads to single-hidden layer neural networks able to solve any problem in classifying a finite number of patterns. The performances of the proposed algorithm have been tested on some benchmark problems, and they have been compared with those of different approaches.