Dynamic topology generation of an artificial neural network of the multilayer perceptron type

Abstract

This paper deals with an approximate constructive method to find architectures of artificial neuronal network (ANN) of the type Multi-Layer Percetron (MLP) which solves a particular problem. This method is supplemented with the technique of the Forced search of better local minima. The training of the net uses an algorithm basic descending gradient (BDG). Techniques such as repetition of the training and the early stopping (cross validation) are used to improve the results. The evaluation approach is based not only on the learning abilities but also on the generalization of the specific generated architectures of a domain. Experimental results are presented in order to prove the effectiveness of the proposed method. These are compared with architectures found by other methods.

Keywords

• Artificial neural networks
• multi-layer perceptron
• topology
• architecture