A Constructive Algorithm for the Training of a Multilayer Perceptron Based on the Genetic Algorithm

A genetic algorit hm is proposed for the training and construction of a multilayer perceptron. The genetic algorit hm works on a layer-by-layer basis. For each layer, it automat ically chooses the number of neurons required, computes the synaptic weights between the present layer of neurons and the next layer, and gives a set of training patterns for the succeeding layer. The algorithm presented here const ruct s networks with neurons that implement a -threshold act ivation funct ion. This architecture is suitable for classificat ion problems with a single binary outp ut . The method is appl ied to the XOR problem, the n-bit pari ty problems, and the lVIONK's problems, and. its performance is found to be comparable to that of other techniques.