A Constructive Algorithm for the Training of a Multilayer Perceptron Based on the Genetic Algorithm
Hans Christian Andersen
Ah Chung Tsoi
Department of Electrical Engineering,
University of Queensland,
St. Lucia, Queensland 4072, Australia
Abstract
A genetic algorithm is proposed for the training and construction of a multilayer perceptron. The genetic algorithm works on a layer-by-layer basis. For each layer, it automatically 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 constructs networks with neurons that implement a threshold activation function. This architecture is suitable for classification problems with a single binary output.
The method is applied to the XOR problem, the n-bit parity problems, and the MONK's problems, and its performance is found to be comparable to that of other techniques.
