Training multilayered neural networks by replacing the least fit hidden neurons

The author discusses a supervised-learning algorithm, called GenLearn, for training multilayered neural networks. GenLearn uses techniques from the field of genetic algorithms to perform a global search of weight space and, thereby, to avoid the common problem of getting stuck in local minima. GenLearn is based on survival of the fittest hidden neuron. In searching for the most fit hidden neurons, GenLearn searches for a globally optimal internal representation of the input data. A big advantage of the GenLearn procedure over the generalized delta rule (GDR) in training three-layered neural nets is that, during each iteration of GenLearn, each weight in the first matrix is modified only once, whereas, in the GDR procedure, each weight in the first matrix is modified once for each output-layer neuron. What makes this such a big advantage is that, although GenLearn often reaches the desired mean square error in about the same number of iterations as the GDR, each iteration takes considerably less time