Evolving heterogeneous neural networks for classification problems
André L. V. Coelho, Daniel Weingaertner, Fernando José Von Zuben · 2001
This paper describes research investigating the behavior of feedforward neural networks with different neuron types when applied to classification problems. For this purpose, a hierarchical evolutionary technique (HCGA) was employed towards the automatic design of heterogeneous neural nets. In an upper level, a genetic algorithm keeps in charge of building the net topology by choosing its hidden layer neurons (possibly with distinct features). Neural nets compete against each other across the GA generations. In a bottom level, a coevolutionary approach was selected in order to train the network by adjusting both the activation function parameters of the hidden neurons and its incoming input weights. This tuning is done by means of a cooperative process where the neurons receive their fitnesses according to the average fitness of the networks in which they participate.