Genetic synthesis of piecewise-linear neural networks

A genetic algorithm is used to partially synthesize the structure of a piecewise-linear neural network appropriate for a given set of input-output examples. The networks studied are composed of linear weighted sum units and weightless switching operators, such as maximum and minimum. The procedure allows one to find the weight, type and location of every linear weighted sum unit but left unsolved the weightless switch nodes. The module can be useful as preprocessor. An application example to the modeling of the normality of electrocardiograms is provided.<<ETX>>