Improvement on function approximation capability of backpropagation neural networks

To increase the approximation accuracy of a network with a given number of units at the hidden layer, the authors propose a method in which the activation functions are trained as well as the weights by error backpropagation. They generalize the sigmoid activation function with several parameters, and derive a set of learning rules for the parameters into the form of error backpropagation. They show the usefulness of the method by a simulation example