Toward generating neural network structures for function approximation