Using evolution to improve neural network learning: pitfalls and solutions