Derivative-free optimization and neural networks for robust regression