Optimizing the Reservoir of a Recurrent Neural Network Using Gradient-Free Methods