Hill climbing in recurrent neural networks for learning the anbncn language