Adapting modularity during learning in cooperative co-evolutionary recurrent neural networks