Stanley and his colleagues proposed a neuroevolution method, called NEAT, that evolves both recurrent neural network (RNN) topologies and weights simultaneously. Experimental comparisons show that the method is much faster than previous ones on some challenging benchmark problems. The generation alternation model employed by NEAT, however, has a drawback; it proceeds without intensively searching neighborhood of parent individuals and hence it often fails in identifying solutions with minimal structures. We propose an alternative for the model and show its effectiveness through the application to the double pole balancing without velocities problem.
