Modularity provides advantages in neuroevolution by improving evolvability or efficiency of connections. Many techniques that leverage modularity, either by utilizing human knowledge or adding additional evolutionary objectives, have been studied. In this work, we reim-plemented and explored an existing encoding based method, MENNAG, that appears promising but has received little attention. The algorithm is tested on four tasks that are expected to receive different levels of benefits from modularity. The results show that this method is able to produce modular neural networks without directly optimizing modularity, as long as the problem has some degree of modular nature.
展开▼