Neuroevolutive Strategies for Topology and Weights Adaptation of Artificial Neural Networks
Lucas Muniz, Carla Négri Lintzmayer, Christian Jutten, Denis G. Fantinato · 2022
Among the methods for training Multilayer Perceptron networks, backpropagation is one of the most used ones on problems of supervised learning. However, it presents some limitations, such as local convergence and the a priori choice of the network topology. Another possible approach for training is to use Genetic Algorithms to optimize the weights and topology of networks, which is known as neuroevolution. In this work, we compare the efficiency of training and defining topology with a modified neuroevolution approach using two different metaheuristics with backpropagation on 5 classification problems. The network’s efficiency is assessed through Mutual Information and Information plane. We concluded that neuroevolution found simpler topologies, while backpropagation showed higher efficiency at updating the weights.