Use of backpropagation and differential evolution algorithms to training MLPs

Artificial Neural Networks (ANNs) are often used (trained) to find a general solution in problems where a pattern needs to be extracted, such as data classification. Feedforward (FFNN) is one of the ANN architectures and multilayer perceptron (MLP) is a type of FFNN. Based on gradient descent, backpropagation (BP) is one of the most used algorithms for MLP training. Evolutionary algorithms can be also used to train MLPs, including Differential Evolution (DE) algorithm. In this paper, BP and DE are used to train MLPs and they are both compared in four different approaches: (a) backpropagation, (b) DE with fixed parameter values, (c) DE with adaptive parameter values and (d) a hybrid alternative using both DE+BP algorithms. © 2013 IEEE