An Algorithm for Determining Neural Network Architecture Using Differential Evolution

Artificial neural networks (ANNs) have been applied to a variety of classification and learning tasks. The use of evolutionary algorithms (EA) as one of the fastest, robust and efficient global search techniques has allowed different properties of artificial neural networks to be evolved. This paper proposes the possibility of using differential evolution for determining an ANN architecture (DNNA). We explain how to use differential evolutionpsilas application for determining an ANN architecture. The approach we describe is innovative and has only been successfully applied and implemented for the first time, although the idea of differential evolution has been applied in various fields since the last decade. In this work, we proposed an algorithm based on differential evolution that uses a minimum number of user specified parameters in determining an ANN architecture. By using backpropagation algorithm to train the ANN architecture partially during the evolution process, DNNA is evaluated on five benchmark classification problems, namely, Cancer, Diabetes, Heart Disease, Thyroid, and the Australian Credit Card problem. Through performance analysis and simulation studies, we show that DNNA can produce ANN architecture with good generalization abilities, but with less number of training cycles when compared with an evolutionary programming approach and standard backpropagation.