Self-Organizing Polynomial Neural Networks Based on Matrix Inversion and Differential Evolution

Abstract

Although Artificial Neural Networks (ANNs) have been extensively used to solve forecasting problems, defining their architectures has commonly been a very difficult task. Self-Organizing Polynomial Neural Networks can be used to alleviate this problem. However, it causes an increase in the computational cost and the addition of other parameters. This first drawback can be mitigated by using a matrix inversion technique as training algorithm, while the second, by using Differential Evolution. The method developed in this study combines those techniques in order to simultaneously search for the best parameters, the network architecture and weights. Finally, one can observe that in most databases the proposed method outperformed the Backpropagation, the most commonly used training algorithm in ANNs.