Cascaded Adaptive Nonlinear Functional Link Networks for Modeling and Predicting Crude Oil Prices Time Series Data

Abstract

Contrast to traditional neural networks functional link neural network (FLNN) is preferred for its single layer structural design, lower computational complexity and higher convergence rate. It achieves high dimensional representation space of input patterns through functional expansion of input signals. However, its nonlinear approximation capability is limited up to certain extent. Further improvement in performance may require enlargement in dimensionality of the input pattern, which increases the computational overhead significantly. Chebyshev FLNN (CFLNN) is a special case of FLNN and has universal approximation capacity along with faster convergence. Legendre neural network (LeNN) uses simple polynomial expansion functions and posses computational gain over FLNN. This paper develops two cascaded neural networks in order to improve the performance of FLNN. The first model combines the input expansion capacity of FLNN and better approximation of CFLNN to develop a model termed as CCFLNN. Similarly, the second model takes the advantages of FLNN and LeNN to develop another model termed as CLeFLNN. The weight and bias vectors are adjusted by gradient descent based back propagation learning method. The proposed models are evaluated on forecasting crude oil prices. Extensive simulation outcome and comparative performance investigation suggests suitability of the proposed model.